Self-evolving AI could close the loop in six months, but the model will quietly drift
The inflection point for RSI is models becoming capable of long-horizon tasks. Apodex's chief scientist says the self-evolution loop could close in as little as six months, but recursive drift and the model's taste will need human gatekeeping for a long time yet.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The ‘R’ in RSI only works now that tasks run long
Self-evolution is not a new concept. Two or three years ago there were already approaches like LM-as-optimizer and agent optimizer, but back then a model could only optimize for two or three hours, and one small mistake compounded on itself until the model could not keep iterating. What is different this year is that models have got stronger: they can finish long-horizon tasks that take a human 1.5 hours or even 12 hours, so a model can correct itself several times within a single task, and the ‘R’ in recursive finally holds. Eighty percent of Anthropic's code is written by Claude — it is precisely this coding ability that makes a model training itself feel natural.
— Bei BingLong-horizon tasks are three stacked problems, not just a context limit
Long-horizon work is not simply a matter of context length. First, standard attention is quadratic in complexity, so inference over a million tokens is enormously resource-hungry and needs architectural innovation. Second, supporting a long context is not enough on its own — the model has to have been trained on very long data, otherwise it goes out-of-distribution. Third, the training and inference infra has to change too, including GPU optimization and kernel-level code. Stack the three difficulties and the difficulty of long-horizon tasks rises exponentially. Simon (西蒙) adds that even with a million-token context, genuinely ultra-long-horizon tasks still need agent architecture and memory mechanisms to digest them.
— Bei BingRecursive drift is the fundamental obstacle to self-evolution
When a model generates its own training data, reasoning errors accumulate and amplify across generation after generation — that is recursive drift. Apodex's answer is verification: code and math are easy to verify, with test scripts or formal proofs, so they are relatively controllable. But the tests themselves can be too broad or too narrow, and subtle drift still occurs. So Apodex trains verification ability specifically, and during reinforcement learning has the judge learn alongside the model, to keep it from picking up bad behaviors such as reward hacking.
— Bei BingVerification comes from sub-agents checking each other, plus redundancy
The name Apodex comes from the Greek for ‘proof’. Verification methods divide by domain: unit tests for code, formal proofs for math — but more problems depend on human judgment. Apodex 1.0 uses an agent team: first decompose the problem, have one sub-agent solve it and a separate sub-agent verify it, so the context does not get contaminated; then have several agents work the same problem to produce redundancy, with a global agent judging which answer is more accurate. In training, agents are also taught to judge how trustworthy a source is — a forum is less reliable than a textbook.
— SimonThe three self-evolution paths have very different moats
Bei Bing (贝冰) divides self-evolution into three paths: pre-training, post-training, and the harness (the scaffolding). Evolving on the harness costs the least — you can do it by tuning API calls, the open-source community and academia have produced plenty of results there, and the moat is shallowest. Post-training is the more central and harder battlefield: you have to diagnose the model's defects first, manufacture targeted problems, and verify that the training worked, which takes a great deal of resources. Pre-training is currently mostly toy-scale tasks and has not really scaled to product level. Apodex works all three, with post-training and the scaffolding bootstrapping each other — left foot on right foot — iterating in alternation. To judge a company's capability, first ask which layer it evolves at.
— Bei BingA team of agents that check each other beats one smart AI
Simon says it flatly: one smart AI cannot beat a team that checks each other. The reason lies in the nature of self-attention — the longer the context, the worse attention performs, and linear attention does not solve ultra-long context either. A single model has a ceiling under current technology, the same way a person's working memory is limited and they need notes. So you want multiple agents dividing the work, sharing information through a memory mechanism, in order to handle ultra-long-horizon problems. Model capability and agent engineering are one plus one greater than two; agent teams in vertical domains still have a moat, it is just that the scaffolding has to be tuned as models iterate every month.
— SimonFor discovery models the hard part is verification, not generation
Generative models already have a mature training recipe, but a discovery model's goal is not to generate an answer — it is to propose an out-of-distribution hypothesis and verify it. The first step requires a creative hypothesis, and the difficulty is that it sits far away from the pre-training data. The second step is judging whether the hypothesis holds, which comes back to the verification problem. You can use deep research to gather existing information to help with hypotheses, and you can write code to run simulations to strengthen confidence. But unknown territory often has no standard answer and no trustworthy simulation environment, so self-evolution is Apodex's main method for getting to a Heavy Duty Solver.
— Bei BingSix months to the first loop; drift will take years
Bei Bing expects the first closed loop of self-evolution to run in six months to a year at the earliest: in some domain (coding, for instance) the model finds its own problems, manufactures its own training data, verifies itself, and completes one round of iteration. But the recursive part of RSI has to run many rounds, and recursive drift is still unsolved; he is confident it will be completely solved within a few years. What keeps him up at night most is not knowing whether the model has gone off course while self-evolving — both the safety objective and the performance objective can drift. For now the only option is to read model outputs by hand every day and intervene in agent behavior; in future, ‘distilling yourself’ could inject that monitoring ability into the model and speed the process up.
— Bei BingIn their own words · checked verbatim
A model that likes to flatter you actually has a very hard time proposing a bold hypothesis.
一个喜欢拍马屁的模型，其实很难提出一个大胆的假设。
Bei Bing42:58
One question that keeps me up at night is how we know that while this model is self-evolving, it is meeting human needs and has not gone off course.
让我睡不着觉的一个问题，就是我们怎么知道这个模型在自我进化的时候，它是满足人类的需求，它没有跑偏。
Bei Bing58:33
I'm not writing code one function at a time any more; I'm at a higher dimension, a higher level, monitoring different agents writing code.
我不在一个function一个function的写代码，而是我会在更高的一个维度，更高的一个层面监控不同的智能体携代码。
Bei Bing1:02:53
Taste has no right or wrong to it, and no good or bad either — only fitting and not fitting.
品味没有对或者是错的区别，也没有好或者坏的区别，就只有适合和不适合的区别。
Bei Bing1:05:00
I think our taste is only what — if you put it in the language of training — you'd call a war start; it just gave it a warm start.
我觉得我们的品位只是一个我们如果放在训练这个语言里面叫做war start，就只是给他热启动了一下。
Bei Bing1:06:02
Figures
| Share of Anthropic's code generated by Claude | 80% | 0:00 |
| Length of human task Claude 3 Ops can handle | about 4 minutes | 8:25 |
| Length of human task Claude 3.7 Sonnet can handle | about 1.5 hours | 8:25 |
| Length of human task Claude 4.6 Ops can handle | about 12 hours | 8:25 |
| Industry-accepted doubling period for the task length models can handle | doubles every 7 months | 9:27 |
| Apodex's expected timeline for closing the self-evolution loop | six months to a year | 57:29 |
| Target reduction in the recursive drift rate | from 10% drift per pass down to 1%, and eventually to 0.01% | 1:03:58 |
Glossary
- RSI (Recursive Self-Improvement)
- A model's ability to train itself and improve in a loop.
- Recursive Drift
- Errors compounding generation after generation as a model trains on data it generated itself, gradually pulling it away from the target.
- Heavy Duty Solver
- Apodex's goal — a general model that solves hard problems for which humans have no standard answer.
- Discovery Model
- A model that proposes and verifies new hypotheses rather than relying on known answers.
- Agent Team
- A collaborative system of sub-agents that divide the work and verify each other, used for ultra-long tasks and for verification.
How to listen
AI founders, large-model researchers, investors, and engineers who follow agent architecture.
The roughly five minutes at the end comparing the two founders' styles can be skipped; it has little to do with the main thread.