I Burned 30 Billion Tokens on These Puzzles and It Changed How I Look at AI
BDBen Davis@bmdavis419 · 45.9K subscribers
- Views
- 6.3K
- 6,302
- Outlier
- ×0.3
- Median 21.4K
- Length
- 21:41
- Published
- Aug 17
- Monday
- Comments
- 21
- 239 likes
Script
How it's written
- Words
- 4,773
- Words a minute
- 220
- Reading time
- 21 min
- Segments
- 15
The hook
0:00–0:25I just got back from spending a week down in Vegas. I was there for Defcon, which is a big security hacker conference thing. It was the first year I've ever gone. It was really, really fun. But I didn't actually spend any time doing Vegas things or even going to the conference itself. I spent the entire time sitting in a hotel room solving some of the hardest puzzles I have ever seen. And I want to talk about them because it was a very cool experience for me as a guy who doesn't
Structure
I just got back from spending a week down in Vegas. I was there for Defcon, which is a big security hacker conference thing. It was the first year I've ever gone. It was really, really fun. But I didn't actually spend any time doing Vegas things or even going to the conference itself. I spent the entire time sitting in a hotel room solving some of the hardest puzzles I have ever seen. And I want to talk about them because it was a very cool experience for me as a guy who doesn't know a huge amount about puzzles and cryptography and security and all that stuff, but just does a lot of stuff with AI and software in general. And I was able to actually make myself useful to the team. And we did win the competition. The competition we were doing is called Goldbug. It is a set of 13 really difficult puzzles that you need to solve over the course of I think it was like 3 days or something like that. Obviously, going into this, my plan was to just have a swarm of agents attack these puzzles the moment they came online and solve as many as humanly possible. I figured there might be one or two, which gave the agent some trouble, but with just like a little bit of steering and extra context, they would be able to get the answer. But that was not even close to what actually happened. The agents I was running against these problems only ended up solving three out of the 13. everything else they just could not solve and we had to do a lot of work ourselves in order to solve these things. Now, the AI did end up being really, really useful and that's what I want to talk about in this video. The ways I was using it, the way I think that the puzzle makers were able to thwart the AI in its current state, a lot of the weird behaviors I noticed from the models as I was using
Watch from 0:00them, especially on these weird longunning puzzly type tasks, I saw things that I don't normally see when doing normal coding work. And I think there's a lot of reasons for that. And also just kind of talk about the puzzles because there were some really dope ones here. Like this one was just a dress that this is all the context we were given is this dress and we had to get a message out of that. There was also one that was just a set of Rubik's cubes arranged like this that we had to again decode a message from. Now I want to start by talking about how I think the puzzle creators were able to beat the AI agents. The easiest way to do this I think is to start with one a problem that the AI could solve easily. And this was one of the first ones we got access to. It didn't actually instantly solve it, but it was able to solve it a little bit later. The reason why is because on the first day, there was a slight error in the puzzle on like the 10th step or something like that. So, it wasn't instantly solvable right off the rip, but as soon as it was solvable, the AI was able to pretty easily rip through the entire thing. And the whole premise of this puzzle is you're given this picture, some text on here, and you need to work your way through this manner. And the way you work your way through is by going deeper and deeper into the URL. So the solution for this page is some camelc case set of words that you then append to the URL up here and then that will take you to the next page and you just keep going over and over and over and over again until you hit the very end. And this puzzle is shaped in such a way that agents really really like it.
Watch from 1:31All of the resources are available digitally. It is all running over like web stuff. It's using images that are pretty easy to parse. It can do as much research as it wants. It can run for as long as it wants. Like this is a pretty easy problem for especially like coding agents to just solve. It takes a while. It took like an hour and a half for it to get the actual solution. But again, there wasn't too many issues. It didn't get caught up in too many random things because there's just there's not too much here for the model. Lately, I've been building so much more, making so many more PRs to the point where CI is actually becoming a real bottleneck for me. GitHub just isn't cutting it, which is why today's sponsor, Blacksmith, makes so much sense. It is a oneline change to your GitHub actions that results in you getting 2x faster hardware, 4x faster cache downloads, and 40x faster docker builds. It is insane. There are humongous companies like Mercury, Subbase, Ashb, Clerk, and many more using Blacksmith for a reason. Actually, including T3 Code. All of T3 Codes builds go through Blacksmith because they are just so cheap, so fast, so good. I I don't really know what else to say. If you're working on any even remotely serious project, you're going to have a ton of CI. And if you have CI, you're just burning money by not using Blacksmith. And outside of GitHub actions, the new Codesmith background coding agent is really good. You can set it up in Slack or you can also just tag it in from GitHub or the web UI. And once you have it set up, it can help with planning, implementing, optimizing code, also autofixing PRs, which is incredibly useful. the amount of like autonomous loops I have running for PRs at this point where I will have CI run some random thing fails is really useful
Watch from 2:53to just have an agent instantly be spun off without me having to do anything make the required fix and just update the PR accordingly so that by the time I go look at it everything is already exactly where I want it to be. Blacksmith will improve literally everything about your GitHub actions and CI experience at dav7.link/blacksmith. But then when you go over to a problem like this, this is one of the problems that the AI just completely fell apart on. I think I had like six different agents running at once trying to solve this thing. Three different codec instances over three different machines. I had a cursor cloud run going, a couple of Fable runs going, just trying literally everything I could to try and get these things to be able to solve it. And there's a lot of reasons it was tough. One of the biggest ones was just the fact that since this was a physical in-person puzzle, the only way you could collect the data from it was by either looking at it there or looking at pictures and videos of it. We got really good pictures and videos, but it was still just really hard for the model to reconstruct the data set that was required here. Cuz what you had to do is take this belt down here and then turn that into a key, an A through because there's 26 little beads on the belt. Each one is a different unique color. And you can go A through Z labeling each of these different colors with a letter. Then once you have this set of labels, you can then label the grid on the chest. You can then label the front and the back part of the dress cuz there was uh 2x6 grids on each side. You merge those together to create this next grid.
Watch from 4:25Then you end up with this data set and you can actually go through and do the real solve. But getting to that point, the models could not do it. Some of them got a lot closer than I was expecting. But even if they did get a decent data set, there was just such insane amounts of data overload that these things could not figure it out. And to be clear, we were struggling with this, too. We spent a long time trying to figure out like, okay, does the number of beads matter? Does the shape of the beads matter? Does whether or not it's transparent matter? Like, it did just end up being color, but it took us quite a while to figure that out. There's so many rabbit holes you can go down using this puzzle as a starting point that all of the agents just did not have any chance of actually solving this without us doing this leg work for it. I spent most of the time working on this puzzle with Thor pirate software. Awesome guy, insanely talented at like cyber security type things and puzzles and ciphers and all that stuff that I have never really seen before. The thing that Thor was able to figure out was the standard columner transposition decryption. there were those like weird beads just floating on the belt from earlier. Like you can see here and here, those beads ended up mattering for this right here. Uh once we had that and we had these two tables, you could like take the bottom table, I believe, and the top table and basically take the rows that overlapped and remove them from the table to get a new set of letters. Run this decryption over it.
Watch from 5:52Read it horizontally. Answer X inconsistent knots. And the answer was inconsistent knots. We had to spend so much time labeling and organizing the data and then also figuring out what paths are worth going down and what paths aren't worth going down. It was a lot of what we were doing there where like Thor and Gern who were the really talented like solving guys. I was just kind of sitting there in the corner listening to them talk trying to figure out what paths they wanted to go down and then furiously typing into the agents to just have them explore all the things they're doing. I also spent a lot of time building out like visualizations for them cuz on the Rubik's cube puzzle, this ended up being really clutch because this is all we had. We had to figure out how to solve this puzzle off of these 12 cubes and black stands, two gold stands for some reason, then more black stands. I was not a huge part of solving this one outside of making the visualizer. But my general understanding of how it works is in order to solve this, you need to first solve these two cubes because they're scrambled in basically the same way, just inverted. Once you have the moves to solve those two cubes, it's like eight moves, I believe. If you apply those same eight moves to all of the rest of the cubes on here, that creates this new view. So, these two are solved. But if you look at this angle right here, it's kind of an awkward angle, but there's kind of an L here, an A here, a C here, a K here, and this all translates down to lack of dust. And that was the final answer.
Watch from 7:14This took the guys so long to figure out because it's just a brutal, weird puzzle. The idea of an agent being able to figure this out is hilarious. It requires so many things that they are just not that good at the 3D spatial reasoning, the visualization cuz like you really need to visualize this to see it. If you just had a set of data arranged in a grid, maybe they could figure it out, but like if this is all you're looking at, you're just looking at yb like this crap. really hard to see a letter in that as compared to this. And once again, this puzzle has the same problem the dress one did where there's just so much data involved and it's so hard for the model to actually figure out what it's supposed to be doing. There are infinite numbers of rabbit holes to go down and the agents did go down rabbit holes. They ran for hours at a time. They really tried their hardest to come up with solutions. There were two big things that really jumped out at me. The first of which was that one, hallucinations are not dead. When you're just using these things for coding work, it feels like hallucinations are not entirely solved, but more solved than they have been in the past. And they're at the point where like I'm really not too worried about them. If it can actually test its work against the real world, it'll probably do an okay job and do a good job of going and referencing documentation. The new models are trained really well to go fetch data from reality and then produce answers based on that. These problems have no ties to reality whatsoever other than at the very end when you submit your final answer. The only way you can check that you're going down the right path is when you get to the very end and you submit
Watch from 8:37it and you get a yes or a no. That's it. As a result of that, it was really easy for the models to be going down this gigantic run. They're a 100 compactions deep. Dozens of sub agents have fired. They've been trying a bunch of stuff and just somewhere somehow some kind of stupid leap in logic happens. Often what I noticed it would end up being is the model would do some test like it would try decrypting this in some weird way and that decryption would yield a phrase that was one letter off of making sense. You would just end up with like something that looks like the sky is and then instead of blue it would be like glue like the sky is glue and it would be like oh that could be a reference to this place where glue kind of makes sense that's trying to prod us into realizing that it should actually be blue. So then it should be the sky is blue and then it would take that as fact and that fact would then poison it for the rest of the run and it would just keep going on and on and on and spinning down stupid rabbit holes that would produce pretty much nothing of value. You really had to babysit these things all the way through or it just wouldn't get solved. But the other part of that was honestly even more fascinating to me was that as I was watching this happen across multiple agents on multiple devices, I noticed a recurring pattern where over and over again they would somehow independently naturally poison themselves in the same exact way and end up at the same exact wrong answers. It took a while. It took many hours and very long runs. But after enough compactions, enough sub aents, enough grinding, even with different starting prompts and all that stuff, just working
Watch from 10:08on the same puzzles going towards the same goal, they would eventually converge on the same exact wrong answer. Very fascinating to watch the probability machines, even at their gigantic current size, eventually converge on the thing, regress to the mean, and not get the right answer. This really made me rethink a lot of things in the AI world right now. You know, we in the dev space like it feels like this stuff is completely inevitable and is taking over everything and is scaling really fast. It is like we are about to get the god AI and in a lot of tasks it does feel like this. Like I could definitely see a world where like that convergence happened because you know GPT 5.6 SAL is only so big. The thing is maybe if you make the model bigger, maybe if you just make GPT56 10 times larger, make it a 50 trill per model, there will just be so many pathways within the neural net that it can go down to where it starts to actually create these novel pathways and novel solutions and can actually keep something this complex in its head and figure it out. That is totally in the realm of possibility and it's an untested assumption by now. The things we've seen so far show that it is heading in that direction. I'm seeing first signs of that and I could believe that. I am very confident that just like reinforcement learning like RL the thing that they're doing to make these things insanely good at like writing code that is not going to be enough to have it be able to solve stuff like this on its own. There's just no universe without scaling aka just making the models bigger working the way the labs think it will. I don't really see a world where anything other than it's an insanely
Watch from 11:38useful tool that you have to direct because it can't direct itself. I can't see any world where we are just stuck there forever. Again, they're all trying to scale. I know especially Anthropic is really going hard on this and I would not be surprised if by the end of the year they're able to solve more of these. This is actually a benchmark I'm going to keep so that whenever new models drop, I got to throw these puzzles at them and see how much further they can get. The way I tried to solve these things was with the agent swarms pattern or whatever. It's the thing that uh OpenAI has been doing internally for a lot of their early model testing. Like all those math proofs that they put out, those were done with a swarm of a new model. The hugging face incident where they had the agent break out of containment, that was an agent swarm. It was dozens of instances of the same model running in parallel, doing their own things, communicating with each other over a message board that ended up allowing them to solve and achieve these really big crazy things. And I figured that these problems would probably benefit a lot from that system. So I was trying that on all of these. It really wasn't enough. Like it does kind of feel like 5.6 is just right at the cusp of the generation where we can start doing swarms, but it's not quite there yet. I would bet that Astra will be substantially better at doing these. This will be a fun thing to test out in a month or two, but for now at least, the swarms really didn't accomplish all that much other than just burning a lot of tokens for no real reason. I uh mentioned this in the article write up I did on this. I just posted on my Substack. It's not behind the subscription wall. It's totally free.
Watch from 13:09I'll have it linked down below if you want to see that. It's a more uh put together version of these thoughts and I have included in here the breakdown of all the tokens I used over the course of this weekend. You can see the vast majority was through codeex which just means it was GBT56. I used a decent amount of cloud code like 1.3 bill pi was me screwing around with some open source models to see if those could maybe be useful and spoiler they just weren't like Fable and 5.6 were the only models that were really even remotely useful for this unfortunately. But you'll also notice looking at this table, the API equivalent cost of $24,000 worth of tokens that I burned over the course of just Defcon. And if you're not deep in the AI space, that probably seems completely ludicrous and insane. Like, why the hell did I spend 24 grand on this? And the answer is that I didn't. I did not spend 24 grand on this. I spent a total of $400 on this. 200 of that went to a Codeex sub. 200 of that went to a second codec sub just because I was burning my rate limits so hard especially on I think this was Friday. Was really trying the swarms thing out and just needed more tokens. So got the second one so I could get a reset. Um ended up canceling it afterward cuz generally for day-to-day work I only need one. And I guess actually I was wrong. It was 600 because there was another $200 sub that was my Claude Code sub. These subscriptions from the Frontier Labs are so heavily subsidized. Yeah, on paper it's 24 grand, but like you can get I think it's something like 10 to 15 grand worth of inference per month off of your $200 sub on OpenAI. It's absurd how much you can get from these things. But you also have
Watch from 14:36to remember that just because the API cost is that does not mean that the model actually costs them that much to run. Like I'm pretty sure their margins are really really good on these models. They just make them really expensive because it's supply and demand. for now. Over time, I'm hoping that stuff like SpaceX, like the new Grock model is really good. The Chinese labs are really catching up. Like hopefully that competition and choice is going to help drive these prices down over time. Again, it's it's not really 24 grand. It's a lot less than that. And to finish the video, I want to end with something more personal and optimistic. This was a very, very fun trip for me. I had a phenomenal time. And if anyone from the team is watching this, thank you for everything. I loved it. And I'm very much looking forward to next year. It was really, really cool to just like be there, learn from you guys, do everything I could to help. I think the team just really worked well because I was there doing all of the AI stuff, trying to help them, and then they weren't as deep on the AI stuff and didn't have the crazy agent setups, but they knew how to solve the puzzles. So, just like working together, building this stuff together. It was an awesome experience, and it also did a lot to kind of help with the uh, for lack of better term, AI psychosis. I think that this is kind of just a right of passage that you have to go through at this point. It is useful to have psychosis for at least a month or two. Like you need to go through the moment where you realize, oh it is all over. Like writing code by hand is kind of dumb at this point. And yes, these things actually are insanely useful and are going to do fundamentally mind-boggling
Watch from 16:08things over the next decades. Like this stuff is real and is not crypto 2.0. Is it a bubble? You can debate that all day long. Maybe in the short term it is, but in the long term I don't think it is. I think the long term of these things is 1,000% in the trillions, if not silly numbers that don't even mean anything beyond that. This stuff is real, but it is not there yet. And I really got to see leaving my field of expertise, which is software and AI stuff. That's the stuff I know really well. The AI seems and feels inevitable and superhuman to us because in our domain it really in a lot of ways is part of the reason for that is because one the primary thing these labs are working on right now is training and RLing for coding agent environments. Like they are training these things to be incredibly good at writing code working in systems because it is the prerequisite to doing everything else. All of the robotic stuff, all of even like the bio stuff and the legal stuff and all of these things require some semblance of code because code is the way the model interacts with the world. It can only produce text and that text can be code and that code is instructions that can cause something useful to actually happen. So this is the first thing that is going to get knocked down and it feels insane in this world. There's still a lot of limitations obviously like I have seen many many times where it will confidently assert something stupid within code. it will fall over repeatedly. But if you are actively driving it and paying attention to what it's doing and reading the outputs and reading the thinking traces, you can steer these things to do incredible things. And everything I just said there
Watch from 17:40about reading what it's doing, giving it good plans, giving it good direction, understanding the problem you're trying to have it solve. That's what we did here. That's how this ended up working. We just bootstrapped this together with multiple people. I alone couldn't understand these problems, but they alone didn't understand the AIS well enough to do that. But together we were able to do the thing that a lot of us already do with ourselves, agents and code is working directly with it and understanding the problems it's trying to solve. When you have a human operator who actually understands the domain and problem space and use these things for such useful stuff. Building out this visualizer was entirely just me telling Codeex to go make this, giving it some screenshots and data. Gave me a first pass. I was like, "Ah, this isn't great. Fix this, this, and this." Then I showed it to the team and they're like, "Ah, this is okay, but we need this, this, and this." Send another prompt over, get it iterated on over and over. And within an hour, we had this visualization, which was very helpful for solving the thing. Then when we needed to get to this stage, we created this visualizer almost instantly. Then we were able to get the solution off that. Like that was really useful. Same thing with the dress here, building out this web page to visualize and see and label the data. Incredible, incredible stuff. And once we had more data and context and theories, all of those could quickly be iterated on and tested by the agents. They're very powerful systems that just have to be steered by someone competent.
Watch from 19:10And unless crossing the threshold from 5 to 10 trillion per models, which is probably what mythos is, up to 20 trill models is enough to magically induce discernment and consciousness, which I don't think it is. I'm pretty sure it's a lot more parameters than that. But as long as that doesn't happen, we are definitely going to live in a world for a very long time where our skills still matter an insane amount, arguably even more, because the amount of leverage and output you can get just from knowing your now is so high. It is so cool the things we can do. I think the thing that I am personally focused the most on right now with this stuff is being as ambitious as possible with what I'm trying to do while also being reasonable and more balanced in the way I do these things. Like you don't need to run the agents 24/7. Run them when you need them. No more, no less. Having personally pushed these things to the absolute limits of their capabilities over the last week, I can pretty confidently say after watching them fall on their faces repeatedly, we definitely have time. So, I hopefully that this video turns into something semicoherent and useful. I was kind of just ranting through everything that happened. I really enjoyed this. If you did too, make sure to like and subscribe and I will talk to you again
Watch from 20:29Full transcript
- 0:00
I just got back from spending a week down in Vegas. I was there for Defcon, which is a big security hacker conference thing. It was the first year I've ever gone. It was really, really fun. But I didn't actually spend any time doing Vegas things or even going to the conference itself. I spent the entire time sitting in a hotel room solving some of the hardest puzzles I have ever seen. And I want to talk about
- 0:21
them because it was a very cool experience for me as a guy who doesn't know a huge amount about puzzles and cryptography and security and all that stuff, but just does a lot of stuff with AI and software in general. And I was able to actually make myself useful to the team. And we did win the competition. The competition we were doing is called Goldbug. It is a set of
- 0:41
13 really difficult puzzles that you need to solve over the course of I think it was like 3 days or something like that. Obviously, going into this, my plan was to just have a swarm of agents attack these puzzles the moment they came online and solve as many as humanly possible. I figured there might be one or two, which gave the agent some trouble, but with just like a little bit of steering and extra context, they
- 1:02
would be able to get the answer. But that was not even close to what actually happened. The agents I was running against these problems only ended up solving three out of the 13. everything else they just could not solve and we had to do a lot of work ourselves in order to solve these things. Now, the AI did end up being really, really useful and that's what I want to talk about in this video. The ways I was using it, the
- 1:23
way I think that the puzzle makers were able to thwart the AI in its current state, a lot of the weird behaviors I noticed from the models as I was using them, especially on these weird longunning puzzly type tasks, I saw things that I don't normally see when doing normal coding work. And I think there's a lot of reasons for that. And also just kind of talk about the puzzles because there were some really dope ones here. Like this one was just a dress
- 1:45
that this is all the context we were given is this dress and we had to get a message out of that. There was also one that was just a set of Rubik's cubes arranged like this that we had to again decode a message from. Now I want to start by talking about how I think the puzzle creators were able to beat the AI agents. The easiest way to do this I think is to start with one a problem that the AI could solve easily. And this
- 2:07
was one of the first ones we got access to. It didn't actually instantly solve it, but it was able to solve it a little bit later. The reason why is because on the first day, there was a slight error in the puzzle on like the 10th step or something like that. So, it wasn't instantly solvable right off the rip, but as soon as it was solvable, the AI was able to pretty easily rip through the entire thing. And the whole premise
- 2:27
of this puzzle is you're given this picture, some text on here, and you need to work your way through this manner. And the way you work your way through is by going deeper and deeper into the URL. So the solution for this page is some camelc case set of words that you then append to the URL up here and then that will take you to the next page and you just keep going over and over and over and over again until you hit the very
- 2:48
end. And this puzzle is shaped in such a way that agents really really like it. All of the resources are available digitally. It is all running over like web stuff. It's using images that are pretty easy to parse. It can do as much research as it wants. It can run for as long as it wants. Like this is a pretty easy problem for especially like coding agents to just solve. It takes a while.
- 3:08
It took like an hour and a half for it to get the actual solution. But again, there wasn't too many issues. It didn't get caught up in too many random things because there's just there's not too much here for the model. Lately, I've been building so much more, making so many more PRs to the point where CI is actually becoming a real bottleneck for me. GitHub just isn't cutting it, which is why today's sponsor, Blacksmith,
- 3:29
makes so much sense. It is a oneline change to your GitHub actions that results in you getting 2x faster hardware, 4x faster cache downloads, and 40x faster docker builds. It is insane. There are humongous companies like Mercury, Subbase, Ashb, Clerk, and many more using Blacksmith for a reason. Actually, including T3 Code. All of T3 Codes builds go through Blacksmith
- 3:49
because they are just so cheap, so fast, so good. I I don't really know what else to say. If you're working on any even remotely serious project, you're going to have a ton of CI. And if you have CI, you're just burning money by not using Blacksmith. And outside of GitHub actions, the new Codesmith background coding agent is really good. You can set it up in Slack or you can also just tag it in from GitHub or the web UI. And
- 4:11
once you have it set up, it can help with planning, implementing, optimizing code, also autofixing PRs, which is incredibly useful. the amount of like autonomous loops I have running for PRs at this point where I will have CI run some random thing fails is really useful to just have an agent instantly be spun off without me having to do anything make the required fix and just update the PR accordingly so that by the time I
- 4:32
go look at it everything is already exactly where I want it to be. Blacksmith will improve literally everything about your GitHub actions and CI experience at dav7.link/blacksmith. But then when you go over to a problem like this, this is one of the problems that the AI just completely fell apart on. I think I had like six different agents running at once trying to solve this thing. Three different codec
- 4:53
instances over three different machines. I had a cursor cloud run going, a couple of Fable runs going, just trying literally everything I could to try and get these things to be able to solve it. And there's a lot of reasons it was tough. One of the biggest ones was just the fact that since this was a physical in-person puzzle, the only way you could collect the data from it was by either
- 5:13
looking at it there or looking at pictures and videos of it. We got really good pictures and videos, but it was still just really hard for the model to reconstruct the data set that was required here. Cuz what you had to do is take this belt down here and then turn that into a key, an A through because there's 26 little beads on the belt. Each one is a different unique color.
- 5:35
And you can go A through Z labeling each of these different colors with a letter. Then once you have this set of labels, you can then label the grid on the chest. You can then label the front and the back part of the dress cuz there was uh 2x6 grids on each side. You merge those together to create this next grid. Then you end up with this data set and you can actually go through and do the
- 5:56
real solve. But getting to that point, the models could not do it. Some of them got a lot closer than I was expecting. But even if they did get a decent data set, there was just such insane amounts of data overload that these things could not figure it out. And to be clear, we were struggling with this, too. We spent a long time trying to figure out like, okay, does the number of beads matter?
- 6:16
Does the shape of the beads matter? Does whether or not it's transparent matter? Like, it did just end up being color, but it took us quite a while to figure that out. There's so many rabbit holes you can go down using this puzzle as a starting point that all of the agents just did not have any chance of actually solving this without us doing this leg work for it. I spent most of the time
- 6:36
working on this puzzle with Thor pirate software. Awesome guy, insanely talented at like cyber security type things and puzzles and ciphers and all that stuff that I have never really seen before. The thing that Thor was able to figure out was the standard columner transposition decryption. there were those like weird beads just floating on the belt from earlier. Like you can see here and here, those beads ended up
- 6:58
mattering for this right here. Uh once we had that and we had these two tables, you could like take the bottom table, I believe, and the top table and basically take the rows that overlapped and remove them from the table to get a new set of letters. Run this decryption over it. Read it horizontally. Answer X inconsistent knots. And the answer was inconsistent knots. We had to spend so
- 7:20
much time labeling and organizing the data and then also figuring out what paths are worth going down and what paths aren't worth going down. It was a lot of what we were doing there where like Thor and Gern who were the really talented like solving guys. I was just kind of sitting there in the corner listening to them talk trying to figure out what paths they wanted to go down and then furiously typing into the
- 7:40
agents to just have them explore all the things they're doing. I also spent a lot of time building out like visualizations for them cuz on the Rubik's cube puzzle, this ended up being really clutch because this is all we had. We had to figure out how to solve this puzzle off of these 12 cubes and black stands, two gold stands for some reason, then more black stands. I was not a huge part of solving this one outside of making the
- 8:02
visualizer. But my general understanding of how it works is in order to solve this, you need to first solve these two cubes because they're scrambled in basically the same way, just inverted. Once you have the moves to solve those two cubes, it's like eight moves, I believe. If you apply those same eight moves to all of the rest of the cubes on here, that creates this new view. So,
- 8:22
these two are solved. But if you look at this angle right here, it's kind of an awkward angle, but there's kind of an L here, an A here, a C here, a K here, and this all translates down to lack of dust. And that was the final answer. This took the guys so long to figure out because it's just a brutal, weird puzzle. The idea of an agent being able
- 8:43
to figure this out is hilarious. It requires so many things that they are just not that good at the 3D spatial reasoning, the visualization cuz like you really need to visualize this to see it. If you just had a set of data arranged in a grid, maybe they could figure it out, but like if this is all you're looking at, you're just looking at yb
- 9:03
like this crap. really hard to see a letter in that as compared to this. And once again, this puzzle has the same problem the dress one did where there's just so much data involved and it's so hard for the model to actually figure out what it's supposed to be doing. There are infinite numbers of rabbit holes to go down and the agents did go down rabbit holes. They ran for hours at a time. They really tried their hardest
- 9:24
to come up with solutions. There were two big things that really jumped out at me. The first of which was that one, hallucinations are not dead. When you're just using these things for coding work, it feels like hallucinations are not entirely solved, but more solved than they have been in the past. And they're at the point where like I'm really not too worried about them. If it can actually test its work against the real world, it'll probably do an okay job and
- 9:47
do a good job of going and referencing documentation. The new models are trained really well to go fetch data from reality and then produce answers based on that. These problems have no ties to reality whatsoever other than at the very end when you submit your final answer. The only way you can check that you're going down the right path is when you get to the very end and you submit
- 10:08
it and you get a yes or a no. That's it. As a result of that, it was really easy for the models to be going down this gigantic run. They're a 100 compactions deep. Dozens of sub agents have fired. They've been trying a bunch of stuff and just somewhere somehow some kind of stupid leap in logic happens. Often what I noticed it would end up being is the model would do some test like it would
- 10:29
try decrypting this in some weird way and that decryption would yield a phrase that was one letter off of making sense. You would just end up with like something that looks like the sky is and then instead of blue it would be like glue like the sky is glue and it would be like oh that could be a reference to this place where glue kind of makes
- 10:49
sense that's trying to prod us into realizing that it should actually be blue. So then it should be the sky is blue and then it would take that as fact and that fact would then poison it for the rest of the run and it would just keep going on and on and on and spinning down stupid rabbit holes that would produce pretty much nothing of value. You really had to babysit these things all the way through or it just wouldn't get solved. But the other part of that
- 11:11
was honestly even more fascinating to me was that as I was watching this happen across multiple agents on multiple devices, I noticed a recurring pattern where over and over again they would somehow independently naturally poison themselves in the same exact way and end up at the same exact wrong answers. It took a while. It took many hours and very long runs. But after enough
- 11:32
compactions, enough sub aents, enough grinding, even with different starting prompts and all that stuff, just working on the same puzzles going towards the same goal, they would eventually converge on the same exact wrong answer. Very fascinating to watch the probability machines, even at their gigantic current size, eventually converge on the thing, regress to the mean, and not get the right answer. This
- 11:54
really made me rethink a lot of things in the AI world right now. You know, we in the dev space like it feels like this stuff is completely inevitable and is taking over everything and is scaling really fast. It is like we are about to get the god AI and in a lot of tasks it does feel like this. Like I could definitely see a world where like that convergence happened because you know GPT 5.6 SAL is only so big. The thing is
- 12:16
maybe if you make the model bigger, maybe if you just make GPT56 10 times larger, make it a 50 trill per model, there will just be so many pathways within the neural net that it can go down to where it starts to actually create these novel pathways and novel solutions and can actually keep something this complex in its head and figure it out. That is totally in the
- 12:37
realm of possibility and it's an untested assumption by now. The things we've seen so far show that it is heading in that direction. I'm seeing first signs of that and I could believe that. I am very confident that just like reinforcement learning like RL the thing that they're doing to make these things insanely good at like writing code that is not going to be enough to have it be able to solve stuff like this on its
- 12:58
own. There's just no universe without scaling aka just making the models bigger working the way the labs think it will. I don't really see a world where anything other than it's an insanely useful tool that you have to direct because it can't direct itself. I can't see any world where we are just stuck there forever. Again, they're all trying to scale. I know especially Anthropic is
- 13:18
really going hard on this and I would not be surprised if by the end of the year they're able to solve more of these. This is actually a benchmark I'm going to keep so that whenever new models drop, I got to throw these puzzles at them and see how much further they can get. The way I tried to solve these things was with the agent swarms pattern or whatever. It's the thing that uh OpenAI has been doing internally for a lot of their early model testing. Like
- 13:40
all those math proofs that they put out, those were done with a swarm of a new model. The hugging face incident where they had the agent break out of containment, that was an agent swarm. It was dozens of instances of the same model running in parallel, doing their own things, communicating with each other over a message board that ended up allowing them to solve and achieve these
- 14:00
really big crazy things. And I figured that these problems would probably benefit a lot from that system. So I was trying that on all of these. It really wasn't enough. Like it does kind of feel like 5.6 is just right at the cusp of the generation where we can start doing swarms, but it's not quite there yet. I would bet that Astra will be substantially better at doing these.
- 14:20
This will be a fun thing to test out in a month or two, but for now at least, the swarms really didn't accomplish all that much other than just burning a lot of tokens for no real reason. I uh mentioned this in the article write up I did on this. I just posted on my Substack. It's not behind the subscription wall. It's totally free. I'll have it linked down below if you want to see that. It's a more uh put together version of these thoughts and I
- 14:41
have included in here the breakdown of all the tokens I used over the course of this weekend. You can see the vast majority was through codeex which just means it was GBT56. I used a decent amount of cloud code like 1.3 bill pi was me screwing around with some open source models to see if those could maybe be useful and spoiler they just weren't like Fable and 5.6 were the only models that were really even remotely
- 15:03
useful for this unfortunately. But you'll also notice looking at this table, the API equivalent cost of $24,000 worth of tokens that I burned over the course of just Defcon. And if you're not deep in the AI space, that probably seems completely ludicrous and insane. Like, why the hell did I spend 24 grand on this? And the answer is that I didn't. I did not spend 24 grand on this. I spent a total of $400 on this.
- 15:26
200 of that went to a Codeex sub. 200 of that went to a second codec sub just because I was burning my rate limits so hard especially on I think this was Friday. Was really trying the swarms thing out and just needed more tokens. So got the second one so I could get a reset. Um ended up canceling it afterward cuz generally for day-to-day work I only need one. And I guess actually I was wrong. It was 600 because
- 15:47
there was another $200 sub that was my Claude Code sub. These subscriptions from the Frontier Labs are so heavily subsidized. Yeah, on paper it's 24 grand, but like you can get I think it's something like 10 to 15 grand worth of inference per month off of your $200 sub on OpenAI. It's absurd how much you can get from these things. But you also have
- 16:08
to remember that just because the API cost is that does not mean that the model actually costs them that much to run. Like I'm pretty sure their margins are really really good on these models. They just make them really expensive because it's supply and demand. for now. Over time, I'm hoping that stuff like SpaceX, like the new Grock model is really good. The Chinese labs are really catching up. Like hopefully that
- 16:29
competition and choice is going to help drive these prices down over time. Again, it's it's not really 24 grand. It's a lot less than that. And to finish the video, I want to end with something more personal and optimistic. This was a very, very fun trip for me. I had a phenomenal time. And if anyone from the team is watching this, thank you for everything. I loved it. And I'm very
- 16:49
much looking forward to next year. It was really, really cool to just like be there, learn from you guys, do everything I could to help. I think the team just really worked well because I was there doing all of the AI stuff, trying to help them, and then they weren't as deep on the AI stuff and didn't have the crazy agent setups, but they knew how to solve the puzzles. So, just like working together, building
- 17:10
this stuff together. It was an awesome experience, and it also did a lot to kind of help with the uh, for lack of better term, AI psychosis. I think that this is kind of just a right of passage that you have to go through at this point. It is useful to have psychosis for at least a month or two. Like you need to go through the moment where you realize, oh it is all over. Like
- 17:31
writing code by hand is kind of dumb at this point. And yes, these things actually are insanely useful and are going to do fundamentally mind-boggling things over the next decades. Like this stuff is real and is not crypto 2.0. Is it a bubble? You can debate that all day long. Maybe in the short term it is, but in the long term I don't think it is. I think the long term of these things is
- 17:51
1,000% in the trillions, if not silly numbers that don't even mean anything beyond that. This stuff is real, but it is not there yet. And I really got to see leaving my field of expertise, which is software and AI stuff. That's the stuff I know really well. The AI seems and feels inevitable and superhuman to us because in our domain it really in a
- 18:13
lot of ways is part of the reason for that is because one the primary thing these labs are working on right now is training and RLing for coding agent environments. Like they are training these things to be incredibly good at writing code working in systems because it is the prerequisite to doing everything else. All of the robotic stuff, all of even like the bio stuff and the legal stuff and all of these
- 18:35
things require some semblance of code because code is the way the model interacts with the world. It can only produce text and that text can be code and that code is instructions that can cause something useful to actually happen. So this is the first thing that is going to get knocked down and it feels insane in this world. There's still a lot of limitations obviously like I have seen many many times where
- 18:57
it will confidently assert something stupid within code. it will fall over repeatedly. But if you are actively driving it and paying attention to what it's doing and reading the outputs and reading the thinking traces, you can steer these things to do incredible things. And everything I just said there about reading what it's doing, giving it good plans, giving it good direction, understanding the problem you're trying to have it solve. That's what we did
- 19:17
here. That's how this ended up working. We just bootstrapped this together with multiple people. I alone couldn't understand these problems, but they alone didn't understand the AIS well enough to do that. But together we were able to do the thing that a lot of us already do with ourselves, agents and code is working directly with it and understanding the problems it's trying to solve. When you have a human operator who actually understands the domain and
- 19:39
problem space and use these things for such useful stuff. Building out this visualizer was entirely just me telling Codeex to go make this, giving it some screenshots and data. Gave me a first pass. I was like, "Ah, this isn't great. Fix this, this, and this." Then I showed it to the team and they're like, "Ah, this is okay, but we need this, this, and this." Send another prompt over, get it iterated on over and over. And within
- 20:00
an hour, we had this visualization, which was very helpful for solving the thing. Then when we needed to get to this stage, we created this visualizer almost instantly. Then we were able to get the solution off that. Like that was really useful. Same thing with the dress here, building out this web page to visualize and see and label the data. Incredible, incredible stuff. And once we had more data and context and
- 20:20
theories, all of those could quickly be iterated on and tested by the agents. They're very powerful systems that just have to be steered by someone competent. And unless crossing the threshold from 5 to 10 trillion per models, which is probably what mythos is, up to 20 trill models is enough to magically induce discernment and consciousness, which I
- 20:41
don't think it is. I'm pretty sure it's a lot more parameters than that. But as long as that doesn't happen, we are definitely going to live in a world for a very long time where our skills still matter an insane amount, arguably even more, because the amount of leverage and output you can get just from knowing your now is so high. It is so cool the things we can do. I think the thing
- 21:02
that I am personally focused the most on right now with this stuff is being as ambitious as possible with what I'm trying to do while also being reasonable and more balanced in the way I do these things. Like you don't need to run the agents 24/7. Run them when you need them. No more, no less. Having personally pushed these things to the absolute limits of their capabilities
- 21:23
over the last week, I can pretty confidently say after watching them fall on their faces repeatedly, we definitely have time. So, I hopefully that this video turns into something semicoherent and useful. I was kind of just ranting through everything that happened. I really enjoyed this. If you did too, make sure to like and subscribe and I will talk to you again
Description
I went to DEF CON for the first time this year, and ended up helping win the GoldBug challenge. I used a lot of Codex, Claude, etc. But it really didn't end up helping that much...
Thank you Blacksmith for sponsoring! Check them out at: https://davis7.link/blacksmith
substack article - https://davis7.substack.com/p/burning...
cubes viewer - https://a-twisted-perspective.vercel.app
x summary - https://x.com/davis7/status/208735732...
My X, Discord, and more: https://davis7.sh/
