> As a mirror to real-world agent design: the limiting factor for general-purpose agents is the legibility of their environments, and the strength of their interfaces. For this reason, we prefer to think of agents as automating diligence, rather than intelligence, for operational challenges.
> The only other notable setback was an accidental use of the word "revert" which Codex took literally, and ran git revert on a file where 1-2 hours of progress had been accumulating.
Amazing that these tools don't maintain a replayable log of everything they've done.
Although git revert is not a destructive operation, so it's surprising that it caused any loss of data. Maybe they meant git reset --hard or something like that. Wild if Codec would run that.
I have had codex recover things for me from its history after claude had done a git reset hard, codex is one of the more reliable models/harneses when it comes to performing undo and redo operations in my experience.
You're correct for an actual git revert, but it seems pretty clear that the original authors have mangled the story and it was actually either a "git checkout" or "git reset". The "file where 1-2 hours of progress had been accumulating" phrasing only makes sense if those were uncommitted changes.
And the reason jj helps in that case is that for jj there is no such thing as an uncommitted change.
Yes, this exact scenario has happened to me a couple times with both Claude and Codex, and it's usually git checkout, more rarely git reset. They immediately realize they fucked up and spend a few minutes trying to undo by throwing random git commands at it until eventually giving up.
Start with env args like AGENT_ID for indicating which Merkle hash of which model(s) generated which code with which agent(s) and add those attributes to signed (-S) commit messages. For traceability; to find other faulty code generated by the same model and determine whether an agent or a human introduced the fault.
Then, `git notes` is better for signature metadata because it doesn't change the commit hash to add signatures for the commit.
And then, you'd need to run a local Rekor log to use Sigstore attestations on every commit.
Sigstore.dev is SLSA.dev compliant.
Sigstore grants short-lived release attestation signing keys for CI builds on a build farm to sign artifacts with.
So, when jujutsu autocommits agent-generated code, what causes there to be an {{AGENT_ID}} in the commit message or git notes? And what stops a user from forging such attestations?
I love the interview at the end of the video. The kubectl-inspired CLI, and the feedback for improvements from Claude, as well as the alerts/segmentation feedback.
You could take those, make the tools better, and repeat the experience, and I'd love to see how much better the run would go.
I keep thinking about that when it comes to things like this - the Pokemon thing as well. The quality of the tooling around the AI is only going to become more and more impactful as time goes on. The more you can deterministically figure out on behalf of the AI to provide it with accurate ways of seeing and doing things, the better.
Ditto for humans, of course, that's the great thing about optimizing for AI. It's really just "if a human was using this, what would they need"? Think about it: The whole thing with the paths not being properly connected, a human would have to sit down and really think about it, draw/sketch the layout to visualize and understand what coordinates to do things in. And if you couldn't do that, you too would probably struggle for a while. But if the tool provided you with enough context to understand that a path wasn't connected properly and why, you'd be fine.
I would’ve walked for days to a CompUSA and spent my life savings if there was anything remotely equivalent to this when I was learning C on my Macintosh 4400 in 1997
Did you actually learn C? Be thankful nothing like this existed in 1997.
A machine generating code you don't understand is not the way to learn a programming language. It's a way to create software without programming.
These tools can be used as learning assistants, but the vast majority of people don't use them as such. This will lead to a collective degradation of knowledge and skills, and the proliferation of shoddily built software with more issues than anyone relying on these tools will know how to fix. At least people who can actually program will be in demand to fix this mess for years to come.
If you don't see a difference between a compiler and a probabilistic token generator, I don't know what to tell you.
And, yes, I'm aware that most compilers are not entirely deterministic either, but LLMs are inherently nondeterministic. And I'm also aware that you can tweak LLMs to be more deterministic, but in practice they're never deployed like that.
Besides, creating software via natural language is an entirely different exercise than using a structured language purposely built for that.
We're talking about two entirely different ways of creating software, and any comparison between them is completely absurd.
> There was a time when you had to know ‘as’, ‘ld’ and maybe even ‘ar’ to get an executable.
No, there wasn't: you could just run the shell script, or (a bit later) the makefile. But there were benefits to knowing as, ld and ar, and there still are today.
I don't understand how OP thinks that being oblivious how anything work underneath is a good thing. There is a threshold of abstraction to which you must know how it works to effectively fix it when it breaks.
You can be a super productive Python coder without any clue how assembly works. Vibe coding is just one more level of abstraction.
Just like how we still need assembly and C programmers for the most critical use cases, we'll still need Python and Golang programmers for things that need to be more efficient than what was vibe coded.
But do you really need your $whatever to be super efficient, or is it good enough if it just works?
It would’ve been nice to have a system that I could just ask questions to teach me how it works instead of having to pour through the few books that existed on C that was actually accessible to a teenager learning on their own
Going to arcane websites, forum full of neckbeards to expect you to already understand everything isn’t exactly a great way to learn
The early Internet was unbelievably hostile to people trying to learn genuinely
I had the books (from the library) but never managed to get a compiler for many years! Was quite confusing trying to understand all the unix references when my only experience with a computer was the Atari ST.
Since there are no humans involved, it's more like growing a tree. Sure it's good to know how trees grow, but not knowing about cells didn't stop thousands of years of agriculture.
I wouldn't say it is a tree as such as at least trees are deterministic where input parameters (seed, environment, sunlight) define the output.
LLM outputs are akin to a mutant tree that can decide to randomly sprout a giant mushroom instead of a branch. And you won't have any idea why despite your input parameters being deterministic.
The Gas Town piece reminded me of this as well. The author there leaned into role playing, social and culture analogies, and it made a lot more sense than an architecture diagram in which one node is “black box intelligence” with a single line leading out of it…
Its not like tree at all because tree is one and done.
Code is a project that has to be updated, fixed, etc.
So when something breaks - you have to ask the contractor again. It may not find an issue, or mess things up when it tries to fix it making project useless, etc.
Its more like a car. Every time something goes wrong you will pay for it - sometimes it will get back in even worse shape (no refunds though), sometimes it will cost you x100 because there is nothing you can do, you need it and you can't manage it on your own.
Except that the tree is so malformed and the core structure so unsound that it can't grow much past its germination and dies of malnourishment because since you have zero understanding of biology, forestry and related fields there is no knowledge to save it or help it grow healthy.
Also out of nowhere an invasive species of spiders that was inside the seed starts replicating geometrically and within seconds wraps the whole forest with webs and asks for a ransom in order to produce the secret enzyme that can dissolve it. Trying to torch it will set the whole forest on fire, brute force is futile. Unfortunately, you assumed the process would only plagiarize the good bits, but seems like it also sometimes plagiarizes the bad bits too, oops.
Interesting article but it doesn’t actually discuss how well it performs at playing the game. There is in fact a 1.5 hour YouTube video but it woulda been nice for a bit of an outcome postmortem. It’s like “here’s the methods and set up section of a research paper but for the conclusion you need to watch this movie and make your own judgements!”
It does discuss that? Basically it has good grasp of finances and often knows what "should" be done, but it struggles with actually building anything beyond placing toilets and hotdog stalls. To be fair, its map interface is not exactly optimal, and a multimodal model might fare quite a bit better at understanding the 2D map (verticality would likely still be a problem).
Follow up Q: what are you supposed to do when the context becomes too large? Start a new conversation/context window and let Claude start from scratch?
Either have Claude /compact or have it output things to a file it can read in on the next session. That file would be a summary of progress for work on a spec or something similar. Also good to prime it again with the Readme or any other higher level context
I ask it to write a markdown file describing how it should go about performing the task. Then have it read the file next time. Works well for things like creating tests for controller methods where there is a procedure it should follow that was probably developed over a session with several prompts and feedback on its output.
It feels like one could produce a digest of the context that works very similarly but fits in the available context window - not just by getting the LLM to use succinct language, but also mathematically; like reducing a sparse matrix.
There might be an input that would produce that sort of effect, perhaps it looks like nonsense (like reading zipped data) but when the LLM attempts to do interactive in it the outcome is close to consuming the context?
I corroborate that spatial reasoning is a challenge still. In this case, it's the complexity of the game world, but anyone who has used Codex/Claude with complex UIs in CSS or a native UI library will recognize the shortcomings fairly quickly.
I actually think it would be pretty fun to code something to play video games for me, it has a lot of overlap with robotics. Separately, I learned about assembly from cheat engine when I was a kid.
You do you. I find this exceedingly cool and I think it's a fun new thing to do.
It's kind of like how people started watching Let's Plays and that turned into Twitch.
One of the coolest things recently is VTubers in mocap suits using AI performers to do single person improv performances with. It's wild and cool as hell. A single performer creating a vast fantasy world full of characters.
LLMs and agents playing Pokemon and StarCraft? Also a ton of fun.
This is a cool idea. I wanted to do something like this by adding a Lua API to OpenRCT2 that allows you to manipulate and inspect the game world. Then, you could either provide an LLM agent the ability to write and run scripts in the game, or program a more classic AI using the Lua API. This AI would probably perform much better than an LLM - but an interesting experiment nonetheless to see how a language model can fare in a task it was not trained to do.
This sounds as expected to me as a heavy user of Opus. Claude absolutely has a "personality" that is a lot less formal and more willing to "play along" with more creative tasks than Codex. If you want an agent that's prepared to just jump in, it's a plus. If you want an agent that will be careful, considered and plan things out meticulously, it's not always so great - I feel that when you want Claude to do reptitive, tedious tasks, you need to do more work to prevent it from getting "bored" and try to take shortcuts or find something else to do, for example.
People have been botting on Runescape since the early 2000s. Obviously not quite at the Claude level :). The botting forums were a group of very active and welcoming communities. This is actually what led me to Java programming and computer science more broadly--I wrote custom scripts for my characters.
I still have some parts of the old Rei-net forum archived on an external somewhere.
The opening paragraph I thought was the agent prompt haha
> The park rating is climbing. Your flagship coaster is printing money. Guests are happy, for now. But you know what's coming: the inevitable cascade of breakdowns, the trash piling up by the exits, the queue times spiraling out of control.
Surely it must have digested plenty of walkthroughs for any game?
A linear puzzle game like that I would just expect the ai to fly through first time, considering it has probably read 30 years of guides and walkthroughs.
Interesting this is on the ramp.com domain? I'm surprised in this tech market they can pay devs to hack on Rollercoaster Tycoon. Maybe there's some crossover I'm missing but seems like a sweet gig honestly.
I’ve been doing game development and it starts to hallucinate more rapidly when it doesn’t understand things like the direction it placing things or which way the camera is oriented
Gemini models are a little bit better about spatial reasoning, but we’re still not there yet because these models were not designed to do spatial reasoning they were designed to process text
In my development, I also use the ascii matrix technique.
Spatial awareness was also a huge limitation to Claude playing pokemon.
It really seems to me that the first AI company getting to implement "spatial awareness" vector tokens and integrating them neatly with the other conventional text, image and sound tokens will be reaping huge rewards.
Some are already partnering with robot companies, it's only a matter of time before one of those gets there.
I disagree. With opus I'll screenshot an app and draw all over it like a child with me paint and paste it into the chat - it seems to reasonably understand what I'm asking with my chicken scratch and dimensions.
As far as 3d I don't have experience however it could be quite awful at that
I wonder if they could integrate a secondary "world model" trained/fine-tuned on Rollercoaster Tycoon to just do the layout reasoning, and have the main agent offload tasks to it.
This was an interesting application of AI, but I don't really think this is what LLMs excel at. Correct me if I'm wrong.
It was interesting that the poster vibe-coded (I'm assuming) the CTL from scratch; Claude was probably pretty good at doing that, and that task could likely have been completed in an afternoon.
Pairing the CTL with the CLI makes sense, as that's the only way to gain feedback from the game. Claude can't easily do spatial recognition (yet).
A project like this would entirely depend on the game being open source. I've seen some very impressive applications of AI online with closed-source games and entire algorithms dedicated to visual reasoning.
Was able to have AI learn to play Mario Kart nearly perfectly. I find his work to be very impressive.
I guess because RCT2 is more data-driven than visually challenging, this solution works well, but having an LLM try to play a racing game sounds like it would be disastrous.
Not sure if you clocked this, but the Mario Kart AI is not an LLM. It's a randomized neural net that was trained with reinforcement learning. Apologies if I misread.
Crusader Kings is a franchise I really could see LLMs shine. One of the current main criticisms on the game is that there's a lack of events, and that they often don't really feel relevant to your character.
An LLM could potentially make events far more aimed at your character, and could actually respond to things happening in the world far more than what the game currently does. It could really create some cool emerging gameplay.
In general you are right, I expect something like this to appear in the future and it would be cool.
But isn't the criticism rather that there are too many (as you say repetitive, not relevant) events - its not like there are cool stories emerging from the underlying game mechanics anymore ("grand strategy") but players have to click through these boring predetermined events again and again.
You get too many events, but there aren't actually that many different events written, so you repeat the same ones over and over again. Eventually it just turns into the player clicking on the 'optimal' choice without actually reading the event.
Although git revert is not a destructive operation, so it's surprising that it caused any loss of data. Maybe they meant git reset --hard or something like that. Wild if Codec would run that.
And what would that reason be? You can git revert a git revert.
And the reason jj helps in that case is that for jj there is no such thing as an uncommitted change.
Then, `git notes` is better for signature metadata because it doesn't change the commit hash to add signatures for the commit.
And then, you'd need to run a local Rekor log to use Sigstore attestations on every commit.
Sigstore.dev is SLSA.dev compliant.
Sigstore grants short-lived release attestation signing keys for CI builds on a build farm to sign artifacts with.
So, when jujutsu autocommits agent-generated code, what causes there to be an {{AGENT_ID}} in the commit message or git notes? And what stops a user from forging such attestations?
> you can manually stage against @-: [with jujutsu]
You could take those, make the tools better, and repeat the experience, and I'd love to see how much better the run would go.
I keep thinking about that when it comes to things like this - the Pokemon thing as well. The quality of the tooling around the AI is only going to become more and more impactful as time goes on. The more you can deterministically figure out on behalf of the AI to provide it with accurate ways of seeing and doing things, the better.
Ditto for humans, of course, that's the great thing about optimizing for AI. It's really just "if a human was using this, what would they need"? Think about it: The whole thing with the paths not being properly connected, a human would have to sit down and really think about it, draw/sketch the layout to visualize and understand what coordinates to do things in. And if you couldn't do that, you too would probably struggle for a while. But if the tool provided you with enough context to understand that a path wasn't connected properly and why, you'd be fine.
what a world!
People don’t appreciate what they have
A machine generating code you don't understand is not the way to learn a programming language. It's a way to create software without programming.
These tools can be used as learning assistants, but the vast majority of people don't use them as such. This will lead to a collective degradation of knowledge and skills, and the proliferation of shoddily built software with more issues than anyone relying on these tools will know how to fix. At least people who can actually program will be in demand to fix this mess for years to come.
There was a time when you had to know ‘as’, ‘ld’ and maybe even ‘ar’ to get an executable.
In the early days of g++, there was no guarantee the object code worked as intended. But it was fun working that out and filing the bug reports.
This new tool is just a different sort of transpiler and optimiser.
Treat it as such.
And, yes, I'm aware that most compilers are not entirely deterministic either, but LLMs are inherently nondeterministic. And I'm also aware that you can tweak LLMs to be more deterministic, but in practice they're never deployed like that.
Besides, creating software via natural language is an entirely different exercise than using a structured language purposely built for that.
We're talking about two entirely different ways of creating software, and any comparison between them is completely absurd.
No, there wasn't: you could just run the shell script, or (a bit later) the makefile. But there were benefits to knowing as, ld and ar, and there still are today.
Just like how we still need assembly and C programmers for the most critical use cases, we'll still need Python and Golang programmers for things that need to be more efficient than what was vibe coded.
But do you really need your $whatever to be super efficient, or is it good enough if it just works?
Going to arcane websites, forum full of neckbeards to expect you to already understand everything isn’t exactly a great way to learn
The early Internet was unbelievably hostile to people trying to learn genuinely
Assembly programmers from years gone by would likley be equally dismissive of the self-aggrandizing code block stitchers of today.
(on topic, RCT was coded entirely in assembly, quite the achievement)
LLM outputs are akin to a mutant tree that can decide to randomly sprout a giant mushroom instead of a branch. And you won't have any idea why despite your input parameters being deterministic.
Code is a project that has to be updated, fixed, etc.
So when something breaks - you have to ask the contractor again. It may not find an issue, or mess things up when it tries to fix it making project useless, etc.
Its more like a car. Every time something goes wrong you will pay for it - sometimes it will get back in even worse shape (no refunds though), sometimes it will cost you x100 because there is nothing you can do, you need it and you can't manage it on your own.
Also out of nowhere an invasive species of spiders that was inside the seed starts replicating geometrically and within seconds wraps the whole forest with webs and asks for a ransom in order to produce the secret enzyme that can dissolve it. Trying to torch it will set the whole forest on fire, brute force is futile. Unfortunately, you assumed the process would only plagiarize the good bits, but seems like it also sometimes plagiarizes the bad bits too, oops.
Maybe this is obvious to Claude users but how do you know your remaining context level? There is UI for this?
There might be an input that would produce that sort of effect, perhaps it looks like nonsense (like reading zipped data) but when the LLM attempts to do interactive in it the outcome is close to consuming the context?
i enjoy playing video games my own self. separately, i enjoy writing code for video games. i don't need ai for either of these things.
https://bansostudio.etsy.com
It's still a neat perspective on how to optimize for super-specific constraints.
It's kind of like how people started watching Let's Plays and that turned into Twitch.
One of the coolest things recently is VTubers in mocap suits using AI performers to do single person improv performances with. It's wild and cool as hell. A single performer creating a vast fantasy world full of characters.
LLMs and agents playing Pokemon and StarCraft? Also a ton of fun.
Am I reading a Claude generated summary here?
> "This was surprising, but fits with Claude's playful personality and flexible disposition."
I still have some parts of the old Rei-net forum archived on an external somewhere.
https://ubos.tech/mcp/runescape-mcp-server-rs-osrs/
> The park rating is climbing. Your flagship coaster is printing money. Guests are happy, for now. But you know what's coming: the inevitable cascade of breakdowns, the trash piling up by the exits, the queue times spiraling out of control.
A linear puzzle game like that I would just expect the ai to fly through first time, considering it has probably read 30 years of guides and walkthroughs.
not just make up bullshit about events
pretty heavy/slow javascript but pretty functional nonetheless...
Gemini models are a little bit better about spatial reasoning, but we’re still not there yet because these models were not designed to do spatial reasoning they were designed to process text
In my development, I also use the ascii matrix technique.
It really seems to me that the first AI company getting to implement "spatial awareness" vector tokens and integrating them neatly with the other conventional text, image and sound tokens will be reaping huge rewards. Some are already partnering with robot companies, it's only a matter of time before one of those gets there.
As far as 3d I don't have experience however it could be quite awful at that
It was interesting that the poster vibe-coded (I'm assuming) the CTL from scratch; Claude was probably pretty good at doing that, and that task could likely have been completed in an afternoon.
Pairing the CTL with the CLI makes sense, as that's the only way to gain feedback from the game. Claude can't easily do spatial recognition (yet).
A project like this would entirely depend on the game being open source. I've seen some very impressive applications of AI online with closed-source games and entire algorithms dedicated to visual reasoning.
I'm still trying to figure out how this guy: https://www.youtube.com/watch?v=Doec5gxhT_U
Was able to have AI learn to play Mario Kart nearly perfectly. I find his work to be very impressive.
I guess because RCT2 is more data-driven than visually challenging, this solution works well, but having an LLM try to play a racing game sounds like it would be disastrous.
HN second-chance pool shenanigans.
Genuinely interested.
An LLM could potentially make events far more aimed at your character, and could actually respond to things happening in the world far more than what the game currently does. It could really create some cool emerging gameplay.
But isn't the criticism rather that there are too many (as you say repetitive, not relevant) events - its not like there are cool stories emerging from the underlying game mechanics anymore ("grand strategy") but players have to click through these boring predetermined events again and again.