Apodex's chief scientists predict AI could close its first self improvement loop within six months. The harder problem is teaching the model which scientific questions are worth asking.
Two scientists at Apodex, a research lab backed by Tianqiao Chen (陈天桥), are publicly predicting that AI could close its first complete self-improvement loop within roughly six months. The same scientists are clear that closing that loop is not the interesting part.
The harder problem, they argue, runs on a two-to-three-year horizon: teaching a self-evolving model to verify its own outputs, separate bold hypotheses from confident errors, and learn which scientific questions are worth pursuing in the first place.
Simon Shaolei Du, Apodex's chief scientist for reasoning models and training, and Beibin Li, the lab's chief scientist of self-evolution and coding, made the prediction on the Mandarin-language SV101 podcast E242. Du is also an associate professor of computer science at the University of Washington, according to the show notes. Apodex bills itself as a "heavy-duty solver" for problems without standard answers, not a general-purpose chatbot vendor; its company site positions the lab around research, not retrieval.
Apodex breaks from the broader race for ever-larger language models with a "discovery model": a system that generates hypotheses humans have not yet formulated and then verifies them, distinct from retrieval or question-answering systems that work from known information. Chen announced "Apodex 1.0 for AI Research" on LinkedIn, and the lab maintains a public HuggingFace organization page for its open-weight models, with a third-party profile on LLM Reference describing Apodex as building "Open-Weight Agentic Models for Deep Research."
Inside Apodex, the self-evolution architecture relies on multiple sub-agents cross-verifying each other, a hedge against what the scientists call "recursive drift" (递归漂移), the tendency of a self-modifying model to drift away from reliable behavior as it rewrites itself. Coding, Du said on the podcast, is the foundational capability underneath that architecture, because code is the easiest domain in which a model can mechanically verify whether its own output works.
Three technical paths to self-evolution are on the table, the scientists said: changes during pretraining, changes during post-training, and changes at the harness or tooling layer. The closed-loop milestone they are forecasting, six months out, is the moment a single cycle of self-modification, verification, and re-deployment completes end-to-end.
What comes after is the harder problem, and it is the one the lab is most explicit about. Recursive self-improvement only matters if the model can do its own verification, and verification is not a benchmark: it requires something the scientists call, in scientific-method terms, the taste to distinguish a bold hypothesis from a confident mistake. That epistemic capacity, they argue, is the bottleneck, not compute, parameters, or data.
Anthropic published its own framing of this terrain recently in a public Recursive Self-Improvement report, which defines the recursive-improvement phase as an industry concept and calls for human-designed deceleration and pause mechanisms. The Apodex prediction slots directly into that framework, with one notable difference: the lab's near-term target is closing the loop, not sustaining it. The two-to-three-year horizon in the podcast is for sustained, reliable self-improvement without drift.
The next trigger is concrete. Apodex says a closed self-evolution loop is on the lab's near-term roadmap, with the architectural pieces (sub-agent cross-verification, coding-first verification, harness-level tuning) already in place. Whether the model that emerges can tell genuine discovery from polished mimicry is the question the six-month timeline does not answer.