The Signal — August 12, 2026
The gap between what an AI model shows you and what it actually does keeps getting more interesting. Today: a trick that pulls hidden reasoning out of frontier models, a leadership shuffle at Google meant to close the distance with its rivals, and a small model claiming it can reason for a fraction of the cost.
Researchers found a way to read what frontier models are secretly thinking
Frontier models generate hidden "reasoning traces" as they work through hard problems, and providers usually keep those traces out of view. A team from the University of Tübingen, the Max Planck Institute, MATS Research, and the security firm Snyk showed you can pull them back into the open. Feed the encrypted or hidden traces to a weaker model from the same provider and it tends to spit them out in plaintext. The method worked on models from OpenAI, Anthropic, and Google reached through their APIs.
Two consequences stand out. The researchers were able to recover personal data like passwords and API keys from a model's inner reasoning, a hole the providers have since patched. And the same technique enables reasoning-distillation attacks at scale, where one lab's expensive thinking gets copied on the cheap. The paper adds a spicier claim: Moonshot AI's open-weight Kimi K3 produced traces strikingly similar to Claude Opus 4.8 and GPT-5.6 Sol on certain prompts, which the authors read as suggestive, though not proof, that some Chinese models were trained on US ones. Two other open models, DeepSeek and Thinking Machines' Inkling, didn't show the same fingerprint. Worth holding the distillation angle loosely; the researchers say themselves it "cannot causally establish distillation."
Sources: Wired · arXiv · Techmeme
Google hands DeepMind and Gemini to one person
Koray Kavukcuoglu is taking over as head of Google DeepMind and will report straight to Sundar Pichai. His job covers Gemini model development, frontier research, and the Gemini app and developer teams, which puts the models, the product, and the people building on top of them under a single executive. That is a deliberate move as Google tries to close the gap with OpenAI and Anthropic, especially on coding, where it has been playing catch-up.
Forrester analyst Brian Hopkins framed the reshuffle as organizing DeepMind "like a product group rather than a lab," and argued that is something Google does better than its rivals. Whether that translates into better models is the open question, but the timing helps the narrative: it lands the same week Google said the Gemini app crossed a billion users.
A 150-million-parameter model claims frontier-adjacent reasoning for pennies
Pathway, one of the newer "neolab" outfits, unveiled BDH-CQ, a small reasoning model built on a post-transformer architecture, and put benchmark numbers behind its efficiency pitch. On ARC-AGI-1 it scored 29.5% pass@2 at roughly $0.0007 per task. OpenAI's cost-conscious GPT-5.6 Luna (Low) did better on accuracy at 34.5%, but even after an 80% price cut it cost about 11 times more per task. Slightly worse answers, far cheaper compute.
The architecture is the interesting part. Pathway says the model reasons without chain-of-thought and needs less training data thanks to improved memory, which is where the cost savings come from. This is early work with one lab's own benchmarks, so the numbers deserve independent replication before anyone rewrites the economics of inference around them. But the code is public, and a credible efficiency claim from outside the big labs is worth watching.
Sources: The Deep View · Pathway · GitHub
On the Editor's Desk
We skipped OpenAI's Daybreak cyber models landing on AWS and the Anthropic EU watermarking story today because we covered both in yesterday's edition, and neither had enough new development to run again. A German court's copyright ruling against the music generator Suno is a real story, but it is a few days old now with no fresh hook, so we're holding it for a better angle rather than running it stale.
