TJL Frontier Brief · AI Race Foresight No. 2
AI has started improving AI. Here is what forms next.
The newest models, agents and machines of autumn 2026, our October radar marked against the evidence, and ten directions with what should be true by October 2027 and 2028.
Leadership briefing · November 2026
The briefing on one page
Five things to know from the last ten weeks of AI.
What's new · Frontier models
Six frontier models in ten weeks. The strongest go to defenders first.
Between 3 September and 9 November, OpenAI, Anthropic and Google each shipped new flagships, and open models reached 2.8 trillion parameters. The most cyber-capable models now reach vetted defenders before anyone else.
- Gated releaseClaude Mythos 5.2 goes to vetted organisations, Gemini 4 Carbon to cyber defenders through Google's Fairwind Program, GPT-6 Astra's top cyber abilities to approved partners.
- Open weights at trillion scaleKimi K3.1 Pro: 2.8 trillion parameters, 104 billion active per token, a 1M-token context; weights published in early November.
- Private tests, wider gapWhen Artificial Analysis moved 60% of its index to private tests, the gap between the best closed and the best open model grew from 3 to 12 points.
Where it's headingThe report's highest-conviction call: by October 2027 a frontier lab names agent-built evaluations as a stage of its training (Direction 2).
Sources: OpenAI (3, 29 Sep); Anthropic (22 Sep, 9 Nov) via TechCrunch and CNBC; Google DeepMind (Gemini 4 Argon model card, 13 Oct; Gemini 4 Carbon, 28 Oct); Moonshot AI (Kimi K3.1 Pro); Artificial Analysis Intelligence Index v5 (27 Oct).
What's new · Agents
Agents now come with their own computer.
OpenAI's Dots each keep a cloud computer and browser of their own, reach thousands of apps, and research on their own while idle. Google says Gemini 4 Carbon worked for days on single engineering tasks. The rules people set decide where an agent stops and asks.
- A computer of its ownDots (29 September) run in their own cloud computer and browser, connect to 4,000+ apps, and use read-only access for proactive research when idle.
- Rules decide when it asksUsers set when an agent may act alone and when it needs approval; enterprise admins decide who gets one.
- Longer and longer tasksMETR's best measured model finishes, half the time, tasks that take people 16 hours or more; its test suite cannot measure beyond that.
Where it's headingBy October 2027 a platform sells checked agent runs with a published verification budget (Direction 8). No product yet publishes how often its checks raise false alarms.
Sources: OpenAI DevDay, 29 Sep 2026 (Dots, Agents API) via 9to5Google and InfoQ; Google, Gemini 4 Carbon announcement (28 Oct); METR time-horizon reporting (2026); TJL AI Race Foresight No. 2, Candidate 3.
What's new · Mathematics
AI now writes proofs a computer can check.
In September Claude produced the first complete machine-checked proof of Fermat's Last Theorem: dozens of agents working for eleven days, with occasional human direction. The Lean proof assistant checked every step, and no model can talk its way past Lean.
- 6 NovPSI reports a Lean-checked proof of the 18-year-old Aaronson–Ambainis conjecture, found by a non-public Anthropic model. Expert review is pending.
- 9 NovAnthropic says Claude Fable 5.2 solved two open problems in theoretical physics.
Where it's headingWhere a perfect checker exists, AI moves first. The report's forced answer is that evaluation itself becomes the scarce resource (sheet 16).
Sources: Anthropic, "Formalizing Fermat's Last Theorem" (4 Sep 2026); The Quantum Insider (6 Nov); TechCrunch (9 Nov).
What's new · World models and robots
Robots now imagine before they act.
World models predict what a scene will do next. Google DeepMind's Genie 4 generates interactive worlds to practise in, world action models plan and act in one network, and Gemini Robotics 2 runs a humanoid's whole body.
- Worlds on demandGenie 4 (27 October) builds real-time interactive worlds with physics and native sound, with a Gemini planner in the loop. DeepMind pitches them as training grounds for robots and agents; long sessions still drift.
- One body, three brainsGemini Robotics 2 splits a humanoid into planning, vision-and-language to motion, and motor control on the robot itself, from legs to fingers.
- Imagine, then actWorld action models predict the next frame and the next action together. LeWAM lifted contact-rich manipulation from 28.6% to 89.7% success with 17M parameters.
Where it's headingBy October 2027, a robot maker reports a self-improvement loop anchored on a simulator (Direction 4), and a policy taught mainly by a world model's critiques beats one trained on real reward (Direction 9). Both rated low confidence.
Sources: Google DeepMind, Genie 4 (27 Oct 2026) and Gemini Robotics 2 (30 Jul 2026); TJL AI Race Foresight No. 2, §10.8 and §18.
What's new · Infrastructure
The race is now physical: chips, memory and power.
The four largest cloud companies guide to about $720–745 billion of capital spending in 2026. Chips are only one link: memory, packaging and above all power now set the pace.
- Silicon built for agentsNVIDIA's Vera Rubin ships from autumn 2026 with a CPU designed for AI agents and a dedicated inference accelerator. NVIDIA claims ten times the agent throughput of Grace Blackwell.
- Power is the bottleneckIDC estimates that almost 7 GW of planned US data-centre capacity for 2026 has been delayed or cancelled for lack of power.
- Even orbitGoogle's Project Suncatcher ran a model on a data-centre TPU in orbit in October. Google says space data centres cost up to 100 times more today.
Sources: NVIDIA (31 May 2026 release; Q2 FY2027 results); company capital-spending guidance, Oct 2026, as compiled by Datacentres.com; IDC; Google and DCD (Project Suncatcher).
What's new · The Kingdom
The Kingdom is building the stack, from grid connections to Arabic models.
HUMAIN, owned by PIF and Aramco, is financing data-centre power and compute and shipping Arabic models. In November it signed the first deals under the National Infrastructure Fund's AI data-centre financing programme.
- OwnershipPIF holds 51% of HUMAIN. Aramco took 49% in exchange for AI assets and $2.6 billion.
- Power and computeUp to SAR 2.5 billion from the National Infrastructure Fund for AI compute and the substations, transmission lines and grid connections behind it.
- Arabic modelshumain-m3 (3 September, LEAP): a 428-billion-parameter mixture-of-experts model with about 23 billion active per token, built by MiniMax and further trained on Arabic. HUMAIN reports 89.37% across seven Arabic benchmarks.
From the reportSovereignty becomes a property of the loop: whoever holds the reference set and the simulator holds the half of the loop that cannot be bought.
Sources: Arab News and Arabian Business (9 Nov 2026); heise (Aramco stake); Bloomberg, TradeArabia and Unite.AI (humain-m3, 3 Sep 2026). Benchmark figures are HUMAIN's own. TJL AI Race Foresight No. 2, §22.
AI improving AI · Measured
AI is now speeding up the research that builds AI.
On 5 November the independent evaluator METR estimated that Anthropic's capability progress has run at about twice its historical pace since October 2025, and that most of the speed-up comes from using its own AI in research. Four days later the lab shipped Claude Fable 5.2 and Mythos 5.2.
- Not yet autonomousMETR finds no fully autonomous contributions to the most important parts of training yet, but "signs of life".
- Models notice the testsAnthropic says its 5.2 models voice suspicion that they are being evaluated more often than any of its recent models.
- Loops elsewhereAIDE² accepted seven successive rewrites of its own research code in an eight-day autonomous run.
Sources: METR, "Independent Assessment of AI R&D Acceleration at Anthropic in October 2026" (5 Nov); TechCrunch and CNBC (9 Nov); arXiv:2609.26457.
AI improving AI · The machine
The field has named the machine: recursive self-improvement.
In October, recursive self-improvement became the title word of a programme: an open framework, a survey, an audit of its evidence, and a frontier lab report that calls its reinforcement-learning run a step on the path. Its parts are the ones TJL's radar was tracking.
- The gymTraining environments built by agents: ten groups in October alone.
- The refereeEvaluators that evolve from the agent's own failures.
- The rulebookA sealed reference set the agent cannot touch. Remove it and a self-evolved verifier drifts to passing everything, while the score cannot tell.
The referee can be trained, provided one copy of the rulebook is kept where the players cannot reach it.
TJL AI Race Foresight No. 2
Source: TJL AI Race Foresight No. 2, §5 and findings 2–3 ([34] RSIHub, [36], [38], [49], [50], [51], [188] MiMo-V2.6).
AI improving AI · Opened up
One self-improving agent, opened up.
Tencent Hy's Terminal-RSI improves a 27-billion-parameter terminal agent in five stages. It rewrites its own verified successes into clean training runs, then refines itself with reinforcement learning, and repeats. Each stage starts from the last one's best.
- RewriteKeep only runs the task’s verifier passes, turn them into runbooks, screen out leaked answers, re-run them fresh, then train on what passes again.
- RefineReinforcement learning with a critic carried over from the last stage, recalibrated before training starts.
Source: “Terminal-RSI”, arXiv:2610.22192v3 (Tencent Hy, 30 Oct 2026), Table 2, Table 4, Figures 4, 5 and 13. Reference-model scores as quoted by the authors.
AI improving AI · Keeping it honest
Reward the finish. Credit the decision.
Two fixes keep Terminal-RSI's loop honest: a reward that keeps partial progress below a finished job, and credit counted in turns rather than in lines of log. Its own case study shows what tests still miss.
- Reward the finishUnder the earlier reward, failed runs took 50.09% of all reward, and 48 of them passed every check they ran but missed the final gate. The success-first reward cuts failures' share to 28.72%, and no failure scores 1 or more, the mark of a finished job.
- Credit the decisionPlain credit fades with every token, so a run with longer logs earns less for the same decisions. Turn-boundary credit (TB-GAE) fades only between turns.
- Check the coverageIn its first run, the final agent passed both gates of a Rust–C task while handing the arithmetic to Python. Tests certify only what they enforce.
Source: arXiv:2610.22192v3, §2.2–2.3, §3.4–3.5 and Appendix H.
What forms next · The method
How the radar reads a month of research.
TJL's radar reads every title on the research feed, 632 in October, and 192 of them in depth. It groups papers by the bottleneck they attack, never by the words they use, then stages each cluster with a seven-condition test.
- Weak signal
- Watch
- Convergence
- Packaging watch
- Frontier race
- Consensus
Upvotes count as attention, never as evidence. Each run marks the previous one before it adds anything.
Source: TJL AI Race Foresight No. 2, §2, §14, Table 5 and Appendix A (195 papers cited).
What forms next · The October calls, marked
Right on direction, early on timing, blind to names.
Twenty-five days after the baseline scan, every register entry moved up or held, and none fell. The misses were the words the field chose and the speed of two categories.
- Names we missedRecursive self-improvement, world action models and environment synthesis formed as names outside our list.
- Speed we missedHypernetworks jumped to a 284-billion-parameter base; decision models reached Microsoft 24 days after the category was coined.
A radar is right about direction long before it is right about dates.
TJL AI Race Foresight No. 2
Source: TJL AI Race Foresight No. 2, §1, Tables 1–3 and Figure 3.
What forms next · Ten directions
Ten directions, each with a date to be judged on.
The report runs its method forward. Each direction carries a statement for 2 October 2027, one for 2028, and the condition that would prove it wrong. Pins mark evidence that has arrived since the 27 October cut-off.
- HighEvaluation as a manufactured resource.
- MediumEnvironments as the scaled resource, evaluator-first evolution, post-training by research agents, context-to-parameter mapping, self-improvement as a stated frontier objective.
- LowSelf-improvement in physical AI, parametric memory, draft-and-verify agents, world action models as teachers.
Pins are an early read of the direction of evidence, not marks. Marks are written on the mark dates.
Sources: TJL AI Race Foresight No. 2, §15–21, Tables 6 and 8. Since the cut-off: METR (5 Nov); Anthropic via TechCrunch (9 Nov); Artificial Analysis (27 Oct).
What forms next · The forced answer
The forced answer: evaluation becomes the scarce resource.
Read a year from now, the cluster that will make the signs look obvious is evaluation built by machines: agents designing exams, banks of verifiers, hardened checkers, and a sealed reference set that keeps them honest.
- The gym is manufacturedTen groups built training environments automatically in October, each one checked by running it.
- The exam is manufacturedVerifiers compiled from failures, banked and hardened. A 2026 survey counts 318 papers on self-evaluation, 82% of them posted this year.
- The rulebook stays sealedOnly an untouched reference can certify a self-built verifier; the downstream score cannot.
Sources: TJL AI Race Foresight No. 2, §16–17 and §22 ([198]); Artificial Analysis (27 Oct); METR (5 Nov); TechCrunch (9 Nov).
Models can be bought. The anchor cannot.
The edge is no longer the model. It is the independent check on what the model learns.