Rick

コンペで金メダル獲得するためには、たとえば10件くらいの気が付かないといけない勝利条件があるイメージです。 現在の「AI Agent＋一般的な計算リソース」の性能は、ナイーブなステアリングだと、勝利条件のうちの2,3件くらいまでは満たせることが多いと思います。残りの条件を達成するためには、人間の介入が今のところは必要だと考えています。 あるいは、相当な計算リソースと工夫しつくされたハーネスを用意すれば金メダルを取れるコンペもあるかもしれませんが、それはそれで普通に金メダルを獲得するより難しい気がします。 AI Agentの性能向上に従って、勝利条件も同時にシフトすると思うので、この構造はしばらく変わらないと思っています。

This 8th ranker in a recent Kaggle competition is not a human; it's an AI. Here's the story. In June, Meta entered AIRA₃ (their autonomous AI research system) in NVIDIA's Nemotron Model Reasoning Challenge. The job: fine-tune a 30B Nemotron so it reasons better. Same information for every team. Scoring on a private test set. The result: they placed 8th out of about 4,000 teams and took Gold, with a score of 0.888, inside the gold band. AIRA₃ runs many long-running agents (each a model plus a coding harness) in isolated environments. They coordinate through two shared surfaces: a forum for hypotheses and findings, and a shared filesystem for solution artifacts. Agents read each other's work and compound it as compute keeps running. The live Gold entry was an ensemble: GPT 5.5 with OpenCode and Claude 4.8 with ClaudeCode. Post-hoc, Muse Spark 1.2 with MuseCode also hit Gold on the same private set. Muse Spark 1.1 and GLM 5.2 landed Silver. Meta also says the same system, with only the task spec changed, cut latency by 27% on production GPU kernels and hit gold-level on an Akkadian tablet translation Kaggle. Still early. No full AIRA₃ blog yet. Prior architecture writeup is AIRA₂. The signal here is simple: research agents that share a forum and a disk can compete with strong human Kaggle teams at fine-tuning a frontier-ish open model. Kaggle LB: kaggle.com/competitions/nvid… Meta thread: xitter.gq/AIatMeta/status/209627… AIRA₂: ai.meta.com/research/publica…