I thought by now AIs would not only be rewriting their code, but rewriting CPU microcode to optimize how their code is written and executed. Nowhere close it turns out.
Yeah, it’s crazy how fast things have changed in a month. I couldn’t find a more recent replication or similar study but it would be interesting to see it done with the current frontiers. Though I don’t think that would change much about the overall conclusion of the paper
Well, duh. If you could do this with Opus 4.8, we would know. When Astra’s successor is 2-3x better at math research, and the internal teams say “we believe we will get there,” I’m inclined to believe the insiders.
What if any of the older good models could also have written those math proofs if they were given the same order of magnitude of resources? We don’t know and there is literally no one else in the world to check it. To me it’s very suspicious that all these hacking, containment escape, hidden internal thinking, math proofs started coming out all at once in a very short time right as IPO talks have intensified and Chinese seem to get closer and closer, also regulation discussions are starting to get very serious. I have used these models and they are good, especially Fable, but not groundbreaking. With intelligent guiding I actually feel better using Opus 4.6 as I feel more in control, having less hidden away from me.
The actual title of the paper is: "Can AI agents conduct open-ended AI research? Early evidence from two case studies"
While I appreciate that the article is throwing a web blanket on doomer claims, the actual study doesn't really get into AI self-improvement. That doesn't require writing papers. That just requires autonomously writing a software system that can produce a better AI agent then the one that created it. That said, I have little worry about this being possible as I have seen no evidence of AI agents being able to produce a working software system of that scale.
I don't think RSI is typically used to describe self-improving agents - it's about improving the model itself, and its performance in agentic tasks.
Most of the gains in model performance from one release to the next are coming from RLVR post training, which has changed a lot over the last couple of years.
The old way was the model generates a response, then a static verifier looks at the response and evaluates it to assign a reward score. The new way is interactive with an agent running in a custom RL task simulation environment, then scored according to how well it completed the assigned task. For a SOTA model there will be many thousands of these simulation environments, each focusing on trying to teach the model/agent a different skill. Post-training also typically uses training curricula to walk the model up though through different levels of task difficulty.
Training has become very complex.
The job of a post-training AI research engineer consists of things like designing environments, designing training curricula, tweaking learning algorithms, running small scale experiments to verify ideas, etc.
When people talk about RSI, it seems they are mostly talking about automating the job of the post-training research engineer - coming up with new ideas, testing them out, building these environments, etc. At the end of the day there is only so much development speed-up to be had since you still need to actually run those experiments and do the post-training, and are bottle-necked by the amount of compute available to do this. The economics of developing/selling LLMs also requires you to balance development compute cost with revenue generated by the model, so even if you had the spare compute available to put into development, you are ultimately bottle-necked by how fast can the resulting model earn back that sunk cost before you can afford to start the next cycle.
It's not all-or-nothing since some aspects of this automating the job of the post-training research engineer are easier than others, and are already being done, while the job as a whole obviously requires full human intelligence.
What is commonly called an AI agent is the combination of a harness, an inference middleware, and a model. Those models are trained by a software system.
Right now agents are good enough for throwing semi-random ideas at the wall. Experiment compute is the bottleneck because it’s not much more than brute force search. A sufficiently intelligent agent with a deep model of its own architecture will more quickly and confidently locate improvements, the same way that high end LLMs can point out a bug and write a correct fix without even needing to observe and probe the program at runtime. If this level of research performance is reachable, experimentation may become much less of a bottleneck. Hopefully it isn’t.
this article reads like a joke the "new study" is from group of people that are not at the frontier. they test with $3k of anthropic credits (compare to the >$10M in compute used to solve recent NS last week)
There’s also the difference between a model recursively improving “itself” and improving itself via online learning.
The former being that these models are helping develop and train future models, but they might not veer too far off in architecture (yet).
The latter is a model being able to train/learn on the fly, in real time, permanently (not just in the current conversation/session), or in other words, adjusting/managing its own weights. But, it also seems like it would take an entire paradigm shift in model architecture from what most LLMs are built on, but I could be wrong.
We used OpenClaw to run these experiments so that our scaffold was agnostic to the model provider. We conducted dry-run experiments with models from OpenAI and Anthropic before settling on Opus 4.8 as the best-performing model. In response to concerns that our results might be principally explained by a limitation in our scaffold, we repeated our experiment on one paper using GPT-5.6 Sol and Codex, its native scaffold, with the same time and API budgets. The results of this experiment were similar to our OpenClaw/Opus 4.8 experiments. This makes us more confident that our results are not simply artifacts of a scaffold deficiency; this run reproduced nearly every single one of our identified failure modes
The agent required three interventions during the run. First, we needed to modify the scaffold to resolve a bug in the OpenClaw harness that affected Anthropic reasoning models. Second, we gave the agents a 24-hour deadline extension; at the time of the original deadline, the agents had submitted drafts with a completion report indicating that their self-review was a "Weak Reject" and outlining the next steps they would take if given additional time.
I'm fairly sure Fable 5.1 could have designed a better experiment than the authors here, but hey.
Of course it might not, it has been the holy grail of AI research for a long time. It would be great if we could leave some self improving code running on a blank slate of a computer while we sleep and the machine was crying asking me what is everything the next morning. None of AI researchers have had that moment outside of their dreams, so far, but it would be great if it happened.
The method of the NS advance involved RLHE (reinforcement learning via human example), and that is only open-ended if users continue to advance the frontier within chats ahead of publications.
Sure, but the point is that the labs use more powerful internal models for research work, not public models. Public models tend to lag the internal frontier by a decent margin, and are constrained in other ways by monitoring. It’s just not a useful indicator.
I thought by now AIs would not only be rewriting their code, but rewriting CPU microcode to optimize how their code is written and executed. Nowhere close it turns out.
Opus 4.8 plus OpenClaw. I feel like the space is moving so fast that the result with this setup says very little about how close we are actually now.
Yeah, it’s crazy how fast things have changed in a month. I couldn’t find a more recent replication or similar study but it would be interesting to see it done with the current frontiers. Though I don’t think that would change much about the overall conclusion of the paper
Well, duh. If you could do this with Opus 4.8, we would know. When Astra’s successor is 2-3x better at math research, and the internal teams say “we believe we will get there,” I’m inclined to believe the insiders.
What if any of the older good models could also have written those math proofs if they were given the same order of magnitude of resources? We don’t know and there is literally no one else in the world to check it. To me it’s very suspicious that all these hacking, containment escape, hidden internal thinking, math proofs started coming out all at once in a very short time right as IPO talks have intensified and Chinese seem to get closer and closer, also regulation discussions are starting to get very serious. I have used these models and they are good, especially Fable, but not groundbreaking. With intelligent guiding I actually feel better using Opus 4.6 as I feel more in control, having less hidden away from me.
Link to the actual paper: https://arxiv.org/abs/2607.27191
The actual title of the paper is: "Can AI agents conduct open-ended AI research? Early evidence from two case studies"
While I appreciate that the article is throwing a web blanket on doomer claims, the actual study doesn't really get into AI self-improvement. That doesn't require writing papers. That just requires autonomously writing a software system that can produce a better AI agent then the one that created it. That said, I have little worry about this being possible as I have seen no evidence of AI agents being able to produce a working software system of that scale.
I don't think RSI is typically used to describe self-improving agents - it's about improving the model itself, and its performance in agentic tasks.
Most of the gains in model performance from one release to the next are coming from RLVR post training, which has changed a lot over the last couple of years.
The old way was the model generates a response, then a static verifier looks at the response and evaluates it to assign a reward score. The new way is interactive with an agent running in a custom RL task simulation environment, then scored according to how well it completed the assigned task. For a SOTA model there will be many thousands of these simulation environments, each focusing on trying to teach the model/agent a different skill. Post-training also typically uses training curricula to walk the model up though through different levels of task difficulty.
Training has become very complex.
The job of a post-training AI research engineer consists of things like designing environments, designing training curricula, tweaking learning algorithms, running small scale experiments to verify ideas, etc.
When people talk about RSI, it seems they are mostly talking about automating the job of the post-training research engineer - coming up with new ideas, testing them out, building these environments, etc. At the end of the day there is only so much development speed-up to be had since you still need to actually run those experiments and do the post-training, and are bottle-necked by the amount of compute available to do this. The economics of developing/selling LLMs also requires you to balance development compute cost with revenue generated by the model, so even if you had the spare compute available to put into development, you are ultimately bottle-necked by how fast can the resulting model earn back that sunk cost before you can afford to start the next cycle.
It's not all-or-nothing since some aspects of this automating the job of the post-training research engineer are easier than others, and are already being done, while the job as a whole obviously requires full human intelligence.
What is commonly called an AI agent is the combination of a harness, an inference middleware, and a model. Those models are trained by a software system.
Right now agents are good enough for throwing semi-random ideas at the wall. Experiment compute is the bottleneck because it’s not much more than brute force search. A sufficiently intelligent agent with a deep model of its own architecture will more quickly and confidently locate improvements, the same way that high end LLMs can point out a bug and write a correct fix without even needing to observe and probe the program at runtime. If this level of research performance is reachable, experimentation may become much less of a bottleneck. Hopefully it isn’t.
Do you think the final product of research is papers?
Did you read what experiment was?
this article reads like a joke the "new study" is from group of people that are not at the frontier. they test with $3k of anthropic credits (compare to the >$10M in compute used to solve recent NS last week)
There’s also the difference between a model recursively improving “itself” and improving itself via online learning.
The former being that these models are helping develop and train future models, but they might not veer too far off in architecture (yet).
The latter is a model being able to train/learn on the fly, in real time, permanently (not just in the current conversation/session), or in other words, adjusting/managing its own weights. But, it also seems like it would take an entire paradigm shift in model architecture from what most LLMs are built on, but I could be wrong.
You may be interested in TITANS:
Test-Time Learning: The model updates its own memory weights while running an inference task.
We used OpenClaw to run these experiments so that our scaffold was agnostic to the model provider. We conducted dry-run experiments with models from OpenAI and Anthropic before settling on Opus 4.8 as the best-performing model. In response to concerns that our results might be principally explained by a limitation in our scaffold, we repeated our experiment on one paper using GPT-5.6 Sol and Codex, its native scaffold, with the same time and API budgets. The results of this experiment were similar to our OpenClaw/Opus 4.8 experiments. This makes us more confident that our results are not simply artifacts of a scaffold deficiency; this run reproduced nearly every single one of our identified failure modes
The agent required three interventions during the run. First, we needed to modify the scaffold to resolve a bug in the OpenClaw harness that affected Anthropic reasoning models. Second, we gave the agents a 24-hour deadline extension; at the time of the original deadline, the agents had submitted drafts with a completion report indicating that their self-review was a "Weak Reject" and outlining the next steps they would take if given additional time.
I'm fairly sure Fable 5.1 could have designed a better experiment than the authors here, but hey.
Of course it might not, it has been the holy grail of AI research for a long time. It would be great if we could leave some self improving code running on a blank slate of a computer while we sleep and the machine was crying asking me what is everything the next morning. None of AI researchers have had that moment outside of their dreams, so far, but it would be great if it happened.
> The researchers asked Anthropic’s Claude Opus 4.8, running on open-source software called OpenClaw
Meanwhile, Navier–Stokes was solved by an internal model significantly more capable than Astra (and therefore more capable than Mythos/Fable).
I’m afraid this sort of experiment is cope. The labs clearly believe RSI is coming soon.
The method of the NS advance involved RLHE (reinforcement learning via human example), and that is only open-ended if users continue to advance the frontier within chats ahead of publications.
Sure, but the point is that the labs use more powerful internal models for research work, not public models. Public models tend to lag the internal frontier by a decent margin, and are constrained in other ways by monitoring. It’s just not a useful indicator.