22 Aug 2026

Digest: Ornith-1.5 open models learn from self-generated tasks, Wildeford's plan for the superintelligence scramble

1. Ornith-1.5: From Self-Scaffolding to Self-Improvement

   Ornith-1.5, a new open-weight family (9B dense, 35B and 397B MoE, built on Qwen3.5 and Gemma 4, announced on X), is trained through a closed self-improvement loop: the model proposes progressively harder tasks for itself, builds the scaffold and grading harness for each one, then learns by RL from its own solution rollouts — with rewards for task validity, near-frontier difficulty (targeting ~20% success), novelty, and the generated harness's resistance to reward hacking. The benchmark numbers are self-reported but notable — the 397B flagship matches Claude Opus 4.8 on Terminal-Bench 2.1 (86.1 vs 85.0) and SWE-bench Verified (86.0 vs 85.8), ahead of similar-size open models GLM-5.2 and DeepSeek-V4-Flash — yet the training recipe matters more than the scores: a model manufacturing its own RL curriculum and environments, exploiting the fact that proposing and grading a task can be easier than solving it, is a working small-scale instance of the automated-AI-improvement feedback loop that the recent run of recursive-self-improvement analyses has been trying to characterise.

   ornith.ai
2. DeepSeek-v4-flash-vision-exp

   DeepSeek added vision to its fast-tier model line, shipping deepseek-v4-flash-vision-exp through its API: the experimental model accepts images alongside text — screenshots, charts, documents — and Bloomberg reports the company claims performance nearing Anthropic's frontier Opus 4.8 on visual tasks, a claim that is self-reported and unverified, and worth discounting given Flash is DeepSeek's small/cheap tier rather than its flagship. It still matters: screenshot understanding is the prerequisite for computer-use agents, and this is the leading open-weight lab extending its unusually fast release cadence (V4 Pro exited preview only days ago) into multimodality.

3. The Scramble: getting in position to pace the frontier

   Peter Wildeford of the Institute for AI Policy and Strategy argues that the moment a US government gets serious about superintelligence will look less like a multi-year treaty negotiation and more like the Cuban Missile Crisis — a chaotic weeks-long 'scramble' followed by an interim deal built on a voluntary Pacing Charter among frontier labs, verification via tools governments already trust (spies, satellites, inspections, export data, scrutiny of the few RSI-capable data centers), and an Operation Warp Speed push toward a durable US–China agreement. The provocative claim for the verification community: much current work on high-assurance mechanisms like cryptographic proof-of-training is mis-sequenced, since resources will explode a hundredfold once a scramble starts — what's scarce now is scramble-ready tooling, a map of the existing intelligence toolkit, and the memo for the emergency White House meeting.

   Institute for AI Policy and Strategy via The Power Law (Peter Wildeford)
4. Evaluating Explanations of LLM Behavior In The Wild with Counterfactual Experiments

   A new pipeline called CHIVE automatically discovers unexpected LLM behaviors in real-world usage and explains them with counterfactual prompt edits — and using it as an evaluation produced a notable negative result: agents given activation-reading interpretability tools predicted the outcomes of these counterfactual experiments no better than agents that simply read the transcript. That is a useful corrective datapoint against the recent run of wins for internals-based oversight (activation oracles, deception probes), suggesting today's activation-reading tools add little for explaining in-the-wild behavior even as the same data proved useful for training models to predict behavioral effects; it is a single team's early result, not yet peer-reviewed.

   Adam Karvonen via LessWrong