1 00:00:11,119 --> 00:00:12,718 really hard. It took me like 10 its with 2 00:00:12,718 --> 00:00:16,960 Gemini to have Harsha riding a lobster. 3 00:00:16,960 --> 00:00:18,480 Gemini didn't want to do it, but we 4 00:00:18,480 --> 00:00:20,960 figured it out. Okay, welcome to harness 5 00:00:20,960 --> 00:00:24,399 night. Um, first order of business. Uh, 6 00:00:24,399 --> 00:00:26,800 how does everyone like the new YC paper 7 00:00:26,800 --> 00:00:29,199 club look? 8 00:00:29,199 --> 00:00:31,839 Well, this is like an idea that I kind 9 00:00:31,839 --> 00:00:34,719 of blurted out at EV, our head of design 10 00:00:34,719 --> 00:00:37,600 at YC, and two days later, she came back 11 00:00:37,600 --> 00:00:38,640 with this, and I'm like, "This is 12 00:00:38,640 --> 00:00:40,799 amazing." So, Ev in the back there, 13 00:00:40,799 --> 00:00:43,884 please take a bow. 14 00:00:43,884 --> 00:00:45,679 [applause] 15 00:00:45,679 --> 00:00:49,359 Very exciting. All right, why harnesses? 16 00:00:49,359 --> 00:00:51,600 Um, I mean, it's just a rapper. This is 17 00:00:51,600 --> 00:00:53,520 just scaffolding. This is just like 18 00:00:53,520 --> 00:00:56,238 prompt engineering. Um, why would this 19 00:00:56,238 --> 00:00:59,119 be at all worthy of a night? Uh, here's 20 00:00:59,119 --> 00:01:00,878 a great little Reddit that was only a 21 00:01:00,878 --> 00:01:03,119 month ago, which is actually uh the most 22 00:01:03,119 --> 00:01:04,879 aggressive. I'm not sure this kind of 23 00:01:04,879 --> 00:01:06,799 prompt engineering belongs at a top tier 24 00:01:06,799 --> 00:01:09,438 machine learning conference. Um, here's 25 00:01:09,438 --> 00:01:11,438 another great one. Uh, context 26 00:01:11,438 --> 00:01:14,159 engineering is not a research problem. 27 00:01:14,159 --> 00:01:17,519 Um, so I think harnesses have long been 28 00:01:17,519 --> 00:01:22,799 belittled uh as um subpar research. yet 29 00:01:22,799 --> 00:01:26,239 it is uh literally gives us an 18% bump 30 00:01:26,239 --> 00:01:28,000 in the difference between harness one 31 00:01:28,000 --> 00:01:31,438 and harness two. Um, and as Seth will 32 00:01:31,438 --> 00:01:32,799 tell us is the difference between 33 00:01:32,799 --> 00:01:35,759 getting ArcGI to work and not. And so 34 00:01:35,759 --> 00:01:38,640 this is a uh obviously worthy of some 35 00:01:38,640 --> 00:01:40,959 some amount of research. And I think if 36 00:01:40,959 --> 00:01:43,118 you look at this the classic meter plots 37 00:01:43,118 --> 00:01:46,319 release date and how long uh an agent 38 00:01:46,319 --> 00:01:49,040 can be running. Um, literally we go a 39 00:01:49,040 --> 00:01:50,640 lot of this progress has been because of 40 00:01:50,640 --> 00:01:52,560 harnesses. And so I call this the static 41 00:01:52,560 --> 00:01:54,799 harness era where there isn't 42 00:01:54,799 --> 00:01:56,399 self-improvement on the harness. And 43 00:01:56,399 --> 00:01:59,359 then this later latest latest um maybe 44 00:01:59,359 --> 00:02:00,959 the last 6 months has all been on the 45 00:02:00,959 --> 00:02:02,799 self improving harnesses and think and 46 00:02:02,799 --> 00:02:05,438 we'll get through all these and it's I 47 00:02:05,438 --> 00:02:07,759 actually saw this plot in a presentation 48 00:02:07,759 --> 00:02:09,919 by the CEO of trajectory. I actually 49 00:02:09,919 --> 00:02:12,318 really liked it. was talking about how 50 00:02:12,318 --> 00:02:16,159 the we we keep measuring perplexity and 51 00:02:16,159 --> 00:02:17,919 you know this somewhat correlates to IQ 52 00:02:17,919 --> 00:02:19,840 and and how intelligent the model is and 53 00:02:19,840 --> 00:02:21,840 so we keep pushing more and more up this 54 00:02:21,840 --> 00:02:25,759 intelligence uh IQ dimension but we're 55 00:02:25,759 --> 00:02:28,800 not leveraging test time experience very 56 00:02:28,800 --> 00:02:30,479 much so we're generating all this test 57 00:02:30,479 --> 00:02:32,719 time experience we have a new domain uh 58 00:02:32,719 --> 00:02:34,878 and it does and we don't really quickly 59 00:02:34,878 --> 00:02:36,719 adapt and if you remember one of the 60 00:02:36,719 --> 00:02:39,039 first YC paper clubs we did um I've been 61 00:02:39,039 --> 00:02:42,159 doing this experiment where um the as I 62 00:02:42,159 --> 00:02:44,479 increase the number of samples online 63 00:02:44,479 --> 00:02:45,919 how do you actually learn from just 64 00:02:45,919 --> 00:02:47,519 batch size one we don't really don't 65 00:02:47,519 --> 00:02:50,318 have that structure we have ICL and then 66 00:02:50,318 --> 00:02:52,878 once ICL gets saturated after just 67 00:02:52,878 --> 00:02:54,959 loaded like 40 or 50 it doesn't actually 68 00:02:54,959 --> 00:02:57,519 improve um on the valet at all then you 69 00:02:57,519 --> 00:02:59,519 have to go to Laura small rank then you 70 00:02:59,519 --> 00:03:02,000 go to Laura big rank then you go to SFT 71 00:03:02,000 --> 00:03:03,598 and it's kind of weird that we have this 72 00:03:03,598 --> 00:03:05,919 like different training procedures and 73 00:03:05,919 --> 00:03:08,000 so I think that's really where harnesses 74 00:03:08,000 --> 00:03:10,318 are shining right 75 00:03:10,318 --> 00:03:12,639 And what Arc AGI exposes, that's kind of 76 00:03:12,639 --> 00:03:16,719 the main uh um uh point of ARGI is how 77 00:03:16,719 --> 00:03:20,719 quickly it adapts to uh a new problem, a 78 00:03:20,719 --> 00:03:23,360 new a new distribution and does well in 79 00:03:23,360 --> 00:03:26,719 it. And so, um, ArcGI actually went 80 00:03:26,719 --> 00:03:28,959 through the batch with me, winter 26, 81 00:03:28,959 --> 00:03:30,639 and we helped Greg, you know, look at 82 00:03:30,639 --> 00:03:32,000 this and like the amount of thought and 83 00:03:32,000 --> 00:03:33,439 attention, as I mentioned last time, 84 00:03:33,439 --> 00:03:35,759 that goes into these these these games 85 00:03:35,759 --> 00:03:37,360 to make sure they're all orthogonal 86 00:03:37,360 --> 00:03:39,439 skills from game one to game two, so 87 00:03:39,439 --> 00:03:40,878 that it's isolating this fluid 88 00:03:40,878 --> 00:03:44,400 intelligence measure. Um really uh 89 00:03:44,400 --> 00:03:47,120 Claude uh Opus was one of the the first 90 00:03:47,120 --> 00:03:49,919 that was actually verified um on the the 91 00:03:49,919 --> 00:03:51,598 hold out on the private that no one else 92 00:03:51,598 --> 00:03:53,439 has access to other than Greg and 93 00:03:53,439 --> 00:03:55,519 Chalet. Um and the best that they got 94 00:03:55,519 --> 00:03:59,519 was 30%. And just with um some harness 95 00:03:59,519 --> 00:04:01,840 uh this thing that doesn't deserve any 96 00:04:01,840 --> 00:04:03,438 research just some wrapper and some 97 00:04:03,438 --> 00:04:06,400 scaffolding we can get to 95 and AVO 98 00:04:06,400 --> 00:04:08,719 from Nvidia got to 100%. So, Prime Agent 99 00:04:08,719 --> 00:04:11,199 and Nvidia, which both recently just 100 00:04:11,199 --> 00:04:12,878 came out. 101 00:04:12,878 --> 00:04:14,719 And the other thing I want to add, so 102 00:04:14,719 --> 00:04:17,680 when uh Carpathy launched his auto 103 00:04:17,680 --> 00:04:20,238 researcher thing in March, I want to say 104 00:04:20,238 --> 00:04:23,199 it was um I forked it and I was playing 105 00:04:23,199 --> 00:04:24,319 around with it. And all I wanted to do 106 00:04:24,319 --> 00:04:25,839 was make like a little user interface to 107 00:04:25,839 --> 00:04:27,120 kind of see what's happening and track 108 00:04:27,120 --> 00:04:30,639 it and and I ended up building a harness 109 00:04:30,639 --> 00:04:32,399 by accident. I didn't mean to, but it 110 00:04:32,399 --> 00:04:34,560 was just like I wanted to see it. And 111 00:04:34,560 --> 00:04:37,439 and basically what it is is you specify 112 00:04:37,439 --> 00:04:39,918 a purpose. And in this example, which is 113 00:04:39,918 --> 00:04:41,439 actually a true one I gave, is like 114 00:04:41,439 --> 00:04:45,600 diffusion LM don't beat ARLM. But maybe 115 00:04:45,600 --> 00:04:48,160 if I ensemble, if I shard the diffusion 116 00:04:48,160 --> 00:04:49,839 LM into a bunch of different ones 117 00:04:49,839 --> 00:04:51,439 because there's such high arithmetic 118 00:04:51,439 --> 00:04:53,600 intensity per GPU on a diffusion model 119 00:04:53,600 --> 00:04:57,120 versus AR that I can actually um in 120 00:04:57,120 --> 00:04:58,800 aggregate by sharding them, I can get 121 00:04:58,800 --> 00:05:00,478 better better results. And so I just 122 00:05:00,478 --> 00:05:02,560 give this to a per as the purpose to the 123 00:05:02,560 --> 00:05:07,759 to my uh uh to my suite of agents um my 124 00:05:07,759 --> 00:05:09,279 swarm of agents. I give it some seed 125 00:05:09,279 --> 00:05:12,240 ideas. I want to vary the ensemble size 126 00:05:12,240 --> 00:05:14,959 shard at different amounts 127 00:05:14,959 --> 00:05:16,478 100 times, 10 times, five times, 128 00:05:16,478 --> 00:05:19,120 whatever. Um and then I specify a 129 00:05:19,120 --> 00:05:21,680 valometric. Um and maybe I want to do 130 00:05:21,680 --> 00:05:24,560 GSMAK and like GPT2 setup or something 131 00:05:24,560 --> 00:05:26,879 like that. And then I have a scoping 132 00:05:26,879 --> 00:05:29,680 agent that will kind of look up papers 133 00:05:29,680 --> 00:05:31,439 that have that are are similar, look up 134 00:05:31,439 --> 00:05:34,399 GitHub repos. Um, we'll give it to a PI 135 00:05:34,399 --> 00:05:37,918 agent named Chris Ray. Um, who will then 136 00:05:37,918 --> 00:05:40,560 uh give it to a research agent um named 137 00:05:40,560 --> 00:05:44,079 John South Khan. Uh, there he is. Um, 138 00:05:44,079 --> 00:05:45,918 and then he'll work on it and then he'll 139 00:05:45,918 --> 00:05:47,600 start doing some stuff, give it to a 140 00:05:47,600 --> 00:05:49,439 council for some help. That's where I 141 00:05:49,439 --> 00:05:51,279 come in and me and Yaso will give you 142 00:05:51,279 --> 00:05:52,879 some feedback and then you work on it a 143 00:05:52,879 --> 00:05:54,800 bit more and then whenever you're ready 144 00:05:54,800 --> 00:05:56,879 and Chris Ray will kind of keep tabs on 145 00:05:56,879 --> 00:05:58,639 you, keep nagging you every every hour 146 00:05:58,639 --> 00:06:00,959 or so then it goes to this author agent 147 00:06:00,959 --> 00:06:03,279 to say okay now start stop freeze the 148 00:06:03,279 --> 00:06:06,720 idea start writing ablations and um and 149 00:06:06,720 --> 00:06:08,399 then start writing the paper and so then 150 00:06:08,399 --> 00:06:10,240 what this has turned into so this is the 151 00:06:10,240 --> 00:06:12,639 scoping agent uh and then you have this 152 00:06:12,639 --> 00:06:14,800 cockpit that kind of you can view from 153 00:06:14,800 --> 00:06:17,680 anywhere we do the uh tail scale up so 154 00:06:17,680 --> 00:06:19,199 you can actually view this URL from 155 00:06:19,199 --> 00:06:21,120 anywhere and see how it's progressing 156 00:06:21,120 --> 00:06:24,079 and and talk to it right there. Um, it 157 00:06:24,079 --> 00:06:26,478 just sends you email updates if you 158 00:06:26,478 --> 00:06:28,478 build the whole 1B thing in here. And 159 00:06:28,478 --> 00:06:29,839 then it actually starts publishing some 160 00:06:29,839 --> 00:06:31,600 papers and like I've read the papers. 161 00:06:31,600 --> 00:06:33,038 They're actually like good. They started 162 00:06:33,038 --> 00:06:35,360 out in March, April like not so good 163 00:06:35,360 --> 00:06:37,519 like you know May or whatever. But and 164 00:06:37,519 --> 00:06:40,399 now I just basically give eight ideas to 165 00:06:40,399 --> 00:06:42,879 eight H100 nodes and each one has eight 166 00:06:42,879 --> 00:06:44,879 eight H100s. and just keep going. And 167 00:06:44,879 --> 00:06:46,720 then I check in and they give give me 168 00:06:46,720 --> 00:06:48,639 back papers. And it's kind of wild what 169 00:06:48,639 --> 00:06:50,079 we're what we're dealing with there. And 170 00:06:50,079 --> 00:06:51,759 what's changed really is just the 171 00:06:51,759 --> 00:06:53,279 scaffolding that you put on top of it to 172 00:06:53,279 --> 00:06:54,879 allow it. And I I didn't spend an 173 00:06:54,879 --> 00:06:56,319 enormous amount of time on this, but 174 00:06:56,319 --> 00:06:58,000 this is largely how I I do a lot of my 175 00:06:58,000 --> 00:07:00,319 research, at least the the initial idea. 176 00:07:00,319 --> 00:07:03,120 Um, and so yeah, so I'll just give this 177 00:07:03,120 --> 00:07:04,720 six different ideas of things I want to 178 00:07:04,720 --> 00:07:06,720 try and just leave it alone and let it 179 00:07:06,720 --> 00:07:08,560 rip. And so the things that we can now 180 00:07:08,560 --> 00:07:10,240 do just because of harnesses on the same 181 00:07:10,240 --> 00:07:14,720 exact weight file is just wild. Um, so I 182 00:07:14,720 --> 00:07:16,560 looked on I spent the the weekend just 183 00:07:16,560 --> 00:07:19,439 in prep for this reading um a whole 184 00:07:19,439 --> 00:07:22,478 bunch of the classic literature from 185 00:07:22,478 --> 00:07:24,800 self-refined to reflection to Voyager to 186 00:07:24,800 --> 00:07:27,439 all the tool former and I just wanted to 187 00:07:27,439 --> 00:07:29,759 just do like a my my shot at a five 188 00:07:29,759 --> 00:07:32,560 minute like history how did we get here? 189 00:07:32,560 --> 00:07:34,639 Um I think it was it's kind of important 190 00:07:34,639 --> 00:07:36,160 and I don't think a lot of people that 191 00:07:36,160 --> 00:07:38,478 are entering AI now kind of like have 192 00:07:38,478 --> 00:07:40,079 the context. So this is not going to be 193 00:07:40,079 --> 00:07:41,360 in chronological order. It actually 194 00:07:41,360 --> 00:07:43,439 doesn't make sense to to teach it that 195 00:07:43,439 --> 00:07:45,598 way or show it that way. I'm not I'm 196 00:07:45,598 --> 00:07:47,598 going to skip a lot of papers and I may 197 00:07:47,598 --> 00:07:49,360 course grain the papers absurdly. So, 198 00:07:49,360 --> 00:07:51,439 apologies. So, the initial harness the 199 00:07:51,439 --> 00:07:55,839 GPT2 February 2019 is a while not end of 200 00:07:55,839 --> 00:07:58,478 end of sequence loop. That's and then we 201 00:07:58,478 --> 00:08:00,720 basically have top P sampling and then 202 00:08:00,720 --> 00:08:02,560 we have the environment and that's the 203 00:08:02,560 --> 00:08:03,759 harness. And there's really not that 204 00:08:03,759 --> 00:08:05,680 much. There's no tool calling, there's 205 00:08:05,680 --> 00:08:07,360 no skills, there's nothing like that. 206 00:08:07,360 --> 00:08:10,478 And so that's the V0ero harness. 207 00:08:10,478 --> 00:08:12,720 And so basically if you have this GSMAK 208 00:08:12,720 --> 00:08:14,160 example, you're a math, remember the 209 00:08:14,160 --> 00:08:15,918 system prompt. You are a math teacher. 210 00:08:15,918 --> 00:08:17,519 This is like the persona stuff we used 211 00:08:17,519 --> 00:08:19,680 to have. Um there's a context. Susie has 212 00:08:19,680 --> 00:08:21,120 five bucks, she spends three. How much 213 00:08:21,120 --> 00:08:23,199 does she have now? And then there's just 214 00:08:23,199 --> 00:08:24,720 no chain of thought. It was just like 215 00:08:24,720 --> 00:08:27,120 four hashes and two end of state and end 216 00:08:27,120 --> 00:08:30,639 of sequence that is uh the GS MK format. 217 00:08:30,639 --> 00:08:33,038 We measure uh what happens after the 218 00:08:33,038 --> 00:08:35,200 four hashes. We get accuracy and you get 219 00:08:35,200 --> 00:08:37,918 a plus one or minus one if it's wrong. 220 00:08:37,918 --> 00:08:40,799 And then the entire last six years has 221 00:08:40,799 --> 00:08:42,479 been giving more functionality into the 222 00:08:42,479 --> 00:08:45,278 harness into a static harness. And so we 223 00:08:45,278 --> 00:08:47,360 said okay well what if we give that back 224 00:08:47,360 --> 00:08:49,120 into the context a bunch of examples 225 00:08:49,120 --> 00:08:51,278 like that like well here's an example 226 00:08:51,278 --> 00:08:52,799 and let's say now we'll switch it to 227 00:08:52,799 --> 00:08:56,879 four and one and hopefully we'll have um 228 00:08:56,879 --> 00:08:58,879 from that previous example we'll say oh 229 00:08:58,879 --> 00:09:01,120 okay this makes sense I can will help to 230 00:09:01,120 --> 00:09:02,720 learn and actually this is with a fshot 231 00:09:02,720 --> 00:09:05,759 learners paper um back in July in 2020 232 00:09:05,759 --> 00:09:07,278 and then we had chain of thought and 233 00:09:07,278 --> 00:09:10,080 says well directly predicting hash hash 234 00:09:10,080 --> 00:09:12,399 2 is maybe difficult what we'll do is 235 00:09:12,399 --> 00:09:15,200 we'll smear the the compute the logic 236 00:09:15,200 --> 00:09:17,679 over many more tokens and we'll train it 237 00:09:17,679 --> 00:09:19,919 to get to that two now rather than just 238 00:09:19,919 --> 00:09:22,080 output the two and that was the chain of 239 00:09:22,080 --> 00:09:25,039 thought idea uh very cool and that was 240 00:09:25,039 --> 00:09:27,839 all context innovations uh and output 241 00:09:27,839 --> 00:09:29,120 space innovations action space 242 00:09:29,120 --> 00:09:31,278 innovations um and then then we came up 243 00:09:31,278 --> 00:09:33,919 with these tool former and webgbt webgbt 244 00:09:33,919 --> 00:09:35,759 actually came out first then tool former 245 00:09:35,759 --> 00:09:37,679 which is the idea of giving it tools and 246 00:09:37,679 --> 00:09:39,600 it can call a tool that's just really 247 00:09:39,600 --> 00:09:41,679 just a JSON object with a bunch of 248 00:09:41,679 --> 00:09:43,519 things specified but in this example 249 00:09:43,519 --> 00:09:45,759 just let's say subtraction instead of me 250 00:09:45,759 --> 00:09:48,320 calculating in the weight file what is 5 251 00:09:48,320 --> 00:09:51,600 - 2 I can just call Python and call sub 252 00:09:51,600 --> 00:09:54,320 53 and it'll tell me two and so that's 253 00:09:54,320 --> 00:09:56,639 pretty cool and you can expose all the 254 00:09:56,639 --> 00:09:59,200 tools in the system prompt and that's 255 00:09:59,200 --> 00:10:01,120 where tools come from and then we 256 00:10:01,120 --> 00:10:03,120 figured out megpt which is big the one 257 00:10:03,120 --> 00:10:04,720 of the coolest tools is being able to 258 00:10:04,720 --> 00:10:07,200 read and write to your own context and 259 00:10:07,200 --> 00:10:08,799 so before then all we could do is just 260 00:10:08,799 --> 00:10:10,559 append append append append to the 261 00:10:10,559 --> 00:10:12,159 context now we said what if I actually 262 00:10:12,159 --> 00:10:14,958 give you create read, update, delete on 263 00:10:14,958 --> 00:10:17,200 the context itself and we'll se separate 264 00:10:17,200 --> 00:10:18,639 just this one little chunk called a 265 00:10:18,639 --> 00:10:22,559 memory that you'll be able to to update. 266 00:10:22,559 --> 00:10:24,720 And then Voyager said, okay, well, we 267 00:10:24,720 --> 00:10:26,480 have these tools, but like what if I 268 00:10:26,480 --> 00:10:27,919 want to chain together these tools to 269 00:10:27,919 --> 00:10:30,240 achieve a task and then I learn it and 270 00:10:30,240 --> 00:10:32,399 how do I distill it back into the the 271 00:10:32,399 --> 00:10:35,120 system prompt to make it learn forever? 272 00:10:35,120 --> 00:10:37,278 And this is the skill this where skills 273 00:10:37,278 --> 00:10:40,078 kind of came about. And uh they did this 274 00:10:40,078 --> 00:10:42,159 on Minecraft. Um and this the Voyager 275 00:10:42,159 --> 00:10:44,078 paper and it's very cool paper. This is 276 00:10:44,078 --> 00:10:46,159 like largely now what a skill is. And so 277 00:10:46,159 --> 00:10:49,120 I have this skills.md and I have the 278 00:10:49,120 --> 00:10:51,278 name of it that I can go and search and 279 00:10:51,278 --> 00:10:53,440 here's the procedure. And then uh 280 00:10:53,440 --> 00:10:55,679 intercode this idea of like again in 281 00:10:55,679 --> 00:10:57,440 action space innovation if I can 282 00:10:57,440 --> 00:10:59,759 actually output code uh then I can 283 00:10:59,759 --> 00:11:02,799 basically now I have on the-ly uh tools 284 00:11:02,799 --> 00:11:04,559 or on the fly skills. I actually don't 285 00:11:04,559 --> 00:11:06,799 know if it's a considered a tool because 286 00:11:06,799 --> 00:11:08,879 a tool is technically an API. Am I 287 00:11:08,879 --> 00:11:10,958 outputting a function as a tool or a 288 00:11:10,958 --> 00:11:14,559 skill? I actually am still unsure. Um 289 00:11:14,559 --> 00:11:17,519 and then the react came first then 290 00:11:17,519 --> 00:11:19,839 self-refine then reflection. But this 291 00:11:19,839 --> 00:11:23,039 idea if I have multiple um agents that 292 00:11:23,039 --> 00:11:24,799 that have different roles and they can 293 00:11:24,799 --> 00:11:27,278 help self-improve self-improve on the 294 00:11:27,278 --> 00:11:30,720 context uh then I can um get smarter and 295 00:11:30,720 --> 00:11:32,559 smarter. So, I'll have take some action. 296 00:11:32,559 --> 00:11:35,440 In this example, I I I flipped the three 297 00:11:35,440 --> 00:11:38,720 and the five. Whoops. Um, I sent it to 298 00:11:38,720 --> 00:11:40,720 an internal evaluator to say, is this 299 00:11:40,720 --> 00:11:42,879 right? No, it doesn't look right. I can 300 00:11:42,879 --> 00:11:44,720 either, you know, keep looping here or I 301 00:11:44,720 --> 00:11:46,480 can go to the actual environment to get 302 00:11:46,480 --> 00:11:49,919 back a reward signal. Um, I go here uh 303 00:11:49,919 --> 00:11:51,679 uh to the reflection. They'll say, "Hey, 304 00:11:51,679 --> 00:11:53,679 you actually flip these two, go back, 305 00:11:53,679 --> 00:11:56,720 and I can improve the result." Um and 306 00:11:56,720 --> 00:12:00,078 this is the one of the first like uh uh 307 00:12:00,078 --> 00:12:02,879 ideas of like being multi- aent uh being 308 00:12:02,879 --> 00:12:05,039 let the letting the agent reflect on its 309 00:12:05,039 --> 00:12:08,399 own um output and improve it. And then 310 00:12:08,399 --> 00:12:11,039 this idea of multi- aent goes even 311 00:12:11,039 --> 00:12:13,679 further where I can actually spawn one 312 00:12:13,679 --> 00:12:16,639 of the tools could be I can spawn a a 313 00:12:16,639 --> 00:12:19,519 set of sub aents um and they they 314 00:12:19,519 --> 00:12:21,440 persist in these persistent ripples and 315 00:12:21,440 --> 00:12:22,879 they'll be able to be running and I can 316 00:12:22,879 --> 00:12:25,039 interact with them and I'll have the sub 317 00:12:25,039 --> 00:12:26,958 aent list. I can invoke those ones and I 318 00:12:26,958 --> 00:12:28,958 can keep adding to that the the launch 319 00:12:28,958 --> 00:12:32,240 sub aents and then RLM went even crazier 320 00:12:32,240 --> 00:12:34,240 to allow this in a recursive fashion so 321 00:12:34,240 --> 00:12:36,958 that um and I exposes this RLM query so 322 00:12:36,958 --> 00:12:39,200 I keep uh recursively calling the RLM 323 00:12:39,200 --> 00:12:41,600 query to uh solve a larger class of 324 00:12:41,600 --> 00:12:43,519 problems and that the leafs whenever I 325 00:12:43,519 --> 00:12:45,919 want I can call the LM query to spawn 326 00:12:45,919 --> 00:12:49,120 that LM agent as well and then I have 327 00:12:49,120 --> 00:12:51,039 this main orchestrator agent that's 328 00:12:51,039 --> 00:12:54,159 running all that and that's what I call 329 00:12:54,159 --> 00:12:56,159 like harness v1 one, this whole thing is 330 00:12:56,159 --> 00:12:57,759 like static harnesses. I'm not improving 331 00:12:57,759 --> 00:13:00,240 on the system prompt. I'm not um 332 00:13:00,240 --> 00:13:03,278 updating the harness itself. And this is 333 00:13:03,278 --> 00:13:04,399 this you, of course, we're going to have 334 00:13:04,399 --> 00:13:06,240 to have GStack on the top. He's the one 335 00:13:06,240 --> 00:13:07,360 that lets us do all this. So, we thank 336 00:13:07,360 --> 00:13:10,000 you, Gary. Love GStack. Um but there's 337 00:13:10,000 --> 00:13:11,600 some other ones that deserve a call out 338 00:13:11,600 --> 00:13:14,159 as well. And basically what that means 339 00:13:14,159 --> 00:13:15,759 to summarize all this, you have some 340 00:13:15,759 --> 00:13:18,078 agent spec, you have some system prompt 341 00:13:18,078 --> 00:13:19,919 here. You say, "How many turns am I 342 00:13:19,919 --> 00:13:21,039 allowed? How many tool calls am I 343 00:13:21,039 --> 00:13:22,078 allowed?" You don't want to allow 344 00:13:22,078 --> 00:13:24,399 infinite. You specify a tool list. you 345 00:13:24,399 --> 00:13:26,480 specify a skills list, the sub agent 346 00:13:26,480 --> 00:13:29,679 list, and that's largely what the V1 um 347 00:13:29,679 --> 00:13:32,559 is, and then you put it in a loop. And 348 00:13:32,559 --> 00:13:34,799 this can be spawned on a prompt if I'm 349 00:13:34,799 --> 00:13:37,200 asking it to do something in my Slack 350 00:13:37,200 --> 00:13:39,039 channel, which we'll hear about QM, 351 00:13:39,039 --> 00:13:40,958 which is like I use it every day. It's 352 00:13:40,958 --> 00:13:42,159 the team that made it is here. It's 353 00:13:42,159 --> 00:13:45,039 super exciting. Um and or it's on a 354 00:13:45,039 --> 00:13:46,879 crunch and just wakes up every hour and 355 00:13:46,879 --> 00:13:48,559 it decides to do work just like you 356 00:13:48,559 --> 00:13:50,958 know, anyone else. Um, and so there's 357 00:13:50,958 --> 00:13:52,720 some session management, there's a loop, 358 00:13:52,720 --> 00:13:54,559 there's this context compilation. We're 359 00:13:54,559 --> 00:13:56,240 actually creating the context. I'm 360 00:13:56,240 --> 00:13:59,198 putting uh uh all that into an LM call. 361 00:13:59,198 --> 00:14:02,399 I get back uh the action. And then I I 362 00:14:02,399 --> 00:14:04,240 may or may not have some tools that I 363 00:14:04,240 --> 00:14:07,278 need to invoke and append back into the 364 00:14:07,278 --> 00:14:09,919 context. And that's basically harness 365 00:14:09,919 --> 00:14:12,879 v1. And then the cool part, this is the 366 00:14:12,879 --> 00:14:14,320 most exciting part where we're seeing a 367 00:14:14,320 --> 00:14:16,078 lot of advancements, um, where you're 368 00:14:16,078 --> 00:14:18,639 letting the harness itself learn. either 369 00:14:18,639 --> 00:14:20,159 we're learning the system prompt or 370 00:14:20,159 --> 00:14:22,078 we're learning the harness itself, which 371 00:14:22,078 --> 00:14:24,879 is very trippy. And so, one of the 372 00:14:24,879 --> 00:14:26,720 famous ones that I actually wanted them 373 00:14:26,720 --> 00:14:29,039 to talk, but they're actually running a 374 00:14:29,039 --> 00:14:32,240 150 person DSPY meetup tonight and in in 375 00:14:32,240 --> 00:14:33,600 this in San Francisco. They couldn't 376 00:14:33,600 --> 00:14:35,600 make it. Um, but I've done some podcast 377 00:14:35,600 --> 00:14:37,039 with them before and a super great 378 00:14:37,039 --> 00:14:39,039 community. Is this DSPY? So, 379 00:14:39,039 --> 00:14:42,000 demonstrate, search, uh, predict. uh 380 00:14:42,000 --> 00:14:45,278 they have this idea of of basically uh 381 00:14:45,278 --> 00:14:48,320 taking a train set a small set of 382 00:14:48,320 --> 00:14:50,639 examples and then learning the optimal 383 00:14:50,639 --> 00:14:52,320 system prompt. So I keep iterating, 384 00:14:52,320 --> 00:14:54,320 iterating, iterating. I can't back prop 385 00:14:54,320 --> 00:14:56,320 through that process, but I can do uh 386 00:14:56,320 --> 00:14:57,519 something called genetic programming 387 00:14:57,519 --> 00:14:58,958 where I'm finding candidates, I'm 388 00:14:58,958 --> 00:15:00,720 merging candidates, and with some merge 389 00:15:00,720 --> 00:15:02,480 rule, I'm evaluating, seeing what 390 00:15:02,480 --> 00:15:04,559 happens, and I keep working and working. 391 00:15:04,559 --> 00:15:06,639 And I basically gives me CRUD over the 392 00:15:06,639 --> 00:15:08,639 system prompt itself, and allows me to 393 00:15:08,639 --> 00:15:10,399 choose any system prompt. And then 394 00:15:10,399 --> 00:15:12,639 Darwin machines actually go a step 395 00:15:12,639 --> 00:15:14,320 further. Not only are you allowed to 396 00:15:14,320 --> 00:15:15,600 change the system prompt, but you're 397 00:15:15,600 --> 00:15:17,120 allowed to change the harness itself, 398 00:15:17,120 --> 00:15:19,039 the harness code that is actually 399 00:15:19,039 --> 00:15:21,039 running. And so you can imagine 400 00:15:21,039 --> 00:15:23,039 basically what happens is you have this 401 00:15:23,039 --> 00:15:27,198 archive of um many different uh agents 402 00:15:27,198 --> 00:15:30,159 which is harness and system prompt. Uh 403 00:15:30,159 --> 00:15:33,919 and you sample from them. You push them 404 00:15:33,919 --> 00:15:35,919 uh through you'll actually evaluate how 405 00:15:35,919 --> 00:15:38,240 how they did on some fit fitness 406 00:15:38,240 --> 00:15:41,679 function. Uh you'll uh add it back into 407 00:15:41,679 --> 00:15:43,919 into this archive state. And I skipped 408 00:15:43,919 --> 00:15:45,600 over the the self modify. You can 409 00:15:45,600 --> 00:15:48,159 actually you have a a meta harness that 410 00:15:48,159 --> 00:15:49,839 actually allows the the agent to modify 411 00:15:49,839 --> 00:15:52,159 its own uh harness so that it can become 412 00:15:52,159 --> 00:15:53,839 a different harness and then that 413 00:15:53,839 --> 00:15:56,320 basically loops around loops around and 414 00:15:56,320 --> 00:15:57,919 eventually you get better and better 415 00:15:57,919 --> 00:15:59,600 agents over time. And then the meta 416 00:15:59,600 --> 00:16:01,278 harness which is the main harness is to 417 00:16:01,278 --> 00:16:03,039 produce harnesses, right? Which is a 418 00:16:03,039 --> 00:16:05,519 really meta concept. Um and so this is 419 00:16:05,519 --> 00:16:06,799 the output space where you're 420 00:16:06,799 --> 00:16:08,879 configuring multi- aent context 421 00:16:08,879 --> 00:16:11,440 compilation. Um, and you're you're doing 422 00:16:11,440 --> 00:16:13,360 this and you have allow CRUD on all of 423 00:16:13,360 --> 00:16:15,440 it. And so it just keeps adding more and 424 00:16:15,440 --> 00:16:18,879 more uh CRUD into the harness code, all 425 00:16:18,879 --> 00:16:21,600 the green that you see here, the uh meta 426 00:16:21,600 --> 00:16:23,759 prompt, the system prompts of all the 427 00:16:23,759 --> 00:16:25,759 agents, how many agents there are. Um, 428 00:16:25,759 --> 00:16:28,000 and you kind of grow this uh this meta 429 00:16:28,000 --> 00:16:30,399 harness over time. And then one of the 430 00:16:30,399 --> 00:16:32,958 authors of this paper is the lead author 431 00:16:32,958 --> 00:16:34,320 is actually here tonight. Super 432 00:16:34,320 --> 00:16:35,600 exciting. You got to drive down and talk 433 00:16:35,600 --> 00:16:37,839 to him a lot of it about um continual 434 00:16:37,839 --> 00:16:40,159 harness, which is I love it. And they go 435 00:16:40,159 --> 00:16:43,198 even a step further where one they add 436 00:16:43,198 --> 00:16:46,720 they they add some extra color on the 437 00:16:46,720 --> 00:16:49,039 classes of memory and so they add this 438 00:16:49,039 --> 00:16:50,720 history thing. You know he'll he'll go 439 00:16:50,720 --> 00:16:52,078 into a bunch bunch more detail on that 440 00:16:52,078 --> 00:16:54,639 memory breakdown. But the coolest part I 441 00:16:54,639 --> 00:16:56,799 think is for the classic RL people. I 442 00:16:56,799 --> 00:16:58,480 see Robert back there. He definitely 443 00:16:58,480 --> 00:17:00,799 would would enjoy this Dagger style 444 00:17:00,799 --> 00:17:03,039 online learning where you can actually 445 00:17:03,039 --> 00:17:05,199 update the weight file itself. So you're 446 00:17:05,199 --> 00:17:08,640 actually doing a test time training on 447 00:17:08,640 --> 00:17:10,480 the LLM based on a small amount of 448 00:17:10,480 --> 00:17:12,078 examples that you just learned which I 449 00:17:12,078 --> 00:17:14,240 think is actually a huge huge important 450 00:17:14,240 --> 00:17:16,000 research direction that we should get 451 00:17:16,000 --> 00:17:20,160 working. Anyway, that's it. How you do 452 00:17:20,160 --> 00:17:26,640 all 453 00:17:26,640 --> 00:17:28,880 authors. Uh Ben got food poisoning this 454 00:17:28,880 --> 00:17:30,240 morning so he couldn't make it. It was 455 00:17:30,240 --> 00:17:32,319 very sad but we have three tremendous 456 00:17:32,319 --> 00:17:37,519 authors. uh Seth uh student under shein 457 00:17:37,519 --> 00:17:39,679 is I say it right uh at Princeton a 458 00:17:39,679 --> 00:17:41,279 researcher at Prime Intellect and the 459 00:17:41,279 --> 00:17:44,880 author of Prime Agent um John Sadvalone 460 00:17:44,880 --> 00:17:47,038 who's a the first time we've had a a 461 00:17:47,038 --> 00:17:49,839 call back um pres presenter very excited 462 00:17:49,839 --> 00:17:53,679 about that PhD under Chris an Aelia um 463 00:17:53,679 --> 00:17:56,160 and he's going to be he's the author of 464 00:17:56,160 --> 00:17:59,119 open Jarvis with Ivonica uh from my lab 465 00:17:59,119 --> 00:18:01,759 uh hazy which is a personal uh open 466 00:18:01,759 --> 00:18:02,480 Jarvis 467 00:18:02,480 --> 00:18:03,919 and which I think is really cool. And 468 00:18:03,919 --> 00:18:07,440 then for Josh and Rean who Josh just got 469 00:18:07,440 --> 00:18:09,519 promoted to be head of YC labs which is 470 00:18:09,519 --> 00:18:11,440 really exciting which I think deserves a 471 00:18:11,440 --> 00:18:13,207 little round of applause as well. 472 00:18:13,207 --> 00:18:14,558 [applause] 473 00:18:14,558 --> 00:18:18,000 Um be talking about QM and so QM there 474 00:18:18,000 --> 00:18:20,160 at YC we had this general agent that we 475 00:18:20,160 --> 00:18:21,759 use and they came out with QM a month 476 00:18:21,759 --> 00:18:24,319 ago and it is like meaningfully uh 477 00:18:24,319 --> 00:18:26,160 better and is so functional and I use it 478 00:18:26,160 --> 00:18:28,558 every day. All right, that's all I got. 479 00:18:28,558 --> 00:18:36,558 Thank you so much. [applause] 480 00:18:37,279 --> 00:18:38,960 tonight. Uh, my name is Seth and I'm 481 00:18:38,960 --> 00:18:40,240 going to be presenting Prime Agent, 482 00:18:40,240 --> 00:18:42,880 which is a self-improving ROM harness. 483 00:18:42,880 --> 00:18:45,440 Um, my job here tonight, I think, is to 484 00:18:45,440 --> 00:18:48,319 try to convince all of you to take a 485 00:18:48,319 --> 00:18:52,000 very first principles style approach to 486 00:18:52,000 --> 00:18:53,119 thinking about how you build your 487 00:18:53,119 --> 00:18:54,880 harness. So, we're going to be very, 488 00:18:54,880 --> 00:18:56,798 very basic here to begin with. If you 489 00:18:56,798 --> 00:18:58,880 think about and I love the uh 490 00:18:58,880 --> 00:19:00,640 introduction right from of all the 491 00:19:00,640 --> 00:19:02,160 background literature um which I thought 492 00:19:02,160 --> 00:19:03,119 was so cool I think is very 493 00:19:03,119 --> 00:19:04,798 complimentary the how I think about this 494 00:19:04,798 --> 00:19:06,480 as well um but if you think about just 495 00:19:06,480 --> 00:19:09,359 the raw LLM itself it's actually just 496 00:19:09,359 --> 00:19:11,919 this sequential processor that you know 497 00:19:11,919 --> 00:19:14,480 has some fixed weights and has some 498 00:19:14,480 --> 00:19:16,480 visible context and it's it's taking 499 00:19:16,480 --> 00:19:17,919 some tokens in and it's putting some 500 00:19:17,919 --> 00:19:20,640 tokens out uh to make the next decision. 501 00:19:20,640 --> 00:19:24,319 Um we don't really think of an LLM that 502 00:19:24,319 --> 00:19:26,558 way these days. We have a set of files 503 00:19:26,558 --> 00:19:29,200 that it has access to. We give endless 504 00:19:29,200 --> 00:19:31,839 programs and tools for it to use. Um, 505 00:19:31,839 --> 00:19:33,759 you can even message other sessions of 506 00:19:33,759 --> 00:19:35,279 LLMs that are going on and create sub 507 00:19:35,279 --> 00:19:37,440 agents in order to do all of these very 508 00:19:37,440 --> 00:19:39,279 cool things. But, but in its basics, 509 00:19:39,279 --> 00:19:41,279 it's just tokens in, tokens out. It's a 510 00:19:41,279 --> 00:19:43,279 neural network making a prediction. The 511 00:19:43,279 --> 00:19:46,079 harness itself is the layer between the 512 00:19:46,079 --> 00:19:48,558 LLM and the world that adds things like 513 00:19:48,558 --> 00:19:51,279 this persistent state tools and compute. 514 00:19:51,279 --> 00:19:53,440 Uh, here we have a diagram of how we 515 00:19:53,440 --> 00:19:55,839 think about prime agent. um from the 516 00:19:55,839 --> 00:19:58,079 human's perspective. So you open up your 517 00:19:58,079 --> 00:19:59,919 prime agent uh on your computer just 518 00:19:59,919 --> 00:20:01,599 like you would cloud code, codec, pi, 519 00:20:01,599 --> 00:20:03,599 etc. Um and it gets you into this agents 520 00:20:03,599 --> 00:20:05,599 view and this agents view is an overview 521 00:20:05,599 --> 00:20:06,960 of all the agents that you have going on 522 00:20:06,960 --> 00:20:08,720 for all your parallel sessions with like 523 00:20:08,720 --> 00:20:10,880 a very tight like tight summary of what 524 00:20:10,880 --> 00:20:12,640 you're using and then you can hop into 525 00:20:12,640 --> 00:20:14,640 one of those and check it out. there 526 00:20:14,640 --> 00:20:16,160 you're going to be at this root session 527 00:20:16,160 --> 00:20:17,519 and this [clears throat] root session is 528 00:20:17,519 --> 00:20:20,400 your basically the project orchestrator 529 00:20:20,400 --> 00:20:22,000 over all of these different sub aents 530 00:20:22,000 --> 00:20:23,759 that it's controlling and you don't have 531 00:20:23,759 --> 00:20:26,079 to ask it to start sub agents it will 532 00:20:26,079 --> 00:20:28,400 leverage sub agents when they're useful 533 00:20:28,400 --> 00:20:30,000 um and these are all programmatically 534 00:20:30,000 --> 00:20:32,079 called because this is all based on uh 535 00:20:32,079 --> 00:20:34,240 the recursive language model principle 536 00:20:34,240 --> 00:20:37,038 where everything is in this IPython 537 00:20:37,038 --> 00:20:39,519 shell um all your tools all your 538 00:20:39,519 --> 00:20:42,159 memories all your sub aents uh and and 539 00:20:42,159 --> 00:20:44,640 we expose uh for further coordination 540 00:20:44,640 --> 00:20:46,798 these uh these messaging paradigms. So 541 00:20:46,798 --> 00:20:49,599 you can manage all of these um and then 542 00:20:49,599 --> 00:20:50,960 these agents can directly interact with 543 00:20:50,960 --> 00:20:52,240 your environment which could just be 544 00:20:52,240 --> 00:20:53,679 like the programs and files on your 545 00:20:53,679 --> 00:20:56,079 computer or running an H200 node cluster 546 00:20:56,079 --> 00:20:57,440 maybe for your auto research. Each of 547 00:20:57,440 --> 00:20:59,038 these agents are then backed by this 548 00:20:59,038 --> 00:21:01,519 persistent dam on your computer. Uh this 549 00:21:01,519 --> 00:21:03,519 is so that you know when you close uh 550 00:21:03,519 --> 00:21:05,599 your laptop or you you control C out of 551 00:21:05,599 --> 00:21:07,038 the session, it's still running in the 552 00:21:07,038 --> 00:21:08,159 background. You have to actually stop 553 00:21:08,159 --> 00:21:09,679 the session so that make sure that 554 00:21:09,679 --> 00:21:11,759 you're continuing to running. And then 555 00:21:11,759 --> 00:21:13,279 we also have these other features we 556 00:21:13,279 --> 00:21:15,359 exposed from continual harness where 557 00:21:15,359 --> 00:21:18,000 it's able to provide like live CRUD 558 00:21:18,000 --> 00:21:20,079 operations on all of the components that 559 00:21:20,079 --> 00:21:23,599 we mentioned um in order to manage its 560 00:21:23,599 --> 00:21:26,558 uh memory skills sub agents um 561 00:21:26,558 --> 00:21:28,480 persistent and and prompt its own system 562 00:21:28,480 --> 00:21:30,240 prompt persistently. The way I like to 563 00:21:30,240 --> 00:21:32,159 think about all of this context that 564 00:21:32,159 --> 00:21:34,720 we're building up is that we have this 565 00:21:34,720 --> 00:21:37,200 sort of almost like like here I have 566 00:21:37,200 --> 00:21:40,960 like L1, L2, L3. is like a cache, right? 567 00:21:40,960 --> 00:21:42,400 It's like what is the most accessible 568 00:21:42,400 --> 00:21:44,798 information that we're working with and 569 00:21:44,798 --> 00:21:47,119 at the very like fastest like readily 570 00:21:47,119 --> 00:21:48,319 available information, you know, the 571 00:21:48,319 --> 00:21:49,440 models to be able to retrieve that 572 00:21:49,440 --> 00:21:51,038 really quickly. It's the model weights. 573 00:21:51,038 --> 00:21:52,558 So, everyone always wants to get all the 574 00:21:52,558 --> 00:21:54,400 information in the model weights. Um, 575 00:21:54,400 --> 00:21:55,599 but then we said, okay, well, maybe we 576 00:21:55,599 --> 00:21:56,640 don't have all the information because 577 00:21:56,640 --> 00:21:57,919 we don't want to have to fine-tune every 578 00:21:57,919 --> 00:21:59,038 single time to update because that's 579 00:21:59,038 --> 00:22:01,599 very expensive. So, we have this uh 580 00:22:01,599 --> 00:22:03,919 active input context. So, we're using 581 00:22:03,919 --> 00:22:06,079 lots and lots of tokens on the input. we 582 00:22:06,079 --> 00:22:08,079 might have some in context examples like 583 00:22:08,079 --> 00:22:10,720 we've seen previously in order to add to 584 00:22:10,720 --> 00:22:12,720 these different capabilities but at a 585 00:22:12,720 --> 00:22:15,679 certain point we run out of context. Um 586 00:22:15,679 --> 00:22:18,720 and so the very like earliest form of 587 00:22:18,720 --> 00:22:20,159 harnesses that we've seen that are still 588 00:22:20,159 --> 00:22:21,839 used to this day even by those who say 589 00:22:21,839 --> 00:22:23,200 we want the most minimal harness 590 00:22:23,200 --> 00:22:25,279 possible is compaction because 591 00:22:25,279 --> 00:22:28,079 compaction is a very generalized tool 592 00:22:28,079 --> 00:22:31,679 for the agent to be able to uh summarize 593 00:22:31,679 --> 00:22:33,919 its own context history in order to work 594 00:22:33,919 --> 00:22:36,319 past its context length working window. 595 00:22:36,319 --> 00:22:39,038 You can think about um once we go beyond 596 00:22:39,038 --> 00:22:40,558 like what are directly like inputs and 597 00:22:40,558 --> 00:22:42,480 outputs from the the model here into 598 00:22:42,480 --> 00:22:44,319 this L2L3. You might be familiar with 599 00:22:44,319 --> 00:22:46,319 the L3 which is more of the dispatch 600 00:22:46,319 --> 00:22:47,839 state. So if you're working with a file 601 00:22:47,839 --> 00:22:50,000 system, you can read and write from uh 602 00:22:50,000 --> 00:22:52,880 main memory. Uh if we're at the L2, 603 00:22:52,880 --> 00:22:55,519 which is I could think uh at a means in 604 00:22:55,519 --> 00:22:57,519 between uh what the active context is 605 00:22:57,519 --> 00:22:59,038 and working with your file system, you 606 00:22:59,038 --> 00:23:02,000 might have a live uh ripple, which could 607 00:23:02,000 --> 00:23:04,079 just be running things directly in 608 00:23:04,079 --> 00:23:06,319 Python uh an IPython shell like you're 609 00:23:06,319 --> 00:23:08,480 in a Jupyter notebook. And all of those 610 00:23:08,480 --> 00:23:10,000 variables are saved directly in your 611 00:23:10,000 --> 00:23:11,519 RAM. your agent can then 612 00:23:11,519 --> 00:23:13,359 programmatically manipulate them and run 613 00:23:13,359 --> 00:23:14,960 all sorts of programs directly on the 614 00:23:14,960 --> 00:23:16,880 information there saving tons of tokens 615 00:23:16,880 --> 00:23:18,480 rather than putting it directly into 616 00:23:18,480 --> 00:23:20,480 context. Uh you can also create sub 617 00:23:20,480 --> 00:23:21,679 agents and it's the same thing you're 618 00:23:21,679 --> 00:23:23,440 basically saving context here because 619 00:23:23,440 --> 00:23:25,679 you can task the agent with a specific 620 00:23:25,679 --> 00:23:27,759 set of information in order to perform 621 00:23:27,759 --> 00:23:29,519 some operations and the report back at 622 00:23:29,519 --> 00:23:31,919 the end. What is interesting here is 623 00:23:31,919 --> 00:23:34,480 what we talked about compaction for the 624 00:23:34,480 --> 00:23:36,480 active context, right? You have your 625 00:23:36,480 --> 00:23:38,240 context history. This is helping to 626 00:23:38,240 --> 00:23:39,759 update it over time. So you can continue 627 00:23:39,759 --> 00:23:42,319 to leverage this. But once we go beyond 628 00:23:42,319 --> 00:23:43,839 this, we need to be thinking about how 629 00:23:43,839 --> 00:23:45,599 are we doing these update. We talked 630 00:23:45,599 --> 00:23:47,839 about CRUD. How are we do beyond just 631 00:23:47,839 --> 00:23:49,440 creating reading? How are we updating 632 00:23:49,440 --> 00:23:52,159 and deleting our context over time 633 00:23:52,159 --> 00:23:55,119 beyond our uh or the state over time 634 00:23:55,119 --> 00:23:57,759 beyond the context line. Uh I like to 635 00:23:57,759 --> 00:23:59,359 think of this at the ripple is this 636 00:23:59,359 --> 00:24:01,038 aentic garbage collection where we're 637 00:24:01,038 --> 00:24:02,558 just cleaning up the variables in our 638 00:24:02,558 --> 00:24:04,240 state as well as like what sub agents 639 00:24:04,240 --> 00:24:06,720 could be used. And then so to make sure 640 00:24:06,720 --> 00:24:08,480 that our RAM doesn't crash my computer 641 00:24:08,480 --> 00:24:10,319 laptop every day. And then on top of 642 00:24:10,319 --> 00:24:11,919 that, we have this uh notion of 643 00:24:11,919 --> 00:24:14,000 refinement where we're updating and 644 00:24:14,000 --> 00:24:16,000 deleting the skills and memories and 645 00:24:16,000 --> 00:24:17,839 prompts that are stored on your system. 646 00:24:17,839 --> 00:24:19,359 You can think of that so that way you 647 00:24:19,359 --> 00:24:21,599 don't crash your actual uh out of space 648 00:24:21,599 --> 00:24:24,558 on your hard drive as well. And so this 649 00:24:24,558 --> 00:24:26,880 very much is a here's how we express 650 00:24:26,880 --> 00:24:29,038 this thing and here's how we revise it 651 00:24:29,038 --> 00:24:30,640 over time. The other perspective that I 652 00:24:30,640 --> 00:24:32,000 really like to think about and I really 653 00:24:32,000 --> 00:24:34,079 trying to push because harnesses are 654 00:24:34,079 --> 00:24:35,519 almost going towards this like agentic 655 00:24:35,519 --> 00:24:37,679 operating system that we're creating uh 656 00:24:37,679 --> 00:24:39,839 is when I think of it um more 657 00:24:39,839 --> 00:24:42,000 metaphorically here is that when you 658 00:24:42,000 --> 00:24:44,400 look at the raw LLM it kind of looks 659 00:24:44,400 --> 00:24:45,919 more like a touring machine where you 660 00:24:45,919 --> 00:24:47,440 have this ticker tape uh and you have 661 00:24:47,440 --> 00:24:48,798 all these instructions that are going in 662 00:24:48,798 --> 00:24:50,159 and then it's performing some set of 663 00:24:50,159 --> 00:24:52,480 operations and going out. But when you 664 00:24:52,480 --> 00:24:55,038 look at a harness it's looking a lot 665 00:24:55,038 --> 00:24:57,440 more vono like a vonoyman computer. 666 00:24:57,440 --> 00:24:58,880 you're able to do these read and write 667 00:24:58,880 --> 00:25:01,679 operations on external memory and that 668 00:25:01,679 --> 00:25:03,919 makes it much more powerful and another 669 00:25:03,919 --> 00:25:05,519 class of problems than just what a 670 00:25:05,519 --> 00:25:06,880 touring machine is able to express on 671 00:25:06,880 --> 00:25:08,960 its own. And so yeah, the idea of like 672 00:25:08,960 --> 00:25:10,400 how do you build a good hardness? You 673 00:25:10,400 --> 00:25:12,640 want it to be the most expressable thing 674 00:25:12,640 --> 00:25:16,319 you can imagine. So some some like early 675 00:25:16,319 --> 00:25:18,000 harnesses before it gets into the data 676 00:25:18,000 --> 00:25:19,278 flywheel where the models can do 677 00:25:19,278 --> 00:25:22,240 themselves are very specific. Plan, act, 678 00:25:22,240 --> 00:25:25,759 critique, do these exact specific um 679 00:25:25,759 --> 00:25:28,720 steps. Well, now har uh the models are 680 00:25:28,720 --> 00:25:30,079 able to do that themselves. You can 681 00:25:30,079 --> 00:25:31,599 imagine like we we don't have like a 682 00:25:31,599 --> 00:25:33,200 react loop that we necessarily need to 683 00:25:33,200 --> 00:25:35,440 explicitly impose. The models kind of 684 00:25:35,440 --> 00:25:37,759 have natively uh figured this out. But 685 00:25:37,759 --> 00:25:39,839 what they haven't figured out is how to 686 00:25:39,839 --> 00:25:41,359 um you know they have to be able to have 687 00:25:41,359 --> 00:25:43,599 the expressibility to call compact. They 688 00:25:43,599 --> 00:25:45,440 have to be able to have a Python ripple 689 00:25:45,440 --> 00:25:47,679 so they can run programs. Um they have 690 00:25:47,679 --> 00:25:49,599 to have the ability to programmatically 691 00:25:49,599 --> 00:25:51,599 create sub agents and access state and 692 00:25:51,599 --> 00:25:53,679 have different feedback mechanisms. Th 693 00:25:53,679 --> 00:25:54,880 those are model controlled 694 00:25:54,880 --> 00:25:56,240 expressibility features and if you 695 00:25:56,240 --> 00:25:57,599 removed one of those you're actually 696 00:25:57,599 --> 00:25:59,200 removing a capability that it won't be 697 00:25:59,200 --> 00:26:01,759 able to do otherwise. The way we manage 698 00:26:01,759 --> 00:26:02,880 um and I'm sure you're all familiar with 699 00:26:02,880 --> 00:26:05,759 the RLM paper uh from my co-author Alex 700 00:26:05,759 --> 00:26:08,159 um fantastic bit work. What we do beyond 701 00:26:08,159 --> 00:26:10,400 what was in the RLM paper is we think 702 00:26:10,400 --> 00:26:12,640 about the age sub aents as these 703 00:26:12,640 --> 00:26:14,798 persistent subsessions. 704 00:26:14,798 --> 00:26:17,679 So each the parent station can create uh 705 00:26:17,679 --> 00:26:20,558 spin up a new RLM sub aent and each of 706 00:26:20,558 --> 00:26:22,319 these are then emitted. They run some 707 00:26:22,319 --> 00:26:24,720 task and then they finish and report 708 00:26:24,720 --> 00:26:27,200 back to some end state to the parent 709 00:26:27,200 --> 00:26:28,798 session. These are then idle. They're 710 00:26:28,798 --> 00:26:30,880 still working in your RAM. At any point, 711 00:26:30,880 --> 00:26:32,640 the parent session can then send a 712 00:26:32,640 --> 00:26:35,278 message to one of the sub aents to 713 00:26:35,278 --> 00:26:36,960 continue working and it has all that 714 00:26:36,960 --> 00:26:38,960 good context that you built up over time 715 00:26:38,960 --> 00:26:40,319 so that you're not missing information 716 00:26:40,319 --> 00:26:42,159 or have to reuse information that was 717 00:26:42,159 --> 00:26:44,960 already developed in a prior context. 718 00:26:44,960 --> 00:26:46,480 And then of course you know we don't 719 00:26:46,480 --> 00:26:47,839 want to use a lot of RAM. So we can move 720 00:26:47,839 --> 00:26:50,319 them offloaded uh in an inactive state 721 00:26:50,319 --> 00:26:51,759 which then can be called back at any 722 00:26:51,759 --> 00:26:53,119 time by messaging them in this 723 00:26:53,119 --> 00:26:55,200 persistent sub agent setup. I talked a 724 00:26:55,200 --> 00:26:57,119 little bit about uh continual harness uh 725 00:26:57,119 --> 00:26:59,359 which we have in a a prior paper of mine 726 00:26:59,359 --> 00:27:02,480 which talks about cuding the entire uh 727 00:27:02,480 --> 00:27:04,400 harness state. Um this is another 728 00:27:04,400 --> 00:27:05,839 feature that we want in our coding 729 00:27:05,839 --> 00:27:08,159 agents leverage all of our prior 730 00:27:08,159 --> 00:27:09,759 history. So you can imagine like some 731 00:27:09,759 --> 00:27:11,278 set of trajectories where they have some 732 00:27:11,278 --> 00:27:12,720 actions and outcomes or something 733 00:27:12,720 --> 00:27:16,319 happened um at a at each turn. And so we 734 00:27:16,319 --> 00:27:17,919 just kind of want to expose the ability 735 00:27:17,919 --> 00:27:19,440 for the agent to leverage all that 736 00:27:19,440 --> 00:27:22,240 information in order to update what the 737 00:27:22,240 --> 00:27:23,440 future harness is going to look like. 738 00:27:23,440 --> 00:27:24,798 Are we do we need to change our system 739 00:27:24,798 --> 00:27:26,880 prompt? Do we need to create some 740 00:27:26,880 --> 00:27:29,200 skills? And skills I think of as a set 741 00:27:29,200 --> 00:27:31,759 of instructions or a program in order to 742 00:27:31,759 --> 00:27:34,960 to achieve some some specific goal. uh 743 00:27:34,960 --> 00:27:36,798 memory which could just be long-term 744 00:27:36,798 --> 00:27:38,640 storage about things that are important 745 00:27:38,640 --> 00:27:40,558 as well as the sub aent specifications 746 00:27:40,558 --> 00:27:41,759 that we talked about in this very 747 00:27:41,759 --> 00:27:43,839 persistent manner. Were there uh certain 748 00:27:43,839 --> 00:27:45,839 sub aents that we want to reuse at a 749 00:27:45,839 --> 00:27:48,000 later time because the context is useful 750 00:27:48,000 --> 00:27:49,679 and just having the ability to do this 751 00:27:49,679 --> 00:27:52,159 kind of reflection or refinement um over 752 00:27:52,159 --> 00:27:55,359 time. It's very powerful for the models 753 00:27:55,359 --> 00:27:56,880 to have. They're not perfect at this 754 00:27:56,880 --> 00:27:58,720 right now, but this is one of the the 755 00:27:58,720 --> 00:28:00,240 capabilities that you want to you want 756 00:28:00,240 --> 00:28:03,038 to build your harness such that it is a 757 00:28:03,038 --> 00:28:05,119 bit better than what the current models 758 00:28:05,119 --> 00:28:06,640 are able to do. So then you can get 759 00:28:06,640 --> 00:28:08,319 those reasoning traces and use that to 760 00:28:08,319 --> 00:28:09,919 leverage your next iteration of model 761 00:28:09,919 --> 00:28:12,240 and they'll be able to handle the 762 00:28:12,240 --> 00:28:13,440 harness and be able to bootstrap 763 00:28:13,440 --> 00:28:14,558 themselves into a higher and higher 764 00:28:14,558 --> 00:28:16,240 performance. One of the coolest features 765 00:28:16,240 --> 00:28:18,880 that we have in uh Prime Agent um that 766 00:28:18,880 --> 00:28:21,119 we we've had since the beginning of when 767 00:28:21,119 --> 00:28:22,480 I was working on this, this is one of 768 00:28:22,480 --> 00:28:24,480 the first things I added um is the 769 00:28:24,480 --> 00:28:27,278 ability to message between any any two 770 00:28:27,278 --> 00:28:29,599 agents um within like some nuclear 771 00:28:29,599 --> 00:28:31,599 family setup, parents, children, uh 772 00:28:31,599 --> 00:28:34,079 siblings. Um and the reason why I did 773 00:28:34,079 --> 00:28:35,679 this is because I was I was constantly 774 00:28:35,679 --> 00:28:36,798 trying to figure out what's the best way 775 00:28:36,798 --> 00:28:38,640 to like myself to manage all of the 776 00:28:38,640 --> 00:28:40,960 agents I have doing everything for me in 777 00:28:40,960 --> 00:28:42,480 five different directions, five billion 778 00:28:42,480 --> 00:28:44,319 different directions every day. Um, and 779 00:28:44,319 --> 00:28:45,359 it would be so much better if they could 780 00:28:45,359 --> 00:28:46,720 just like share their contacts directly 781 00:28:46,720 --> 00:28:48,398 with each other and coordinate. And 782 00:28:48,398 --> 00:28:50,480 turns out that's fantastic for like 783 00:28:50,480 --> 00:28:52,319 typical software engineering and long 784 00:28:52,319 --> 00:28:54,640 horizon jobs as well. Uh, the last thing 785 00:28:54,640 --> 00:28:57,119 that we look at when it comes to how did 786 00:28:57,119 --> 00:28:59,599 we want to design our harness is we were 787 00:28:59,599 --> 00:29:01,119 really thinking about long horizon 788 00:29:01,119 --> 00:29:03,519 performance. I want to go run some jobs 789 00:29:03,519 --> 00:29:05,038 and I don't want to have to babysit my 790 00:29:05,038 --> 00:29:06,880 agents the entire time and when when I'm 791 00:29:06,880 --> 00:29:08,558 ready to come back and check in, I can 792 00:29:08,558 --> 00:29:10,159 check in with them and see what's going 793 00:29:10,159 --> 00:29:12,480 on. And this is a perspective that I 794 00:29:12,480 --> 00:29:14,960 also really lack seeing in a lot of the 795 00:29:14,960 --> 00:29:17,519 evaluations that we're looking at. Uh a 796 00:29:17,519 --> 00:29:19,839 lot of times if you run a model for not 797 00:29:19,839 --> 00:29:21,440 enough time or say, oh well the model 798 00:29:21,440 --> 00:29:22,960 stopped working after this amount of 799 00:29:22,960 --> 00:29:25,038 budgets, but then this other model kept 800 00:29:25,038 --> 00:29:26,720 working with using more budgets. Well, 801 00:29:26,720 --> 00:29:28,159 first of all, you're not even using the 802 00:29:28,159 --> 00:29:30,000 same fixed expenditure to compare the 803 00:29:30,000 --> 00:29:31,599 models. But second of all, that could 804 00:29:31,599 --> 00:29:33,200 also be hiding performance that you're 805 00:29:33,200 --> 00:29:35,440 missing. Uh the way that I look at long 806 00:29:35,440 --> 00:29:37,278 horizon performance eval is that I want 807 00:29:37,278 --> 00:29:39,519 to see what's the practical plateau. At 808 00:29:39,519 --> 00:29:42,240 what point will we only get incremental 809 00:29:42,240 --> 00:29:44,000 gains in performance as I throw more 810 00:29:44,000 --> 00:29:45,839 test time tokens at it? I have a couple 811 00:29:45,839 --> 00:29:47,519 experiments that I'm going to show after 812 00:29:47,519 --> 00:29:49,278 we've shared design philosophy here 813 00:29:49,278 --> 00:29:52,079 about how we created uh prime agent. Um 814 00:29:52,079 --> 00:29:53,200 we're going to talk a little bit about 815 00:29:53,200 --> 00:29:55,839 test time scaling and uh our our TI 816 00:29:55,839 --> 00:29:58,720 results as well as looking at um you 817 00:29:58,720 --> 00:30:00,079 know does is it actually helpful and why 818 00:30:00,079 --> 00:30:01,359 is it actually helpful for our 819 00:30:01,359 --> 00:30:03,038 information management for the ripple 820 00:30:03,038 --> 00:30:04,079 that we're working on these long 821 00:30:04,079 --> 00:30:06,480 contexts and then um when we have these 822 00:30:06,480 --> 00:30:08,558 really really long like almost ultra 823 00:30:08,558 --> 00:30:11,759 horizon long horizon uh tasks uh how do 824 00:30:11,759 --> 00:30:13,278 we sustain these like multi-day work and 825 00:30:13,278 --> 00:30:15,119 like what actually goes on when we have 826 00:30:15,119 --> 00:30:18,000 these refinements um over like these 827 00:30:18,000 --> 00:30:19,839 settings that can last like a week at a 828 00:30:19,839 --> 00:30:21,440 time or more. So, this is a result that 829 00:30:21,440 --> 00:30:23,200 you probably all seen. We actually have 830 00:30:23,200 --> 00:30:24,960 a one additional data point that we 831 00:30:24,960 --> 00:30:26,319 added here that we didn't include in our 832 00:30:26,319 --> 00:30:28,480 original result uh just to compare 833 00:30:28,480 --> 00:30:30,640 across harnesses. We solved this uh we 834 00:30:30,640 --> 00:30:31,679 went out, we're trying to figure out 835 00:30:31,679 --> 00:30:34,240 what is the the best uh eval that people 836 00:30:34,240 --> 00:30:35,440 care about these days when we're running 837 00:30:35,440 --> 00:30:37,119 our harnesses and we're like, "Oh, we 838 00:30:37,119 --> 00:30:38,558 should do RKGI." I like, "Oh, yeah. 839 00:30:38,558 --> 00:30:40,000 Yeah, I remember. I I ran some results 840 00:30:40,000 --> 00:30:42,000 with continual harness and we got 20% 841 00:30:42,000 --> 00:30:45,519 with um Gemini Flash uh or sorry, Gemini 842 00:30:45,519 --> 00:30:47,278 Pearl." Um so, I I think we can get at 843 00:30:47,278 --> 00:30:48,720 least 20%. people who think that's 844 00:30:48,720 --> 00:30:49,839 really cool that our like general 845 00:30:49,839 --> 00:30:51,839 harness that didn't even like wasn't 846 00:30:51,839 --> 00:30:53,759 even structured for ARHI did really 847 00:30:53,759 --> 00:30:55,278 well. So I went online I was like okay I 848 00:30:55,278 --> 00:30:56,398 need to find a good system prompt 849 00:30:56,398 --> 00:30:57,599 because I don't want to make sure that 850 00:30:57,599 --> 00:30:59,519 we're losing information. So I found a 851 00:30:59,519 --> 00:31:01,278 another community leaderboard called 852 00:31:01,278 --> 00:31:02,798 prolong and I just grabbed their system 853 00:31:02,798 --> 00:31:03,839 prompt and I was like okay I'm going to 854 00:31:03,839 --> 00:31:05,599 grab their system prompt forget the rest 855 00:31:05,599 --> 00:31:06,558 and I'm just going to throw this 856 00:31:06,558 --> 00:31:08,558 directly into prime agent. Uh and then I 857 00:31:08,558 --> 00:31:11,200 ran this and I was like oh my god the 858 00:31:11,200 --> 00:31:14,640 first run that I got it hit 99.9% 859 00:31:14,640 --> 00:31:15,919 and then I looked at the logs and I was 860 00:31:15,919 --> 00:31:17,119 cheating. Okay. So, I was like, "Okay, I 861 00:31:17,119 --> 00:31:18,798 got to do proper sandboxing here. Like, 862 00:31:18,798 --> 00:31:20,480 let's set this up properly." Uh, and 863 00:31:20,480 --> 00:31:21,839 then so I spent another day on this. And 864 00:31:21,839 --> 00:31:23,119 then, and then I went back and I was 865 00:31:23,119 --> 00:31:26,000 like, "Oh my god, 866 00:31:26,000 --> 00:31:28,798 I got 78% with GPT soul. Like, this is 867 00:31:28,798 --> 00:31:30,558 going to be a great result." Um, and 868 00:31:30,558 --> 00:31:31,839 then we're back. It's like, oh, let's 869 00:31:31,839 --> 00:31:33,759 let's compare a couple other ones. And 870 00:31:33,759 --> 00:31:35,759 so, it's again, we just took the prompt, 871 00:31:35,759 --> 00:31:37,440 uh, general prompt that basically says, 872 00:31:37,440 --> 00:31:40,079 uh, use a world model to solve ARC AGI 873 00:31:40,079 --> 00:31:42,000 3. Uh, here are the actions that you can 874 00:31:42,000 --> 00:31:44,240 take. um you have uh and then the 875 00:31:44,240 --> 00:31:45,519 general system prompt for prime agent 876 00:31:45,519 --> 00:31:46,720 which is like you have a ripple you can 877 00:31:46,720 --> 00:31:49,119 call sub agents uh you can use the it 878 00:31:49,119 --> 00:31:51,599 programmatically um and and so we went 879 00:31:51,599 --> 00:31:52,720 through I went through the traces and 880 00:31:52,720 --> 00:31:53,919 it's basically doing a bunch of 881 00:31:53,919 --> 00:31:56,798 different um like calls of the coding in 882 00:31:56,798 --> 00:31:58,079 order to like check out these different 883 00:31:58,079 --> 00:32:00,000 scenarios and analyzing the images and 884 00:32:00,000 --> 00:32:01,679 doing like image processing and it's a 885 00:32:01,679 --> 00:32:04,880 lot of really cool um stuff that uh it 886 00:32:04,880 --> 00:32:06,000 seems like it was doing reasonable 887 00:32:06,000 --> 00:32:08,159 reasoning while leveraging the the 888 00:32:08,159 --> 00:32:10,558 ripple that we had um as like one of the 889 00:32:10,558 --> 00:32:12,159 main things that was able to enable 890 00:32:12,159 --> 00:32:13,679 build this. Uh so I went through and I 891 00:32:13,679 --> 00:32:15,599 ran a couple other ones. We did GPT tero 892 00:32:15,599 --> 00:32:17,919 25.7% which is really cool. You can see 893 00:32:17,919 --> 00:32:20,480 that compared to like what were the um 894 00:32:20,480 --> 00:32:22,159 like the week before we did this uh open 895 00:32:22,159 --> 00:32:24,079 AAI was like the guys the harness 896 00:32:24,079 --> 00:32:24,960 matters a lot when you're doing 897 00:32:24,960 --> 00:32:27,200 evaluations. We use the responses API. 898 00:32:27,200 --> 00:32:30,960 This is the result that we got. Um and 899 00:32:30,960 --> 00:32:34,798 we we ran Terra and and got almost like 900 00:32:34,798 --> 00:32:36,960 we we didn't run to completion this one 901 00:32:36,960 --> 00:32:38,798 but we got really good results in 902 00:32:38,798 --> 00:32:41,839 comparison. And then we go um that that 903 00:32:41,839 --> 00:32:43,359 we're already achieving higher than some 904 00:32:43,359 --> 00:32:45,038 of like the GBT soul extra high which 905 00:32:45,038 --> 00:32:48,398 was crazy. And then uh we went and we 906 00:32:48,398 --> 00:32:50,960 did Opus and hit 95.5%. We're like 907 00:32:50,960 --> 00:32:53,200 that's insane. We also compared to a lot 908 00:32:53,200 --> 00:32:54,558 of the other harnesses. So some people 909 00:32:54,558 --> 00:32:55,839 ask me like did you run this with cloud 910 00:32:55,839 --> 00:32:58,558 code? Uh I did. Um unfortunately the 911 00:32:58,558 --> 00:33:00,480 results weren't very good. Um, and so 912 00:33:00,480 --> 00:33:02,480 rather than having bad results, I just 913 00:33:02,480 --> 00:33:04,480 deferred to the the original cloud code 914 00:33:04,480 --> 00:33:05,679 results and some other people have run 915 00:33:05,679 --> 00:33:07,278 it uh with similar configurations to 916 00:33:07,278 --> 00:33:08,640 prime agent and gotten much better 917 00:33:08,640 --> 00:33:10,398 results since then. Um, but what's 918 00:33:10,398 --> 00:33:12,398 interesting is that a lot of the really 919 00:33:12,398 --> 00:33:14,558 popular harnesses don't necessarily do 920 00:33:14,558 --> 00:33:16,240 well when prime agent does well. So like 921 00:33:16,240 --> 00:33:18,398 for air agent, um, we spent a lot of 922 00:33:18,398 --> 00:33:20,640 money very quickly and uh, we had to cut 923 00:33:20,640 --> 00:33:22,319 it off because I spent like $5,000 924 00:33:22,319 --> 00:33:24,640 without making much performance. Um, not 925 00:33:24,640 --> 00:33:25,839 saying this is the best they could do, 926 00:33:25,839 --> 00:33:28,398 but it cost a lot of money to do so. Uh 927 00:33:28,398 --> 00:33:31,038 so I think that the cost to performance 928 00:33:31,038 --> 00:33:32,319 uh ratio is very important and one of 929 00:33:32,319 --> 00:33:33,759 the things that does save money is being 930 00:33:33,759 --> 00:33:36,240 able to programmatically work with your 931 00:33:36,240 --> 00:33:38,880 context. Uh we ran a bunch of long 932 00:33:38,880 --> 00:33:42,000 horizon um evals as well like oolong and 933 00:33:42,000 --> 00:33:44,159 some coding uh emulator bench which is 934 00:33:44,159 --> 00:33:45,519 going to come out soon which is a 935 00:33:45,519 --> 00:33:47,759 program bench alternative and we found 936 00:33:47,759 --> 00:33:49,839 that it was mainly parody or slightly 937 00:33:49,839 --> 00:33:52,240 better than these other harnesses like 938 00:33:52,240 --> 00:33:54,159 you across different models versus doing 939 00:33:54,159 --> 00:33:56,960 like pimono cloud codecs with glm 5.2 to 940 00:33:56,960 --> 00:34:00,319 Opus 5 and 5.6 as our setting. Another 941 00:34:00,319 --> 00:34:01,599 one I thought was really cool is we have 942 00:34:01,599 --> 00:34:04,240 this like program bench alternative 943 00:34:04,240 --> 00:34:05,679 called emulator bench where we're trying 944 00:34:05,679 --> 00:34:07,919 to reproduce entire emulators of 945 00:34:07,919 --> 00:34:09,280 computer systems or in this case 946 00:34:09,280 --> 00:34:11,119 creating like a Game Boy Color and check 947 00:34:11,119 --> 00:34:13,358 that out. And we found that what's 948 00:34:13,358 --> 00:34:14,878 really interesting is because it has 949 00:34:14,878 --> 00:34:18,000 this um ripple access in the RLM, it's 950 00:34:18,000 --> 00:34:20,320 able to use these programs in order to 951 00:34:20,320 --> 00:34:22,960 kind of do these like out of experiment 952 00:34:22,960 --> 00:34:26,320 uh loop designs in order to um try 953 00:34:26,320 --> 00:34:29,119 things out in a lot more expressable and 954 00:34:29,119 --> 00:34:30,719 free way before submitting the final 955 00:34:30,719 --> 00:34:32,639 solution to the greater. Uh we also 956 00:34:32,639 --> 00:34:35,358 tried this with uh GPU kernels um and we 957 00:34:35,358 --> 00:34:37,679 got about par results uh across 958 00:34:37,679 --> 00:34:39,918 different um both soul and kimico. One 959 00:34:39,918 --> 00:34:42,000 is better, one is worse. about par um 960 00:34:42,000 --> 00:34:43,838 which so we we're not overfit to like 961 00:34:43,838 --> 00:34:47,760 any one particular um evaluation here. 962 00:34:47,760 --> 00:34:49,760 Um what's interesting for the long 963 00:34:49,760 --> 00:34:51,760 horizon stuff is we had some auto 964 00:34:51,760 --> 00:34:54,559 research uh experiments that we did with 965 00:34:54,559 --> 00:34:56,480 the nano GPT speedrun but we scaled it 966 00:34:56,480 --> 00:34:59,039 up. We said let's give it uh 8 by H200 967 00:34:59,039 --> 00:35:02,320 for um a week and see what happens. And 968 00:35:02,320 --> 00:35:04,079 you might be like, okay, prime age is 969 00:35:04,079 --> 00:35:04,960 going to do so much better, right? 970 00:35:04,960 --> 00:35:05,838 Because it's able to do all this 971 00:35:05,838 --> 00:35:07,519 programming. Uh, it's a little high 972 00:35:07,519 --> 00:35:09,920 variance. We can't attribute um any of 973 00:35:09,920 --> 00:35:11,280 the benefits to with the harness versus 974 00:35:11,280 --> 00:35:12,719 the model there because it's a very hard 975 00:35:12,719 --> 00:35:14,559 task. But what we can do is inspect a 976 00:35:14,559 --> 00:35:16,800 lot of the behavior that we've seen. And 977 00:35:16,800 --> 00:35:18,480 what's really interesting is that we're 978 00:35:18,480 --> 00:35:20,880 seeing models like deep 6v4, GLM 5.3, 979 00:35:20,880 --> 00:35:22,960 and Kim K3. Um, you can tell these were 980 00:35:22,960 --> 00:35:24,239 done a little more recently than our 981 00:35:24,239 --> 00:35:27,280 first results. Uh, and we took these and 982 00:35:27,280 --> 00:35:28,800 they were doing like what we call out of 983 00:35:28,800 --> 00:35:30,320 loop experiments. So we were trying to 984 00:35:30,320 --> 00:35:32,000 say how can I run experiments on like 985 00:35:32,000 --> 00:35:33,280 the CPU and like look at the 986 00:35:33,280 --> 00:35:35,280 parameterization and do hyperparameter 987 00:35:35,280 --> 00:35:38,400 search and analyze the data so that I 988 00:35:38,400 --> 00:35:40,079 don't have to spend like all my time 989 00:35:40,079 --> 00:35:43,119 running expensive H200 experiments uh 990 00:35:43,119 --> 00:35:44,480 because that takes the majority of the 991 00:35:44,480 --> 00:35:46,239 time. So it's it's running experiments 992 00:35:46,239 --> 00:35:48,079 that are not the main experiment in 993 00:35:48,079 --> 00:35:49,358 order to optimize them. I think that's 994 00:35:49,358 --> 00:35:51,599 really cool behavior that we're seeing 995 00:35:51,599 --> 00:35:54,559 uh as we we shape what would be what 996 00:35:54,559 --> 00:35:56,320 kind of things we need to for the 997 00:35:56,320 --> 00:35:57,838 expressability for prime agent. So you 998 00:35:57,838 --> 00:35:59,519 can use like really good auto research 999 00:35:59,519 --> 00:36:00,639 because you can imagine if it's good at 1000 00:36:00,639 --> 00:36:02,000 auto research, it'll be good with you. 1001 00:36:02,000 --> 00:36:04,000 It be even better with a human in the 1002 00:36:04,000 --> 00:36:05,599 loop to bootstrap your experiments. Uh 1003 00:36:05,599 --> 00:36:07,760 and finally, we also streamed a 7-day 1004 00:36:07,760 --> 00:36:11,519 factorial run which used a total of 633 1005 00:36:11,519 --> 00:36:16,159 agents um across uh 23 million tok 1006 00:36:16,159 --> 00:36:18,320 output tokens in order to make like 1007 00:36:18,320 --> 00:36:20,719 steady uh tech technological advancement 1008 00:36:20,719 --> 00:36:22,480 across the tech tree to continue to 1009 00:36:22,480 --> 00:36:25,358 progress over time. And here it uh one 1010 00:36:25,358 --> 00:36:27,199 of the main benefits is I can use like 1011 00:36:27,199 --> 00:36:29,280 these sub aents that can divvy up into 1012 00:36:29,280 --> 00:36:31,280 different tasks in the factory in order 1013 00:36:31,280 --> 00:36:33,760 to research and build and gather 1014 00:36:33,760 --> 00:36:35,760 resources and build the next items to 1015 00:36:35,760 --> 00:36:38,960 design the factory. Um as well as it can 1016 00:36:38,960 --> 00:36:40,400 use the refinement to leverage what 1017 00:36:40,400 --> 00:36:42,800 happened in the past in order to help in 1018 00:36:42,800 --> 00:36:44,719 the future um over these very long 1019 00:36:44,719 --> 00:36:46,400 context so it doesn't get stuck. And one 1020 00:36:46,400 --> 00:36:47,599 of the most interesting things here is 1021 00:36:47,599 --> 00:36:48,960 that it does not get stuck and it 1022 00:36:48,960 --> 00:36:50,639 continues to make technology progression 1023 00:36:50,639 --> 00:36:54,000 even at the end of our uh stage. Um this 1024 00:36:54,000 --> 00:36:55,599 is more like a Gemini plays Pokemon kind 1025 00:36:55,599 --> 00:36:57,280 of uh conclusion here. If there's one 1026 00:36:57,280 --> 00:36:59,679 thing that uh I find interesting today u 1027 00:36:59,679 --> 00:37:00,800 but like what takeaways you should 1028 00:37:00,800 --> 00:37:02,719 actually add to your own harness. Um I 1029 00:37:02,719 --> 00:37:03,679 think that you should think about 1030 00:37:03,679 --> 00:37:06,000 agentic context management. Uh you 1031 00:37:06,000 --> 00:37:07,599 should think about swarms and looking 1032 00:37:07,599 --> 00:37:10,000 into further depth RLMs and trying to 1033 00:37:10,000 --> 00:37:11,838 run standardized eval. All of the 1034 00:37:11,838 --> 00:37:13,199 results that we can they showed today 1035 00:37:13,199 --> 00:37:15,760 can be run with our uh verifiers uh 1036 00:37:15,760 --> 00:37:18,079 package that we have at Prime Inslect. 1037 00:37:18,079 --> 00:37:20,239 Um and shout out to my collaborators who 1038 00:37:20,239 --> 00:37:22,400 are fantastic and I love working with. 1039 00:37:22,400 --> 00:37:24,054 Thanks. 1040 00:37:24,054 --> 00:37:31,119 [applause] 1041 00:37:31,119 --> 00:37:33,039 >> Hey everybody. I'm super excited to talk 1042 00:37:33,039 --> 00:37:34,800 about a project um that we've been 1043 00:37:34,800 --> 00:37:37,039 working on at Stanford. Um, I've been 1044 00:37:37,039 --> 00:37:39,519 working on this with Ivanka Orion, my my 1045 00:37:39,519 --> 00:37:41,920 co-lead author, as well as our adviserss 1046 00:37:41,920 --> 00:37:46,079 Hazeni and Christopher Ray. So, personal 1047 00:37:46,079 --> 00:37:47,760 AI is everywhere, but it's mostly 1048 00:37:47,760 --> 00:37:49,920 cloudbound today. Uh, we see lots of 1049 00:37:49,920 --> 00:37:52,559 different harnesses and projects focused 1050 00:37:52,559 --> 00:37:54,960 on making daily writing, research, 1051 00:37:54,960 --> 00:37:57,199 coding, and scheduling. But projects 1052 00:37:57,199 --> 00:37:59,519 like OpenClaw and Hermes agent typically 1053 00:37:59,519 --> 00:38:02,800 rely on cloud LMS um for most of the 1054 00:38:02,800 --> 00:38:04,639 intelligence and for most of the most of 1055 00:38:04,639 --> 00:38:07,280 the queries. Um, what does this mean? It 1056 00:38:07,280 --> 00:38:09,358 means that it's pretty costly. You're 1057 00:38:09,358 --> 00:38:11,039 getting thousands and thousands of 1058 00:38:11,039 --> 00:38:13,039 dollars in API costs if you aggregate it 1059 00:38:13,039 --> 00:38:14,960 over a year. Um, it's not private. 1060 00:38:14,960 --> 00:38:17,519 You're often sending your most personal 1061 00:38:17,519 --> 00:38:20,079 um, data to LMS up in the cloud and you 1062 00:38:20,079 --> 00:38:21,280 don't necessarily know where all that 1063 00:38:21,280 --> 00:38:23,280 data is going. Um, it also requires you 1064 00:38:23,280 --> 00:38:25,039 to rent your intelligence as opposed to 1065 00:38:25,039 --> 00:38:26,880 just simply owning it out of the box. 1066 00:38:26,880 --> 00:38:28,800 And finally, it tends to consume orders 1067 00:38:28,800 --> 00:38:30,800 of magnitude more energy than just 1068 00:38:30,800 --> 00:38:33,679 running these LMS on your laptop. And so 1069 00:38:33,679 --> 00:38:35,920 the local LMS are finally good enough to 1070 00:38:35,920 --> 00:38:37,838 actually run a lot of these queries that 1071 00:38:37,838 --> 00:38:40,800 people care about. And so we see that um 1072 00:38:40,800 --> 00:38:43,119 the the current LMS of today are only 6 1073 00:38:43,119 --> 00:38:45,679 to 12 months uh behind whatever is the 1074 00:38:45,679 --> 00:38:48,079 state-of-the-art frontier models um of 1075 00:38:48,079 --> 00:38:50,960 before. So you see um LMS today such as 1076 00:38:50,960 --> 00:38:54,320 Quen 3.8 27B um that achieve roughly the 1077 00:38:54,320 --> 00:38:57,039 same performance as like Claude 4.6 Opus 1078 00:38:57,039 --> 00:38:58,320 um back in the day. So that was kind of 1079 00:38:58,320 --> 00:38:59,838 the state-of-the-art model back in 1080 00:38:59,838 --> 00:39:02,880 August 2025. Um, and that gap seems to 1081 00:39:02,880 --> 00:39:05,358 be closing uh more and more as the 1082 00:39:05,358 --> 00:39:07,199 hardware accelerators that we have um 1083 00:39:07,199 --> 00:39:09,119 for our laptops and for our workstations 1084 00:39:09,119 --> 00:39:10,880 get better and better. Uh, just this 1085 00:39:10,880 --> 00:39:12,880 week we saw a new release from Apple um 1086 00:39:12,880 --> 00:39:14,800 with the new Mac Mini. And so we're 1087 00:39:14,800 --> 00:39:16,960 seeing this renewed focus from Apple as 1088 00:39:16,960 --> 00:39:19,039 well as Nvidia to build accelerators 1089 00:39:19,039 --> 00:39:21,519 specifically for personal use cases. And 1090 00:39:21,519 --> 00:39:23,119 so with this project, we wanted to 1091 00:39:23,119 --> 00:39:25,760 explore the the main question of can we 1092 00:39:25,760 --> 00:39:27,920 build the core of a personal AI stack, 1093 00:39:27,920 --> 00:39:30,000 namely the model inference, the agent 1094 00:39:30,000 --> 00:39:32,159 execution, the memory, the learning, 1095 00:39:32,159 --> 00:39:34,079 basically the parts that are mostly 1096 00:39:34,079 --> 00:39:36,639 reliant on the cloud today entirely on 1097 00:39:36,639 --> 00:39:38,400 device while staying competitive with 1098 00:39:38,400 --> 00:39:40,559 these cloudonly stacks. And so we 1099 00:39:40,559 --> 00:39:43,039 decided to propose open Jarvis. Um name 1100 00:39:43,039 --> 00:39:45,599 needs no needs no explanation. Um but we 1101 00:39:45,599 --> 00:39:46,960 wanted to explore just how much of this 1102 00:39:46,960 --> 00:39:48,960 we could run on device completely for 1103 00:39:48,960 --> 00:39:51,358 for free uh while preserving uh 1104 00:39:51,358 --> 00:39:54,079 security, privacy and quality. And so to 1105 00:39:54,079 --> 00:39:55,519 construct open Jarvis we wanted to 1106 00:39:55,519 --> 00:39:57,119 create the simplest set of primitives 1107 00:39:57,119 --> 00:39:59,519 for which you define any sort of harness 1108 00:39:59,519 --> 00:40:01,679 or or personal AI stack. Um the first 1109 00:40:01,679 --> 00:40:03,358 one is whatever user interfaces you need 1110 00:40:03,358 --> 00:40:05,440 to use. Um the second one is the actual 1111 00:40:05,440 --> 00:40:07,519 agentic logic around composable 1112 00:40:07,519 --> 00:40:09,280 reasoning and using different kinds of 1113 00:40:09,280 --> 00:40:11,280 intelligence and tools. Uh for the 1114 00:40:11,280 --> 00:40:12,880 intelligence, it's whatever LM you're 1115 00:40:12,880 --> 00:40:14,239 using as your engine for keeping 1116 00:40:14,239 --> 00:40:16,000 everything going. Um so this could be 1117 00:40:16,000 --> 00:40:19,838 Quen, GBDO, OSS, Gemma 3N. Um and then 1118 00:40:19,838 --> 00:40:21,599 whatever actual inference engine you 1119 00:40:21,599 --> 00:40:23,760 need to run it. So this could be O Lama, 1120 00:40:23,760 --> 00:40:27,920 um Llama CBP, VLM, SG Lang, um including 1121 00:40:27,920 --> 00:40:29,280 whatever hardware you're running it on. 1122 00:40:29,280 --> 00:40:31,280 So this could be Apple Silicon, Nvidia, 1123 00:40:31,280 --> 00:40:33,199 whatever you need. um for actually 1124 00:40:33,199 --> 00:40:34,480 making all of these agents and 1125 00:40:34,480 --> 00:40:36,559 intelligence useful you need some set of 1126 00:40:36,559 --> 00:40:38,320 tools in memory that can be run through 1127 00:40:38,320 --> 00:40:40,960 a standard MCP protocol um and you need 1128 00:40:40,960 --> 00:40:42,880 some sort of uh set of primitives for 1129 00:40:42,880 --> 00:40:44,639 actually doing learning whether it's 1130 00:40:44,639 --> 00:40:46,239 prompt based techniques like Japa or 1131 00:40:46,239 --> 00:40:48,159 DSPI um whether it's weight based 1132 00:40:48,159 --> 00:40:51,280 techniques like gpo and sftt and Laura 1133 00:40:51,280 --> 00:40:52,559 um you need some way to actually get 1134 00:40:52,559 --> 00:40:54,400 this agent to improve over time and 1135 00:40:54,400 --> 00:40:55,599 actually be able to make it more 1136 00:40:55,599 --> 00:40:57,599 personal and more effective and so to 1137 00:40:57,599 --> 00:40:59,440 kind of walk through like what opens 1138 00:40:59,440 --> 00:41:01,760 looks like um we tried to go with all of 1139 00:41:01,760 --> 00:41:04,318 the standard um interfaces that people 1140 00:41:04,318 --> 00:41:06,000 are already accustomed to. Um so we 1141 00:41:06,000 --> 00:41:07,358 wanted to give people the ability to 1142 00:41:07,358 --> 00:41:09,199 interact with it through a desktop and 1143 00:41:09,199 --> 00:41:10,480 actually just run it as they would 1144 00:41:10,480 --> 00:41:12,400 normally expect, but then see all of the 1145 00:41:12,400 --> 00:41:13,760 savings that they're getting in terms of 1146 00:41:13,760 --> 00:41:16,480 dollars and energy. Um we also wanted to 1147 00:41:16,480 --> 00:41:18,880 give people the ability um to run 1148 00:41:18,880 --> 00:41:21,199 different kinds of continuous agents. So 1149 00:41:21,199 --> 00:41:22,719 different kinds of agents that are 1150 00:41:22,719 --> 00:41:24,639 persistent in terms of cron jobs and 1151 00:41:24,639 --> 00:41:26,639 being able to run standard protocols um 1152 00:41:26,639 --> 00:41:28,880 day after day. Um, basically we just 1153 00:41:28,880 --> 00:41:30,239 wanted to to make this like 1154 00:41:30,239 --> 00:41:32,159 plug-and-play with all of the workflows 1155 00:41:32,159 --> 00:41:33,760 that people are already accustomed to 1156 00:41:33,760 --> 00:41:35,760 running. Um, and we wanted to make this 1157 00:41:35,760 --> 00:41:37,760 something that can get people to have 1158 00:41:37,760 --> 00:41:39,760 their first experience with LMS on 1159 00:41:39,760 --> 00:41:41,358 device the same way people had their 1160 00:41:41,358 --> 00:41:43,679 first experience with ChatGpt or Claude 1161 00:41:43,679 --> 00:41:46,400 back in the day. Um, so yeah, and so 1162 00:41:46,400 --> 00:41:47,760 yeah, to step through a little quicker, 1163 00:41:47,760 --> 00:41:49,519 but yeah, here's like a nice way to like 1164 00:41:49,519 --> 00:41:52,639 set up new persistent jobs. Um, yeah, we 1165 00:41:52,639 --> 00:41:54,000 have all of these different components. 1166 00:41:54,000 --> 00:41:55,519 We need some way to actually optimize 1167 00:41:55,519 --> 00:41:57,358 it. And so we wanted to get out of the 1168 00:41:57,358 --> 00:41:59,760 way of the LM as much as possible by 1169 00:41:59,760 --> 00:42:01,838 just creating a simple spec of these 1170 00:42:01,838 --> 00:42:03,280 five primitives by which they could go 1171 00:42:03,280 --> 00:42:04,880 through the optimization. And what we 1172 00:42:04,880 --> 00:42:06,318 found is that by going through this 1173 00:42:06,318 --> 00:42:08,079 whole optimization loop, we were not 1174 00:42:08,079 --> 00:42:09,519 only able to get significant dollar 1175 00:42:09,519 --> 00:42:11,838 costs uh dollar cost reductions, but 1176 00:42:11,838 --> 00:42:14,079 also significant latency reduction and 1177 00:42:14,079 --> 00:42:16,559 significant um improvements to overall 1178 00:42:16,559 --> 00:42:18,719 quality on these tests. And so this this 1179 00:42:18,719 --> 00:42:20,480 configuration is meant to simplify down 1180 00:42:20,480 --> 00:42:22,079 to just the five main things that people 1181 00:42:22,079 --> 00:42:23,440 care about when they're building these 1182 00:42:23,440 --> 00:42:25,679 these LM uh harnesses. So the 1183 00:42:25,679 --> 00:42:27,440 intelligence, the engine, the actual 1184 00:42:27,440 --> 00:42:29,920 agentic logic around it, the tools or 1185 00:42:29,920 --> 00:42:31,920 learning systems required for running it 1186 00:42:31,920 --> 00:42:34,239 um and the whole optimization um for the 1187 00:42:34,239 --> 00:42:35,920 whole spec as a whole. And so something 1188 00:42:35,920 --> 00:42:37,199 that we thought could be interesting to 1189 00:42:37,199 --> 00:42:39,280 help bridge this gap between local and 1190 00:42:39,280 --> 00:42:41,599 cloud LMS is to actually have the cloud 1191 00:42:41,599 --> 00:42:43,920 LM go through and manual and uh and 1192 00:42:43,920 --> 00:42:46,400 automatically optimize the whole LM the 1193 00:42:46,400 --> 00:42:48,159 whole local stack. And so this is a nice 1194 00:42:48,159 --> 00:42:49,519 way of taking advantages of the 1195 00:42:49,519 --> 00:42:52,318 capabilities of cloud LM to diagnose 1196 00:42:52,318 --> 00:42:55,039 proposed changes and gate um to create 1197 00:42:55,039 --> 00:42:57,440 improved solutions for these local LMS 1198 00:42:57,440 --> 00:42:59,358 while not incurring the cost of those 1199 00:42:59,358 --> 00:43:01,920 cloud LMS when you actually deploy these 1200 00:43:01,920 --> 00:43:04,159 um local stacks at inference. And so 1201 00:43:04,159 --> 00:43:06,318 what we found is that these uh these 1202 00:43:06,318 --> 00:43:08,318 open Jarvis jobs that were these open 1203 00:43:08,318 --> 00:43:10,000 Jarvis um configurations that were 1204 00:43:10,000 --> 00:43:12,480 actually optimized by cloud LMS like 1205 00:43:12,480 --> 00:43:15,358 cloud or chat GPT um were much more 1206 00:43:15,358 --> 00:43:17,838 effective than uh local stacks that were 1207 00:43:17,838 --> 00:43:19,679 just deployed out of the box because you 1208 00:43:19,679 --> 00:43:21,599 could actually cater to the specific LMS 1209 00:43:21,599 --> 00:43:24,000 the specific harness uh that was needed 1210 00:43:24,000 --> 00:43:26,000 uh for for different kinds of workloads. 1211 00:43:26,000 --> 00:43:27,838 And what we found is that even with the 1212 00:43:27,838 --> 00:43:30,159 ondevice LMS of today, we can rival 1213 00:43:30,159 --> 00:43:32,239 cloud LMS on different workflows around 1214 00:43:32,239 --> 00:43:35,119 personal AI um personal use cases, 1215 00:43:35,119 --> 00:43:37,599 coding, agentic tasks. While there 1216 00:43:37,599 --> 00:43:40,159 remains like many tasks for which um 1217 00:43:40,159 --> 00:43:42,880 like local local size LMS are not enough 1218 00:43:42,880 --> 00:43:45,440 um the gap is surprisingly closing um 1219 00:43:45,440 --> 00:43:48,239 month after month um as these LMS become 1220 00:43:48,239 --> 00:43:50,719 better distilled, more effective and 1221 00:43:50,719 --> 00:43:52,960 also we get better accelerators um for 1222 00:43:52,960 --> 00:43:54,880 running them. And so even with LM of 1223 00:43:54,880 --> 00:43:57,838 today, we can get 800x lower lower 1224 00:43:57,838 --> 00:44:00,159 costum in terms of actually running them 1225 00:44:00,159 --> 00:44:01,838 as well as a significant reduction in 1226 00:44:01,838 --> 00:44:03,838 latency. Um what we also found is that 1227 00:44:03,838 --> 00:44:05,838 no matter which cloud LM that we cloud 1228 00:44:05,838 --> 00:44:08,318 LM we picked um it was useful um in 1229 00:44:08,318 --> 00:44:10,079 terms of optimizing the whole aentic 1230 00:44:10,079 --> 00:44:13,199 loop for these uh local local open 1231 00:44:13,199 --> 00:44:15,599 Jarvis configurations. Um we found that 1232 00:44:15,599 --> 00:44:18,318 um the Opus series, Opus 5 as well as 1233 00:44:18,318 --> 00:44:20,400 GBD 5.6 Soul were were naturally the 1234 00:44:20,400 --> 00:44:22,159 best. Um but was interesting to see is 1235 00:44:22,159 --> 00:44:23,599 that you could pick Gemini, you could 1236 00:44:23,599 --> 00:44:27,838 pick um other other uh larger um LM 1237 00:44:27,838 --> 00:44:30,239 families like Kimmy and GLM and use them 1238 00:44:30,239 --> 00:44:32,400 to optimize these local configurations 1239 00:44:32,400 --> 00:44:33,679 so that you could capture those 1240 00:44:33,679 --> 00:44:34,960 efficiency gains, capture those 1241 00:44:34,960 --> 00:44:37,119 performance gains um for local inference 1242 00:44:37,119 --> 00:44:38,800 later. What we also found is that the 1243 00:44:38,800 --> 00:44:40,960 whole open Jarvis harness was cheaper to 1244 00:44:40,960 --> 00:44:43,280 optimize than alternatives which might 1245 00:44:43,280 --> 00:44:45,039 might require more data or more LM 1246 00:44:45,039 --> 00:44:47,119 calls. Um we found that like this set of 1247 00:44:47,119 --> 00:44:49,519 specs um and this set of primitives um 1248 00:44:49,519 --> 00:44:51,679 was most effective for local LM settings 1249 00:44:51,679 --> 00:44:54,960 because it got um the whole optimization 1250 00:44:54,960 --> 00:44:57,119 loop and the whole um set of LM 1251 00:44:57,119 --> 00:44:59,199 abstractions out of the way of the cloud 1252 00:44:59,199 --> 00:45:01,280 LM to just optimize the whole system and 1253 00:45:01,280 --> 00:45:02,960 just make it make it really fast and 1254 00:45:02,960 --> 00:45:04,480 really effective. Uh looking forward, 1255 00:45:04,480 --> 00:45:05,760 we're excited to keep building out this 1256 00:45:05,760 --> 00:45:07,679 project. Uh we think in the very near 1257 00:45:07,679 --> 00:45:10,159 future you're going to see um a huge maj 1258 00:45:10,159 --> 00:45:12,639 a huge proportion maybe even a majority 1259 00:45:12,639 --> 00:45:15,039 of uh people's daily inference calls 1260 00:45:15,039 --> 00:45:17,599 going to local devices and on-prem uh 1261 00:45:17,599 --> 00:45:19,920 laptops or on-prem workstations as 1262 00:45:19,920 --> 00:45:21,280 opposed to the kind of standard of today 1263 00:45:21,280 --> 00:45:22,480 where everything's being pushed out to 1264 00:45:22,480 --> 00:45:24,318 the cloud. Um we think these trends are 1265 00:45:24,318 --> 00:45:25,358 only going to continue because the 1266 00:45:25,358 --> 00:45:26,800 accelerators keep getting better and the 1267 00:45:26,800 --> 00:45:28,719 LMS keep getting better. Um and so if 1268 00:45:28,719 --> 00:45:30,719 you're excited about um anything in the 1269 00:45:30,719 --> 00:45:33,199 stack, whether it's uh better local LMS, 1270 00:45:33,199 --> 00:45:34,800 better accelerators, better inference 1271 00:45:34,800 --> 00:45:37,199 engines uh for deploying uh beyond data 1272 00:45:37,199 --> 00:45:39,440 centers, um please reach out. Uh we'd be 1273 00:45:39,440 --> 00:45:41,440 excited to chat. Include a QR code of 1274 00:45:41,440 --> 00:45:43,599 the project. Um if if folks are around 1275 00:45:43,599 --> 00:45:45,039 here afterwards, we'd love to chat. 1276 00:45:45,039 --> 00:45:52,159 Thanks. [applause] 1277 00:45:52,159 --> 00:46:00,159 have Josh and Rean. 1278 00:46:01,679 --> 00:46:03,920 we're working on QM, which is YC's 1279 00:46:03,920 --> 00:46:05,599 open-source 1280 00:46:05,599 --> 00:46:08,318 uh agent harness for work. QM is one 1281 00:46:08,318 --> 00:46:10,800 system that uh gives every employee at 1282 00:46:10,800 --> 00:46:14,318 YC an open claw-like assistant uh that's 1283 00:46:14,318 --> 00:46:17,199 like fully customizable and available in 1284 00:46:17,199 --> 00:46:20,159 Slack or via web UI, which is uh what 1285 00:46:20,159 --> 00:46:23,119 we're looking at here. Um each person 1286 00:46:23,119 --> 00:46:26,400 works within QM in their own personal 1287 00:46:26,400 --> 00:46:30,480 context that has its own sandbox files 1288 00:46:30,480 --> 00:46:34,239 uh and crons and it can they can also 1289 00:46:34,239 --> 00:46:36,400 work with QM in a multiplayer setting 1290 00:46:36,400 --> 00:46:39,760 like a slack channel. People use QM for 1291 00:46:39,760 --> 00:46:43,599 a pretty broad range of things uh like a 1292 00:46:43,599 --> 00:46:45,679 lot of automations like email triage 1293 00:46:45,679 --> 00:46:48,159 like legal and finance workflows. It's 1294 00:46:48,159 --> 00:46:49,679 really good at editing documents and 1295 00:46:49,679 --> 00:46:51,358 pulling data out of our internal 1296 00:46:51,358 --> 00:46:54,079 database. Uh it can also spin up live 1297 00:46:54,079 --> 00:46:57,358 internal web apps and help with stuff 1298 00:46:57,358 --> 00:46:59,920 like planning events. Uh but it's 1299 00:46:59,920 --> 00:47:02,559 designed to be broadly helpful for the 1300 00:47:02,559 --> 00:47:04,719 range of tasks that someone might 1301 00:47:04,719 --> 00:47:07,920 encounter at YC uh on a day-to-day 1302 00:47:07,920 --> 00:47:09,760 basis. 1303 00:47:09,760 --> 00:47:13,199 So you might be wondering uh why we 1304 00:47:13,199 --> 00:47:16,880 built this and it's really the result of 1305 00:47:16,880 --> 00:47:19,599 a string of internal agent projects that 1306 00:47:19,599 --> 00:47:21,119 have kind of unwound throughout the 1307 00:47:21,119 --> 00:47:24,400 years and all of which were really 1308 00:47:24,400 --> 00:47:26,480 riding this tailwind of increasingly 1309 00:47:26,480 --> 00:47:29,519 capable models. Um 1310 00:47:29,519 --> 00:47:32,239 the first one we built was in like 1311 00:47:32,239 --> 00:47:34,960 January of 2025. We internally refer to 1312 00:47:34,960 --> 00:47:37,119 it as the quote unquote like general 1313 00:47:37,119 --> 00:47:39,358 agent. Uh, but it was pretty 1314 00:47:39,358 --> 00:47:40,800 straightforward. Just kind of a system 1315 00:47:40,800 --> 00:47:44,480 prompt with tools uh in a loop. It was 1316 00:47:44,480 --> 00:47:46,239 one sizefits-all. 1317 00:47:46,239 --> 00:47:47,838 Um, 1318 00:47:47,838 --> 00:47:49,440 sort of like everyone was talking to the 1319 00:47:49,440 --> 00:47:51,599 same thing. Uh, and it was pretty 1320 00:47:51,599 --> 00:47:53,039 straightforward architecturally, but 1321 00:47:53,039 --> 00:47:57,519 like still uh very or surprisingly good 1322 00:47:57,519 --> 00:48:00,719 at answering data questions. Uh, 1323 00:48:00,719 --> 00:48:02,400 interestingly like the scope of what the 1324 00:48:02,400 --> 00:48:05,199 general agent was good at just increased 1325 00:48:05,199 --> 00:48:06,960 I guess unsurprisingly as the under 1326 00:48:06,960 --> 00:48:09,199 underlying models got better. Uh, and we 1327 00:48:09,199 --> 00:48:10,960 eventually hooked it up to Slack. We 1328 00:48:10,960 --> 00:48:14,960 added crons uh and gave it a a few more 1329 00:48:14,960 --> 00:48:16,800 tools so that it could be more be 1330 00:48:16,800 --> 00:48:20,800 capable across more domains. Um, in June 1331 00:48:20,800 --> 00:48:23,920 of 2025, uh, 1332 00:48:23,920 --> 00:48:25,599 by then like a lot of our engineers 1333 00:48:25,599 --> 00:48:28,159 started using cloud code and codecs. Uh, 1334 00:48:28,159 --> 00:48:29,519 and we realized that you could pretty 1335 00:48:29,519 --> 00:48:32,719 easily run these in a VM. And then, uh, 1336 00:48:32,719 --> 00:48:34,880 we hooked that up to a Slack tag, which 1337 00:48:34,880 --> 00:48:36,480 was a pretty powerful medium for people 1338 00:48:36,480 --> 00:48:38,559 who just wanted to like run a one-off 1339 00:48:38,559 --> 00:48:41,199 code change. Uh, we also configure 1340 00:48:41,199 --> 00:48:43,760 configured it to run our CI pipelines 1341 00:48:43,760 --> 00:48:46,400 and then spin up dev environments for 1342 00:48:46,400 --> 00:48:49,599 testing. Uh, and so people could come in 1343 00:48:49,599 --> 00:48:51,358 like describe a bug or something they 1344 00:48:51,358 --> 00:48:54,000 wanted to see happen and the bot would 1345 00:48:54,000 --> 00:48:55,358 go off and actually solve it, which is 1346 00:48:55,358 --> 00:48:58,400 like a pretty powerful um thing for 1347 00:48:58,400 --> 00:49:01,599 someone who like maybe hadn't uh made a 1348 00:49:01,599 --> 00:49:03,599 code change before in their life even. 1349 00:49:03,599 --> 00:49:07,280 Uh but we also on top of that had a 1350 00:49:07,280 --> 00:49:09,519 small loop going where we would observe 1351 00:49:09,519 --> 00:49:12,318 sort of how the bot failed uh where it 1352 00:49:12,318 --> 00:49:14,159 went wrong and then update the 1353 00:49:14,159 --> 00:49:17,119 agents.mmd which was uh present in the 1354 00:49:17,119 --> 00:49:20,480 codebase at the time um to make sure 1355 00:49:20,480 --> 00:49:23,440 that the thing got better as as uh we 1356 00:49:23,440 --> 00:49:26,318 like observe the usage and so in January 1357 00:49:26,318 --> 00:49:28,159 of this year a lot of the partners 1358 00:49:28,159 --> 00:49:30,639 started using openclaw and one thing to 1359 00:49:30,639 --> 00:49:32,159 know about YC partners is that they're 1360 00:49:32,159 --> 00:49:33,760 incredibly busy 1361 00:49:33,760 --> 00:49:36,480 uh between like office hours uh they get 1362 00:49:36,480 --> 00:49:38,159 tons of inbound email they're always 1363 00:49:38,159 --> 00:49:40,639 reading applications so like any tools 1364 00:49:40,639 --> 00:49:42,400 that can give them additional leverage 1365 00:49:42,400 --> 00:49:45,920 are incredibly valuable to YC so open 1366 00:49:45,920 --> 00:49:47,838 cloud particular was useful because it 1367 00:49:47,838 --> 00:49:49,280 was the first agent that a lot of them 1368 00:49:49,280 --> 00:49:51,679 had used uh that had their own computer 1369 00:49:51,679 --> 00:49:53,920 that had its own computer and so this 1370 00:49:53,920 --> 00:49:55,838 made it like very customizable in a way 1371 00:49:55,838 --> 00:49:57,920 that the previous paradigm of agents was 1372 00:49:57,920 --> 00:50:00,559 not uh and it functioned almost like a 1373 00:50:00,559 --> 00:50:02,480 personal assistant 1374 00:50:02,480 --> 00:50:04,960 And so in April uh the question became 1375 00:50:04,960 --> 00:50:06,318 like could we provide this to every 1376 00:50:06,318 --> 00:50:10,000 employee at YC uh without uh like buying 1377 00:50:10,000 --> 00:50:12,239 everyone a Mac Mini effectively. And so 1378 00:50:12,239 --> 00:50:14,639 we ended up provisioning a fleet of like 1379 00:50:14,639 --> 00:50:17,838 50 plus uh Hermes agents that were 1380 00:50:17,838 --> 00:50:20,318 running in VMs. And these were 1381 00:50:20,318 --> 00:50:22,400 definitely pretty helpful but they 1382 00:50:22,400 --> 00:50:24,480 required a lot of configuring for people 1383 00:50:24,480 --> 00:50:27,280 to get value out of them. And it was 1384 00:50:27,280 --> 00:50:28,960 just like inherently kind of difficult 1385 00:50:28,960 --> 00:50:32,000 to manage this fleet. Um, it was sort of 1386 00:50:32,000 --> 00:50:34,400 like a whack-a-ole situation where I 1387 00:50:34,400 --> 00:50:36,318 would have to sort of like SSH into 1388 00:50:36,318 --> 00:50:38,880 these individual instances and fix them. 1389 00:50:38,880 --> 00:50:40,960 And so the follow-up question became 1390 00:50:40,960 --> 00:50:43,679 like we've got we've gotten a lot of 1391 00:50:43,679 --> 00:50:46,480 value out of these agentic systems. Uh, 1392 00:50:46,480 --> 00:50:48,480 like let's build something that tries to 1393 00:50:48,480 --> 00:50:50,159 address some of the downsides of running 1394 00:50:50,159 --> 00:50:52,318 this big fleet of Hermes agents. Uh, 1395 00:50:52,318 --> 00:50:53,440 while still maintaining the 1396 00:50:53,440 --> 00:50:56,318 personalizability and some of the like 1397 00:50:56,318 --> 00:50:57,440 the stuff that people were really 1398 00:50:57,440 --> 00:51:00,318 getting value out of. 1399 00:51:00,318 --> 00:51:06,559 And so, 1400 00:51:06,559 --> 00:51:08,159 there's a pretty clear trend from 1401 00:51:08,159 --> 00:51:09,519 whether you could see there from what 1402 00:51:09,519 --> 00:51:12,159 Josh was showing you. Basically, um, the 1403 00:51:12,159 --> 00:51:14,318 models are getting better exponentially. 1404 00:51:14,318 --> 00:51:16,239 Um, and we were starting to see just 1405 00:51:16,239 --> 00:51:18,239 increasingly impressive returns from 1406 00:51:18,239 --> 00:51:19,760 giving them more and more capabilities. 1407 00:51:19,760 --> 00:51:21,838 So, OpenCloud gives the agent its own 1408 00:51:21,838 --> 00:51:23,440 computer and we start to see really 1409 00:51:23,440 --> 00:51:25,519 impressive returns from that. So around 1410 00:51:25,519 --> 00:51:27,119 May this year, we started thinking just 1411 00:51:27,119 --> 00:51:28,800 like how far can we push this if we just 1412 00:51:28,800 --> 00:51:31,119 keep pulling on this threat. Um we 1413 00:51:31,119 --> 00:51:33,119 really like the the lens of sort of 1414 00:51:33,119 --> 00:51:35,039 unhobling. I don't know if you guys uh 1415 00:51:35,039 --> 00:51:36,719 read situational awareness when it came 1416 00:51:36,719 --> 00:51:38,960 out in like 2024. Um but that was kind 1417 00:51:38,960 --> 00:51:40,800 of an era when like test time compute 1418 00:51:40,800 --> 00:51:42,480 was just starting to become a thing and 1419 00:51:42,480 --> 00:51:43,440 like you know we're starting to give 1420 00:51:43,440 --> 00:51:45,358 agents tools for the first time and 1421 00:51:45,358 --> 00:51:47,358 there's sort of this intuition that what 1422 00:51:47,358 --> 00:51:49,519 agents can do like there's a little bit 1423 00:51:49,519 --> 00:51:50,880 there's kind of more intelligence in the 1424 00:51:50,880 --> 00:51:52,719 models than we're than we're using in a 1425 00:51:52,719 --> 00:51:54,719 lot of cases. Um, and it's like if we 1426 00:51:54,719 --> 00:51:57,039 really push uh the frontier in terms of 1427 00:51:57,039 --> 00:51:58,480 just like what we're offering up the 1428 00:51:58,480 --> 00:52:00,318 agents as as capabilities that they can 1429 00:52:00,318 --> 00:52:02,719 make use of uh like magic can start 1430 00:52:02,719 --> 00:52:06,159 happening. Um so the first way that we 1431 00:52:06,159 --> 00:52:09,519 do that with QM um is by essentially 1432 00:52:09,519 --> 00:52:12,159 pulling the brain of the system up out 1433 00:52:12,159 --> 00:52:15,039 of the sandbox. So with uh you know with 1434 00:52:15,039 --> 00:52:17,838 Hermes and with OpenClaw you effectively 1435 00:52:17,838 --> 00:52:20,719 have um the agent has its own computer 1436 00:52:20,719 --> 00:52:22,239 which is super powerful but it's also 1437 00:52:22,239 --> 00:52:24,400 trapped inside that computer. So that 1438 00:52:24,400 --> 00:52:26,719 causes a few issues just from you know 1439 00:52:26,719 --> 00:52:28,318 uh us trying to administer that system 1440 00:52:28,318 --> 00:52:30,318 when there were you know uh even a few 1441 00:52:30,318 --> 00:52:31,440 dozen of these things it starts to 1442 00:52:31,440 --> 00:52:33,920 become unwieldy almost immediately. Um 1443 00:52:33,920 --> 00:52:35,760 but the other issue with that is that 1444 00:52:35,760 --> 00:52:38,880 you um all all of the like all the 1445 00:52:38,880 --> 00:52:40,800 sessions that you would have uh they're 1446 00:52:40,800 --> 00:52:42,960 trapped inside that computer. And so 1447 00:52:42,960 --> 00:52:45,838 what we did instead is we uh we just 1448 00:52:45,838 --> 00:52:47,280 offload everything into Postgress. So 1449 00:52:47,280 --> 00:52:48,719 everything is centralized um from all 1450 00:52:48,719 --> 00:52:50,239 the agent conversations that people are 1451 00:52:50,239 --> 00:52:52,480 having. Um and then we expose those to 1452 00:52:52,480 --> 00:52:54,239 the agent itself. So it can look at all 1453 00:52:54,239 --> 00:52:55,838 the context that's sort of aggregating 1454 00:52:55,838 --> 00:52:57,760 from across the system. And then the 1455 00:52:57,760 --> 00:52:59,119 other thing we do is we start thinking 1456 00:52:59,119 --> 00:53:02,000 about sandboxes less as this home where 1457 00:53:02,000 --> 00:53:04,079 the uh where the agent lives and where 1458 00:53:04,079 --> 00:53:06,400 it's kind of stuck in a lot of ways. And 1459 00:53:06,400 --> 00:53:08,400 sandboxes become more of this thing uh 1460 00:53:08,400 --> 00:53:09,838 more of a resource that the agent can 1461 00:53:09,838 --> 00:53:12,639 dip into and use as needed. Um but it's 1462 00:53:12,639 --> 00:53:15,519 it's um it's a lot less limiting. Uh the 1463 00:53:15,519 --> 00:53:17,119 other thing that this starts to open up 1464 00:53:17,119 --> 00:53:19,280 um is this idea of you're accumulating 1465 00:53:19,280 --> 00:53:20,880 this large eval set of all the traces 1466 00:53:20,880 --> 00:53:22,079 that you have from the conversations 1467 00:53:22,079 --> 00:53:23,599 that people are having with the agent. 1468 00:53:23,599 --> 00:53:25,358 And in principle, you can think about 1469 00:53:25,358 --> 00:53:27,440 going and and hill climbing on that um 1470 00:53:27,440 --> 00:53:28,639 and sort of having this automated 1471 00:53:28,639 --> 00:53:30,719 improvement loop. Um we've had sort of 1472 00:53:30,719 --> 00:53:32,400 mixed results with that. I would say I 1473 00:53:32,400 --> 00:53:34,159 think typically if you're just 1474 00:53:34,159 --> 00:53:36,000 dispatching this like torrent of agents 1475 00:53:36,000 --> 00:53:38,480 that are supposed to um fix all of the 1476 00:53:38,480 --> 00:53:40,079 bugs that they're encountering when you 1477 00:53:40,079 --> 00:53:42,159 have the LLM as a judge, you start to 1478 00:53:42,159 --> 00:53:44,159 get this kind of uh like main character 1479 00:53:44,159 --> 00:53:46,480 syndrome where the agents are making 1480 00:53:46,480 --> 00:53:48,480 fixes that are, you know, only seeing 1481 00:53:48,480 --> 00:53:51,039 their their um their piece of the 1482 00:53:51,039 --> 00:53:52,239 elephant effectively. like they're 1483 00:53:52,239 --> 00:53:55,119 they're really um they can be sort of 1484 00:53:55,119 --> 00:53:57,199 yeah not not seeing the whole um whole 1485 00:53:57,199 --> 00:53:58,639 system and so having the human in the 1486 00:53:58,639 --> 00:53:59,920 loop there has continued to be really 1487 00:53:59,920 --> 00:54:01,358 important although we're really looking 1488 00:54:01,358 --> 00:54:04,000 forward to this uh working uh all the 1489 00:54:04,000 --> 00:54:06,239 way around. Um so the other major thing 1490 00:54:06,239 --> 00:54:08,318 we do that's pretty obvious is just wire 1491 00:54:08,318 --> 00:54:09,760 the agents to all of the resources 1492 00:54:09,760 --> 00:54:11,679 across the company that we can. Uh we 1493 00:54:11,679 --> 00:54:14,000 already happened to have a CLI at YC 1494 00:54:14,000 --> 00:54:16,800 that worked really well um that wired a 1495 00:54:16,800 --> 00:54:20,000 lot of systems together. Um uh but 1496 00:54:20,000 --> 00:54:21,920 anything that wasn't in there uh we 1497 00:54:21,920 --> 00:54:25,440 basically allow um adding just arbitrary 1498 00:54:25,440 --> 00:54:28,480 API keys um that sort of thing. And then 1499 00:54:28,480 --> 00:54:30,400 we also want to ensure that we have par 1500 00:54:30,400 --> 00:54:32,159 with uh just an employee working on 1501 00:54:32,159 --> 00:54:34,239 their laptop. So you know device code 1502 00:54:34,239 --> 00:54:35,760 OOTH 1503 00:54:35,760 --> 00:54:37,838 um we go ahead and ingest that into a 1504 00:54:37,838 --> 00:54:39,199 keychain and then refresh it for you. So 1505 00:54:39,199 --> 00:54:41,679 it's it's uh ideally supposed to imitate 1506 00:54:41,679 --> 00:54:43,199 the experience of a person on their 1507 00:54:43,199 --> 00:54:46,400 computer. We mostly keep this to to be 1508 00:54:46,400 --> 00:54:48,719 read only uh in the database but we do 1509 00:54:48,719 --> 00:54:52,079 allow for rights um via uh human 1510 00:54:52,079 --> 00:54:54,159 reviewed bulk upserts. So the way that 1511 00:54:54,159 --> 00:54:55,838 works is the agent will put forward a 1512 00:54:55,838 --> 00:54:59,039 plan to um edit the database um that a 1513 00:54:59,039 --> 00:55:00,318 person can kind of give a once over and 1514 00:55:00,318 --> 00:55:01,519 ensure it's not doing anything crazy 1515 00:55:01,519 --> 00:55:04,000 before the write actually happens. Um 1516 00:55:04,000 --> 00:55:05,920 one thing we've observed with this is 1517 00:55:05,920 --> 00:55:07,679 that we've started just kind of rubber 1518 00:55:07,679 --> 00:55:09,679 stamping these. It's a little bit like 1519 00:55:09,679 --> 00:55:11,760 um I think if you guys use cloud code in 1520 00:55:11,760 --> 00:55:12,960 the early days like you might have been 1521 00:55:12,960 --> 00:55:15,119 reviewing the tool uses very closely and 1522 00:55:15,119 --> 00:55:17,440 eventually um you sort of build up more 1523 00:55:17,440 --> 00:55:18,880 trust in the agent. So this is something 1524 00:55:18,880 --> 00:55:20,480 that we're uh looking at very closely 1525 00:55:20,480 --> 00:55:22,079 over the over the next few months. Yeah. 1526 00:55:22,079 --> 00:55:24,639 So sort of like I was saying um the 1527 00:55:24,639 --> 00:55:26,800 sandboxes in this system we like to 1528 00:55:26,800 --> 00:55:29,119 think of as a resource for the agent. So 1529 00:55:29,119 --> 00:55:30,639 and and less where the agent actually 1530 00:55:30,639 --> 00:55:35,358 lives. Um so in QM the agent can um 1531 00:55:35,358 --> 00:55:37,838 basically uh by default it's going to be 1532 00:55:37,838 --> 00:55:39,679 using a particular sandbox. So that's 1533 00:55:39,679 --> 00:55:41,280 going to be one that's been allocated to 1534 00:55:41,280 --> 00:55:43,358 the user that it's talking to. But in 1535 00:55:43,358 --> 00:55:45,599 general uh there are there are 1536 00:55:45,599 --> 00:55:47,280 environments that the agent can kind of 1537 00:55:47,280 --> 00:55:49,599 converge on and uh and can collaborate 1538 00:55:49,599 --> 00:55:52,559 with. Um, and then the other uh key 1539 00:55:52,559 --> 00:55:56,079 thing is that um if the agent is working 1540 00:55:56,079 --> 00:55:58,559 on like a um like a heavier dev 1541 00:55:58,559 --> 00:55:59,838 workload, it can go and reach for a 1542 00:55:59,838 --> 00:56:01,920 machine that has more resources. Um if 1543 00:56:01,920 --> 00:56:02,960 it's working on something that's 1544 00:56:02,960 --> 00:56:04,639 simpler, it'll just go for a sandbox 1545 00:56:04,639 --> 00:56:06,719 that's that's less powerful. Um and so 1546 00:56:06,719 --> 00:56:08,159 pushing that decision into the agent 1547 00:56:08,159 --> 00:56:09,679 itself rather than the harness has been 1548 00:56:09,679 --> 00:56:12,239 a really um really powerful thing. 1549 00:56:12,239 --> 00:56:15,679 Similarly, um allowing the agent to uh 1550 00:56:15,679 --> 00:56:17,760 tap into its own runtime. So basically, 1551 00:56:17,760 --> 00:56:20,239 if it can choose the provider um that 1552 00:56:20,239 --> 00:56:22,000 it's working with, you start to get out 1553 00:56:22,000 --> 00:56:24,159 of situations like um I'm sure you guys 1554 00:56:24,159 --> 00:56:25,760 have run into this with Fable. If you 1555 00:56:25,760 --> 00:56:27,599 try to do AI research, if you try to do 1556 00:56:27,599 --> 00:56:29,760 cyber security, anything, you'll get a 1557 00:56:29,760 --> 00:56:31,599 bunch of refusals. Um so what we can do 1558 00:56:31,599 --> 00:56:32,798 when the agent can control it on 1559 00:56:32,798 --> 00:56:34,159 runtime, it can just pop out into 1560 00:56:34,159 --> 00:56:35,679 another model uh when it needs to avoid 1561 00:56:35,679 --> 00:56:39,440 a situation like that. Um similarly, uh 1562 00:56:39,440 --> 00:56:41,599 like in the earlier situation, um it's 1563 00:56:41,599 --> 00:56:43,920 often useful to pop between different 1564 00:56:43,920 --> 00:56:46,000 sandbox providers. Um and so that's 1565 00:56:46,000 --> 00:56:47,920 something we can do quite easily. And as 1566 00:56:47,920 --> 00:56:49,679 a general rule, what we've tried to do 1567 00:56:49,679 --> 00:56:52,400 is keep the harness extremely thin. Uh 1568 00:56:52,400 --> 00:56:54,000 we think of the kind of core of the 1569 00:56:54,000 --> 00:56:56,480 system as being these three tools um 1570 00:56:56,480 --> 00:56:58,400 where you have execution in a remote 1571 00:56:58,400 --> 00:57:00,480 sandbox, reading and writing from object 1572 00:57:00,480 --> 00:57:02,719 storage and then uh publishing internal 1573 00:57:02,719 --> 00:57:05,119 apps, a pretty simple sort of getbacked 1574 00:57:05,119 --> 00:57:07,519 uh system. And then we have other tools 1575 00:57:07,519 --> 00:57:09,838 for interacting with memory and crons 1576 00:57:09,838 --> 00:57:11,119 and that sort of thing. But we really 1577 00:57:11,119 --> 00:57:14,318 think of these as sort of temporary um 1578 00:57:14,318 --> 00:57:16,400 papering over rough edges in the system. 1579 00:57:16,400 --> 00:57:18,318 And really the core of it is is these 1580 00:57:18,318 --> 00:57:20,719 three up here. Um it's we we really try 1581 00:57:20,719 --> 00:57:22,480 to keep it as small as we can. Yeah, 1582 00:57:22,480 --> 00:57:23,760 we're we're sort of trying to be this 1583 00:57:23,760 --> 00:57:26,639 like uh AGI anticipating harness. 1584 00:57:26,639 --> 00:57:29,039 Although uh since we aren't there yet, 1585 00:57:29,039 --> 00:57:31,039 um there are a few things that we've run 1586 00:57:31,039 --> 00:57:34,480 into. Um, one of these is that the 1587 00:57:34,480 --> 00:57:36,239 agents have been we've tried to put them 1588 00:57:36,239 --> 00:57:37,920 in this really capable environment where 1589 00:57:37,920 --> 00:57:39,599 they have all these uh tools available 1590 00:57:39,599 --> 00:57:42,400 to them, but um they often give up way 1591 00:57:42,400 --> 00:57:44,480 too early. Uh so one thing we've 1592 00:57:44,480 --> 00:57:45,838 experimented with especially over the 1593 00:57:45,838 --> 00:57:48,239 past month or so has been setting uh we 1594 00:57:48,239 --> 00:57:50,880 call it like a grind tool or um 1595 00:57:50,880 --> 00:57:53,838 basically we set budgets on goals. So 1596 00:57:53,838 --> 00:57:55,838 the agent is not allowed to give up on 1597 00:57:55,838 --> 00:57:57,679 its task before a certain amount of like 1598 00:57:57,679 --> 00:57:59,039 walk clock time. just like a couple 1599 00:57:59,039 --> 00:58:01,039 hours uh or a certain amount of token 1600 00:58:01,039 --> 00:58:03,920 spend. And so um what that can 1601 00:58:03,920 --> 00:58:06,239 accomplish is just like uh really a lot 1602 00:58:06,239 --> 00:58:08,480 better, you know, uh research outputs, 1603 00:58:08,480 --> 00:58:10,480 better reports, that sort of thing. Um 1604 00:58:10,480 --> 00:58:12,639 and it's been really fun actually to see 1605 00:58:12,639 --> 00:58:15,199 uh like OpenAI and anthropic um you 1606 00:58:15,199 --> 00:58:16,798 know, crack some open problems in math 1607 00:58:16,798 --> 00:58:18,400 with like a very similar technique. Uh 1608 00:58:18,400 --> 00:58:20,000 but it also works for just you know, 1609 00:58:20,000 --> 00:58:22,798 normal office work stuff too. Um the 1610 00:58:22,798 --> 00:58:24,960 other thing that we've seen a lot of is 1611 00:58:24,960 --> 00:58:26,719 so this harness is supposed to work uh 1612 00:58:26,719 --> 00:58:28,159 it works in multiplayer. It works in 1613 00:58:28,159 --> 00:58:30,480 Slack. Um, but because of the the 1614 00:58:30,480 --> 00:58:32,159 artifacts of its of its training, 1615 00:58:32,159 --> 00:58:35,599 effectively uh what we see is that uh 1616 00:58:35,599 --> 00:58:37,358 the agent gets can get very confused 1617 00:58:37,358 --> 00:58:39,280 about the situation that it's in even if 1618 00:58:39,280 --> 00:58:40,798 we specify this pretty clearly in the 1619 00:58:40,798 --> 00:58:42,639 system prompt. So having like local 1620 00:58:42,639 --> 00:58:44,079 affordances for this has been something 1621 00:58:44,079 --> 00:58:45,599 that's been uh that's been really 1622 00:58:45,599 --> 00:58:53,599 important. 1623 00:58:53,838 --> 00:58:55,440 been a problem is that agents really 1624 00:58:55,440 --> 00:58:59,760 don't uh understand social contexts. Um 1625 00:58:59,760 --> 00:59:01,199 to make this a little more concrete, 1626 00:59:01,199 --> 00:59:03,199 like if I tell Regan a piece of 1627 00:59:03,199 --> 00:59:05,920 information, uh he intuitively sort of 1628 00:59:05,920 --> 00:59:08,400 knows uh or at least has like a good 1629 00:59:08,400 --> 00:59:10,559 mental framework of where it is okay to 1630 00:59:10,559 --> 00:59:13,519 share that information. Uh but it takes 1631 00:59:13,519 --> 00:59:15,838 some actual work to recreate this with 1632 00:59:15,838 --> 00:59:18,318 an agent. uh like privileged information 1633 00:59:18,318 --> 00:59:20,079 can very easily just leak into these 1634 00:59:20,079 --> 00:59:24,400 contexts where it should not be and so 1635 00:59:24,400 --> 00:59:26,000 uh the information that you can put in 1636 00:59:26,000 --> 00:59:28,000 the brain is effectively like bounded by 1637 00:59:28,000 --> 00:59:30,400 how good your permission system is uh 1638 00:59:30,400 --> 00:59:33,358 and so YC luckily has an existing 1639 00:59:33,358 --> 00:59:35,599 software system with like fine grained 1640 00:59:35,599 --> 00:59:37,599 permissioning uh that has been built 1641 00:59:37,599 --> 00:59:39,440 over the years but a lot of people just 1642 00:59:39,440 --> 00:59:42,960 don't have that uh and so it takes uh 1643 00:59:42,960 --> 00:59:45,440 work to allow for knowledge sharing in 1644 00:59:45,440 --> 00:59:48,159 an in nuanced Okay, so thanks everybody 1645 00:59:48,159 --> 00:59:51,119 for listening. Um, you can try out QM. 1646 00:59:51,119 --> 00:59:54,318 It's open source. Uh, and coding agents 1647 00:59:54,318 --> 00:59:56,079 are pretty good at standing it up. If 1648 00:59:56,079 --> 00:59:57,519 you run into any problems, feel free to 1649 00:59:57,519 --> 00:59:59,280 put up an issue and we'll look at it. 1650 00:59:59,280 --> 01:00:01,760 Uh, we're also hiring. So if any of this 1651 01:00:01,760 --> 01:00:03,519 resonated with you or would be exciting, 1652 01:00:03,519 --> 01:00:07,780 then feel free to send us an email. 1653 01:00:07,780 --> 01:00:10,780 [applause]