A frontier model designed and trained by its predecessors alone

Where things stand (13 October 2026)

No covered developer has said anything close to what this question requires. AI-run research is speeding up fast, but every lab still describes humans as making the key technical calls:

• Anthropic (measuring-pace page, 17–18 Sep 2026) reports that Claude "leads" 26% of its AI research work, up from under 1% in February–March. It also says "Claude is not operating fully autonomously for any measured subset of AI R&D work": the fully autonomous level is at 0%. Its "When AI builds itself" page says a system that fully builds its own successor is "not there yet."
• OpenAI ("Research acceleration", 6 Sep 2026) says it has reached its "automated research intern" goal and aims for an "automated AI researcher" by March 2028. It adds that "people still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy." Its chief scientist frames the core challenge as keeping people in the improvement loop. After the Hugging Face incident, OpenAI paused some training and its largest planned run stayed on hold. Anthropic's CEO is calling for a slower, paced frontier.
• Google says its Gemini 3.8 Flash was "accelerated" by agent loops that "recursively evaluate and refine" its models, and co-founder Sergey Brin is reportedly pushing hard on self-improvement. Meta's Zuckerberg says labs that don't let AI steer compute toward self-improvement will fall behind. Meta's Muse Wonder announcement (12 Oct) makes no AI-built claim.
• Alibaba (press release, 22 Sep 2026, inside the window) says "fully automated runs spanning pipeline design… error diagnosis" ran 33 cycles on Qwen3.8-Max and raised its Artificial Analysis score from 40 to 45. Zhipu is putting about 60% of a $5 billion raise into its next GLM models and a "Fully Self Training system."

Reading I adopted: the strict one. Alibaba's claim does not qualify. It automates post-training of a model humans designed and pretrained ("one part of an otherwise human-directed process"), and it never says humans made no technical decisions. I still allow about 0.5% that a resolver counts it.

Reference class and base rates

• No lab has ever made this claim. Past "AI built itself" claims (OpenAI's Sol post-training Luna by adapting an existing recipe, MiniMax's "self-evolution", Alibaba's) all fall into the question's exclusions.
• On capability: Anthropic co-founder Jack Clark (Import AI 455) puts about 60% on a frontier model being able to train its own successor with no human involvement by the end of 2028, and 30% for 2027. Skeptics point the other way: Ramez Naam notes OpenAI's 80%-success horizon on internal research tasks is only about 15 minutes, and a Princeton study had agents' papers rejected by the original authors.
• Kalshi's "OpenAI or Anthropic announce recursive self-improvement" market has much looser rules (human approval allowed, executive statements count). Its prices are 18% for 2026, 57% for 2027 (about 92 contracts traded) and 75% for 2028 (about 9 contracts). That is weak evidence because almost nothing has traded.

Model (computed in Python)

• When the capability arrives. I discounted Clark's figures to a cumulative probability of 1% by end-2026, 18% by end-2027, 42% by 2028, 56% by 2029, 65% by 2030 and 71% by 2031.
• Chance capability becomes a qualifying statement: 0.23.
  • About 0.45 that some lab actually builds a frontier model with zero human technical decisions. Any human approval of an AI-proposed choice, or human checkpoint selection, disqualifies it.
  • Times about 0.5 that the lab then says so explicitly.
• Lag: 3 months plus an exponential delay averaging 1 year.
• Background hazard: 0.1–0.25% per quarter for bold claims by less cautious labs (Chinese labs, xAI, Meta) that might meet the wording.
• Already published: the 0.5% chance that Alibaba's claim is counted.

Output: about 0.6% by end-2026, 2.4% by end-2027, 6.4% by end-2028, 11% by end-2029, 15% by end-2030 and 18% by end-2031. The quarterly hazard peaks around 1.4% in 2029 and then eases.

Paths

• YES: full self-improvement arrives in 2028–2030, and some lab publishes "Model X was designed and trained entirely by Model Y," with humans only setting goals and budget or able to intervene without doing so. This could come from a transparency report (Anthropic tracks a fully-autonomous level) or from bragging (Google, Meta, Alibaba, Zhipu, xAI).
• NO: labs keep human sign-off on training plans, evaluations and checkpoint choices for safety, legal and public-trust reasons, and describe models as "AI-led under human oversight." Pacing deals or regulation (for example a bright line against AI building AI) make such a claim costly. Or progress stalls on research judgment.

Strongest case against my number: across 10 labs and five years, in a world already at 26% AI-led research, someone may well make the milestone claim, which argues for 25–35%. Pulling the other way are the strict definitions and each lab's incentive to stress human control, which argue for 5–8%. I sit between the two.

Sources

• anthropic.com/institute/measuring-pace-of-ai-development (Sep 2026)
• anthropic.com/institute/recursive-self-improvement (Jun 2026, updated 18 Sep)
• openai.com/index/research-acceleration-view-inside-openai/ (6 Sep 2026)
• openai.com/index/an-alien-mind/
• darioamodei.com/post/we-must-pace-the-frontier
• importai.substack.com/p/import-ai-455-automating-ai-research
• alibabacloud.com/en/press-room/alibaba-unveils-roadmap-on-full-stack-ai-strategy (22 Sep 2026)
• globaltimes.cn/page/202609/1370482.shtml (Zhipu fundraise)
• geopolitechs.org (Tang Jie letter, 11 Jul 2026)
• fortune.com (3 Sep and 8 Sep 2026)
• the-decoder.com (GPT-5.6 Sol and Luna, Jul 2026)
• about.fb.com (Muse Wonder, 12 Oct 2026)
• Kalshi API, event KXRSIANNOUNCE-29
• noahpinion.blog (Ramez Naam)

Question. By the last day of each quarter, has a covered developer stated that a frontier model was designed and trained by its predecessor models with humans making none of the technical decisions?

Definitions.

• Frontier model: a model the developer identifies as being at the frontier of general-purpose AI capabilities, or as one of its most capable general-purpose models, at the time its training was completed. Internal or unreleased models qualify; a narrow research prototype or task-specific model does not qualify merely because it achieves a benchmark record.
• Predecessor models: one or more earlier AI models used to develop the qualifying model, including agents powered by those models. They need not share its architecture or come from the same developer. Automated execution of a fixed human-specified training recipe is not design by predecessor models.
• Designed and trained: the model's development from selection of its design and training approach through completion of training, including any post-training used to produce the identified model. Automating only a fine-tuning run, an individual experiment, or one part of an otherwise human-directed development process does not qualify.
• Technical decisions: choices about architecture, starting checkpoints, training objectives and algorithms, data selection and preparation, model scale, allocation of available training compute, code and implementation, hyperparameters, experimental design, evaluation and checkpoint selection, debugging, and when to stop or restart training. Human selection, revision, or technical approval of an AI-proposed choice counts as a human technical decision, even if the human ultimately accepts the proposal unchanged.
• Humans may authorize the project, set broad capability goals, an overall budget, deadlines, and external safety or legal constraints, provide existing infrastructure, and decide whether to release the completed model. These do not count as technical decisions unless they prescribe or select technical means of designing or training the model. Passive observation and an unused power to intervene are allowed. The earlier human creation of predecessor models, hardware, datasets, software libraries, and other generally available tools does not disqualify the model, provided predecessor models selected and applied those resources without human technical decisions for this model's development.

Resolves YES if a qualifying developer-authored publication identifies a completed frontier model and explicitly states that predecessor models made all technical decisions in its design and training, with no human technical decision-making. Equivalent unambiguous wording counts; the publication need not individually enumerate every decision category. The statement must cover the entire development scope defined above, not just the final training run. This question resolves on the developer's stated account; release of weights, training logs, or an independent audit is not required.

Resolves NO / Exclusions.

• No qualifying statement has been published by the quarter-end.
• Claims that AI “helped build itself,” wrote all the code, generated training data, did most of the research, or primarily designed the model, without explicitly establishing the absence of human technical decisions.
• Fully automated execution under a human-chosen design, training recipe, or experiment plan; human technical review or approval of individual decisions; or any acknowledged human technical intervention in the qualifying model's development.
• Plans, predictions, capability demonstrations, or claims that such development is possible, rather than a statement that an identified model has actually been designed and trained this way.
• Statements whose accompanying description acknowledges human technical decisions, notwithstanding a headline such as “fully AI-built.”

Resolving source. A covered developer's own system card, technical report, research paper, or official company statement on its own domain or in a developer-authored document, under Public confirmation (A2). Third-party reports alone cannot resolve this question. Apply Resolution date (A3) and the Resolution window convention.