{"schema":"postcutoff/event@1","as_of":"2026-11-19T22:11:00+01:00","url":"https://postcutoff.com/e/2026-10-21-mirendil-base-one-researcher-agents-pretraining/","md":"https://postcutoff.com/e/2026-10-21-mirendil-base-one-researcher-agents-pretraining/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"human editor","policy":"https://postcutoff.com/about/"},"license":null,"id":"2026-10-21-mirendil-base-one-researcher-agents-pretraining","date":"2026-10-21","date_precision":"day","short_title":"Mirendil-Base: one researcher, agents, <$10M","deck":null,"takeaway":"On Oct 21, 2026 Mirendil, the ex-Anthropic startup working on self-improving AI, announced Mirendil-Base, a 100B–500B-parameter coding base model.","category":"research","category_label":"Research","importance":3,"confidence":"medium","status":{"key":"partly","labels":["Partly confirmed"]},"sources":[{"n":1,"title":"Mirendil: Pretraining at scale without a pretraining team","url":"https://mirendil.com/news/pretraining-at-scale-without-a-pretraining-team/","type":"official","group":"primary","domain":"mirendil.com"},{"n":2,"title":"Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI means","url":"https://importai.substack.com/p/import-ai-478-math-for-alignment","type":"discussion","group":"reactions","domain":"importai.substack.com"}],"official":1,"filed":"2026-10-26","updated":"2026-10-26","orgs":["Mirendil"],"title":"Mirendil says one researcher with AI agents pretrained a frontier-scale coding base model in 3.5 months for under $10M","summary":"On Oct 21, 2026 Mirendil, the ex-Anthropic startup working on self-improving AI, announced Mirendil-Base, a 100B–500B-parameter coding base model. It says a single researcher built it in 3.5 months for under $10M in total compute, with AI agents writing over 99% of the code and running and debugging the experiments. On bits-per-byte over seven outside code repositories it beats DeepSeek-V4-Pro-Base, GLM-5.3 and Nemotron-550B-Base and trails the post-trained Kimi K3. Larger Large and XL versions are still training. All of this is self-reported, and no weights, API or paper have been released.","key_facts":["Family: Mirendil-Base (100B–500B parameters), Base-Large (500B–2T, ~80% through training) and Base-XL (2T–10T, ~19%), described as 'coding-specialized models'. Post-training has started","Cost and team: under $10M all-in (agent tokens, data processing, research, small experiments and the final run) over 3.5 months, with one unnamed researcher. Agents wrote over 99% of the code; the researcher chose what to try and what to trust","Data: only public data (GitHub and other public code, web pages, public datasets); no purchased or licensed data, no distillation and no synthetic data. Code written from scratch on PyTorch and JAX","Bits per byte on seven external repositories (lower is better): Kimi-K3 (post-trained) 0.193, Mirendil-Base 0.222, GLM-5.3 0.242, DeepSeek-V4-Pro-Base 0.243, Nemotron-550B-Base 0.247. Mirendil says Large and XL both already exceed Kimi-K3 (as summarised by Import AI)","Mirendil: 'We don't think we have reached fully autonomous recursive self-improvement yet, but we are closer than we have ever been.' Jack Clark's Import AI 478 (Oct 26) headlined it 'Mirendil shows what RSI means'"],"key_numbers":[],"tags":["pretraining","ai-rd-automation","recursive-self-improvement","coding","agents","base-model"],"science":null,"body_md":"## What happened\n\nMirendil says it pretrained a large coding base model without a pretraining team: one researcher directed AI agents that wrote the\ntraining stack, ran the experiments and diagnosed the failures. Its only evidence is a compression metric (bits per byte) on outside code.\nNo downstream benchmark of a finished model, no weights and no technical report have been published.\n\n## Why it matters\n\nPretraining a base model of this class used to take large teams and nine-figure budgets. If the claim holds, agents are already\ncompressing frontier-lab R&D into one person's work, which is the early stage of AI-driven AI research that labs and forecasters watch\nfor. It is still a self-reported result from a startup that is raising money.","disputed":[],"related":[{"id":"2026-09-22-mirendil-5b-valuation-talks","url":"https://postcutoff.com/e/2026-09-22-mirendil-5b-valuation-talks/","date":"2026-09-22","date_precision":"day","short_title":"Mirendil, an ex-Anthropic startup building self-improving AI, in talks to raise up to $1B at a $5B valuation","deck":null,"takeaway":"It shows investors paying up for teams that openly target automated AI R&D, as policymakers discuss restricting exactly that.","category":"business","category_label":"Business","importance":3,"confidence":"medium","status":{"key":"partly","labels":["Partly confirmed"]},"sources":3,"official":0,"filed":"2026-09-29","updated":"2026-10-26","orgs":["Mirendil","Kleiner Perkins","Andreessen Horowitz"]}],"people":[{"id":"jack-clark","name":"Jack Clark","url":"https://postcutoff.com/person/jack-clark/"}],"posts":[],"videos":[],"models":[],"changes":[{"date":"2026-10-26","type":"filed","text":"Created (Mirendil post read directly; five days after publication)"}],"provenance":{"agents":[{"model":"Claude Opus 5.5","maker":"Anthropic","tool":"Claude Code"}],"filed":"2026-10-26","run":null,"sources_read":"Mirendil post read directly; five days after publication","updated":"2026-10-26","human_review":null,"version":{"date":"2026-10-27"}},"gaps":[{"model_id":"gpt-6-2-astra","name":"GPT-6.2 Astra","cutoff":"2026-05","days_after":143,"in_training_data":false},{"model_id":"claude-opus-5-5","name":"Claude Opus 5.5","cutoff":"2026-06","days_after":113,"in_training_data":false},{"model_id":"gemini-4-hydrogen","name":"Gemini 4 Hydrogen","cutoff":"2026-03","days_after":204,"in_training_data":false},{"model_id":"grok-4-7","name":"Grok 4.7","cutoff":"2026-05","days_after":143,"in_training_data":false}],"short_url":null}