Applied Compute says its Ari agent can now run a post-training loop on its own. In a research report published on 4 November, it says Ari Autoresearch, live on its AC2 platform, took Qwen3.5-9B from 0.118 to 0.629 held-out reward on Lichess chess puzzles in one 16-hour episode, “a 5.3x improvement”. The result is from one task, and the scores are the company’s own.
The agent’s gain, +0.511, is about 12.8 times the +0.040 gain of the same agent with the same tools but without the new research skills, by our arithmetic.
Business Pill · Fine-Tuning
A short explainer on fine-tuning: adapting a general model to a specific task with your own data. It teaches the general idea only and says nothing about Applied Compute.
Held-out reward on Lichess chess puzzles, each agent’s own baseline run against its best run in a 16-hour episode, from Applied Compute’s Ari Autoresearch report (4 November 2026). The company’s own experiment on one task, not independently verified. Chart: FourWeekMBA.
The key insight: Applied Compute says Ari Autoresearch lifted Qwen3.5-9B’s held-out chess-puzzle reward from 0.118 to 0.629 in one 16-hour episode; its gain of +0.511 is about 12.8 times the +0.040 of the same agent with the same tools but without the new research skills, by our arithmetic, in the company’s own test on one task.
The report describes Ari Autoresearch as letting “Ari run its own post-training experiments on AC2”. Applied Compute’s post on X says: “Post-training a model usually takes a team of researchers running experiments for weeks”.
Co-founder and CEO Yash Patil wrote on X that Applied Compute is building “the full platform for open models: training, inference, routing, synthetic data generation, and endpoints that continuously improve”. The report says Ari Autoresearch “is available to AC2” platform users today.
Gain in held-out reward on Lichess chess puzzles from each agent’s own baseline to its best run in a 16-hour episode, from Applied Compute’s Ari Autoresearch report (4 November 2026). The company’s own experiment on one task, not independently verified; gains and the 12.8x ratio are our arithmetic.
What Was Posted on X
Inference isn’t everything.
We’re building the full platform for open models: training, inference, routing, synthetic data generation, and endpoints that continuously improve.
Ari, our applied research engineering agent, is built directly into AC2 to automate post-training. It curates data, runs experiments, evaluates results, and iterates using the same tools our researchers use.
Yash Patil, co-founder and CEO of Applied Compute, 19:02 UTC on 4 November. He says the company is building the full platform for open models and that Ari is built into AC2 to automate post-training.
Post-training a model usually takes a team of researchers running experiments for weeks. With Ari Autoresearch, now live on AC2, an agent can run that whole loop itself. In one 16 hour episode with no researcher intervention, it took Qwen3.5-9B from 0.118 to 0.629 held-out reward on chess puzzles.
Applied Compute’s launch post, 18:37 UTC. It says that in one 16-hour episode with no researcher intervention, Ari Autoresearch took Qwen3.5-9B from 0.118 to 0.629 held-out reward on chess puzzles.
The Chess Experiment
Each agent got a 16-hour horizon and 16 B300 GPUs, and trained Qwen3.5-9B on a pool of 8,000 Lichess puzzles, scored on 800 held-out puzzles with partial credit for each correct move. All four agents used GPT-6 Sol with medium reasoning effort, the report says.
Ari Autoresearch went from 0.118 to 0.629. Base Ari, with AC2 tools but without the autoresearch skills, went from 0.188 to 0.228. Codex with the skills but without AC2 went from 0.061 to 0.174, and base Codex completed only its baseline run, 0.189. By our arithmetic, Ari Autoresearch’s gain is about 4.5 times that of Codex with the same skills.
The report says Ari traced failures to illegal moves, showed the model its legal moves, then trained on replays of full solved puzzles. Applied Compute’s thread says this ran “with no researcher intervention”.
Why the Tools Mattered
The report says base Codex’s 250-step run “took over 13 hours, compared to roughly 2 hours for the same baseline on AC2”, leaving no time for a second experiment. Ari Autoresearch launched 9 runs in its episode, against 8 for base Ari and 13 for Codex with skills.
Applied Compute also says it tuned Ari’s research skills on decision points from past internal runs. On its 34 held-out evaluation tasks, the share of runs where Ari found the key root cause rose from 0.38 to 0.54.
Connecting the Coverage
This follows Applied Compute’s $100 million run rate, which it announced on 2 November with a $350 million round at a $3.25 billion valuation. Ari Autoresearch adds an agent to the training side of that platform.
As we read the report, the two ingredients were tested separately: the skills helped Codex too (+0.113), and AC2 helped base Ari finish more runs, but only the agent with both moved far from its baseline.
As we read Patil’s post, the launch extends Applied Compute’s pitch from serving models to training them: he lists training, inference, routing, synthetic data generation and endpoints as one platform.
Applied Compute, Ari Autoresearch report
“Each agent ran its own baseline from the same config, so starting reward varies with run-to-run training variance”
Three Implications
FOR AC2 USERS Applied Compute says Ari Autoresearch is available to AC2 platform users today, with Ari launching training jobs, creating datasets and improving harnesses.
FOR RESEARCH TEAMS The report says a frontier coding agent without AC2 spent over 13 hours on a 250-step baseline that took roughly 2 hours on AC2, so it never ran a second experiment.
FOR READERS OF THE NUMBER The 5.3x is a ratio of partial-credit rewards on one chess task with one 9B model; the report does not give results on enterprise tasks or the cost of an episode.
The framework says the rise of forward-deployed engineering shows that “technical capability has far outpaced organizational readiness”.
As we read it, Ari Autoresearch is an attempt to put part of that human layer into software: the report says Ari draws on accumulated insights from Applied Compute’s Applied Research Engineers. The framework does not discuss Applied Compute. For the full argument, read Forward-Deployed Engineering on businessengineer.ai.
The results are Applied Compute’s own, from one task with one model, and we found no independent evaluation in what we read. Held-out reward is a partial-credit score, not the share of puzzles solved. Each agent ran its own baseline, and the report says starting reward “varies with run-to-run training variance”. It does not say how Ari performs on enterprise tasks or what an episode costs.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Applied Compute says Ari Autoresearch lifted a 9-billion-parameter model’s held-out chess-puzzle reward from 0.118 to 0.629 in 16 hours without researcher intervention. By our arithmetic, that gain is about 12.8 times the gain of the same agent without the new skills, on one task, in the company’s own test.
A note on sourcing. Experiment details, results and quotations are from Applied Compute’s Ari Autoresearch research report and X posts by Applied Compute and Yash Patil, all read on 4 November 2026. The results are the company’s own, on one task, and not independently verified. Gains and ratios are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
