AI self-evolution could close the loop in as little as six months, but models will quietly drift off course
The inflection point for RSI is a model that can handle long-horizon tasks. Apodex's chief scientist says the self-evolution loop could close in as little as six months, but recursive drift and model taste still need humans minding them for a long time to come.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Recursive self-improvement became possible because of long-horizon tasks
Self-evolution is not a new idea. 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; one small mistake would compound on itself and it could not keep iterating. What is different this year is that models have gotten stronger and can complete long-horizon tasks that take a human 1.5 hours or even 12 hours, so they can self-correct many times within a single task, and the "R" in recursive finally holds up. At Anthropic, 80% of code is written by Claude — it is precisely coding ability that makes it natural for a model to train itself.
— Bei BingThree technical bottlenecks stand in the way of long-horizon tasks
Long-horizon tasks are not just a context-length problem. First, standard attention has quadratic complexity, so million-token inference is extremely resource-hungry and needs architectural innovation. Second, long-context support alone is not enough: the model must have been trained on very long data, otherwise it goes out-of-distribution. Third, the training and inference infra also has to change, including GPU optimization and kernel-level code. Stack those three difficulties together and the difficulty of long-horizon tasks rises exponentially. Simon adds that even with a million-token context, truly ultra-long-horizon tasks still need agent architecture and memory mechanisms to digest.
— Bei BingRecursive drift is the fundamental obstacle to self-evolution
When a model generates its own training data, reasoning errors accumulate and amplify across generations — that is recursive drift. Apodex's countermeasure is verification: code and math are easy to verify, using 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 specifically trains verification ability, and during reinforcement learning it has the judge learn alongside, to keep the model from learning bad behaviors like reward hacking.
— Bei BingVerification relies on sub-agents checking each other plus redundancy
The name Apodex comes from the Greek word for "proof." Verification methods differ by domain: code uses unit tests, math uses formal proofs, but more problems depend on human judgment. Apodex 1.0 uses an agent team: first decompose the problem, one sub-agent solves it and another sub-agent verifies it separately, avoiding context contamination; then have multiple agents work the same problem to produce redundancy, with a global agent judging which answer is more accurate. During training it also has agents judge the credibility of information sources — a forum, for instance, is less reliable than a textbook.
— SimonThe three self-evolution paths have very different moats
Beibing divides self-evolution into three paths: pretraining, post-training, and harness (scaffolding). Evolving at the harness level is the cheapest — you can do it just by calling APIs — and the open-source community and academia have produced plenty of results there, so the moat is the shallowest. Post-training is the more core and more difficult battlefield: you first have to diagnose model defects, create targeted problems, and verify training results, which takes a lot of resources. Pretraining is currently mostly toy-level tasks and has not really scaled to product level. Apodex does all three; post-training and scaffolding are like the left foot stepping on the right foot, iterating alternately. To judge a company's capability, first ask which layer it evolves at.
— Bei BingWhy an agent team beats a single smart AI
Simon states it plainly: one smart AI cannot beat a team that checks each other. The reason is the nature of self-attention — the longer the context, the worse attention works, and linear attention does not solve ultra-long context either, so a single model has a ceiling under current technology, just as a person's working memory is limited and needs notes. So multiple agents must divide the work and share information through memory mechanisms 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, though the scaffolding has to be fine-tuned as models iterate every month.
— SimonThe hard part of a discovery model is verification, not generation
Generative models already have mature training recipes, but a discovery model's goal is not to generate answers but to propose and verify out-of-distribution hypotheses. The first step requires a creative hypothesis, and the hard part is that it is far from the pretraining data; the second step is judging whether the hypothesis holds, which brings us back to verification. You can use deep research to gather existing information to assist the hypothesis, and you can write code to run simulations to build confidence. But unknown domains often have no standard answer and no trustworthy simulation environment, so self-evolution is Apodex's main route toward a Heavy Duty Solver.
— Bei BingA six-month loop, but drift will take years
Beibing expects the first self-evolution loop to close in as little as six months to a year: in some domain (coding, say) the model finds its own problems, generates its own training data, verifies itself, and iterates one round. But the recursive part of RSI requires many rounds, and recursive drift is still unsolved; he is confident it can be fully solved within a few years. What keeps him up at night is not knowing whether the model has drifted off course while self-evolving — both safety goals and performance goals can drift. For now the only option is to manually read model outputs every day and intervene in agent behavior; in the future, "distilling yourself" could inject that monitoring ability into the model and speed up the process.
— Bei BingIn their own words · checked verbatim
A model that likes to flatter is actually very unlikely to propose 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
The 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 drifted 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 don't write code function by function; instead I monitor different agents writing code from a higher dimension, a higher level.
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 distinction, and no good or bad distinction — only a suitable and unsuitable distinction.
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 just — if we put it in training language, it's called warm start — it just gives 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 code generated by Claude | 80% | 0:00 |
| Human task duration Claude 3 Ops can handle | about 4 minutes | 8:25 |
| Human task duration Claude 3.7 Sonnet can handle | about 1.5 hours | 8:25 |
| Human task duration Claude 4.6 Ops can handle | about 12 hours | 8:25 |
| Industry-accepted doubling period for task duration 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 recursive drift rate | from 10% drift per round to 1%, and ultimately to 0.01% | 1:03:58 |
Glossary
- RSI / Recursive Self-Improvement
- The ability of a model to train itself and improve in a loop.
- Recursive Drift
- When a model self-generates training data, errors accumulate across generations and it gradually deviates from the goal.
- 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 does not rely on known answers but proposes and verifies new hypotheses.
- Agent Team
- A collaborative system in which multiple sub-agents divide work and verify each other, used for ultra-long tasks and verification.
How to listen
AI founders, large-model researchers, investors, and engineers following agent architectures.
The roughly 5-minute comparison of the two founders' styles at the end can be skipped; it has little to do with the main thread.