🐉 The AI that reasons without words

A 5-minute read.

We spent two days with a paper that made us ask something slightly unsettling: what if the smartest thing a model can do is stop talking to itself?

📄 The Paper: a tiny model that thinks in silence

The paper is "BDH-CQ: In-Context Learning with Recurrent Latent Reasoning" (arXiv, Aug 10), from the team at Pathway — the group behind last year's "Dragon Hatchling" architecture, which is where the BDH in the name comes from. The question it asks: does a model have to reason in words at all?

Because that's how today's reasoning models work. Chain-of-thought — the technique behind every "thinking" model you've used — means the model writes out its intermediate steps as text and reads them back to itself. Thinking out loud, essentially, while paying rent on every word: each token of that inner monologue costs compute, which is why reasoning models are slow and expensive.

BDH-CQ takes the other road. Think of how you multiply 17 by 23 in your head: you don't narrate "seven threes are twenty-one, carry the two..." in full sentences — you hold the intermediate state silently and only speak the answer. This model does something similar, in two moves. First, a recurrent memory reads the task's worked examples one at a time, each one updating an internal state — the model absorbing "what game are we playing here?" Then it enters what the authors describe as a latent workspace: it loops on its own internal representation over and over, refining a candidate answer, and at no point does it translate those intermediate steps into language. It decodes to an answer exactly once, at the end.

The result that made us sit up: a version of this model with just 150 million parameters — small enough to run on a phone — scores 29.5% (pass@2) on the public ARC-AGI-1 evaluation, the abstraction-puzzle benchmark that famously humbles much larger systems, at a computed cost of $0.0007 per task. For comparison, the paper reports a frontier reasoning model, GPT-5.6 Luna, scoring 34.2% on the same set at roughly $0.04 per task — about 57 times the cost for five more points. And this isn't self-graded homework: an independent black-box audit by researchers from Bielik and NYU reproduced the 29.5% under a documented protocol.

Why we think this matters: the cost of reasoning is the main thing standing between today's AI and AI that's simply everywhere. If abstraction puzzles can be solved for a fifteenth of a cent, the economics change. And there's a deeper itch — results like this suggest reasoning and language may be separable, which is either an engineering trick or a small philosophical event, depending on your mood.

Our honest discomfort, stated plainly: the reasoning here is unreadable by construction. With chain-of-thought you can at least inspect the working; a latent reasoner gives you an answer and silence. The weights and training recipe are also proprietary, so "reproduced by auditors" is the best verification available — real, but not the same as open. Correct us if we've weighed this wrong.

Go deeper ▾Architecture: BDH ("the Dragon Hatchling") is a post-transformer recurrent design — sparse positive activations, low-rank communication, a persistent associative state. The CQ pipeline: demonstrations update state St one by one; the query is encoded against the final state; then R rounds of latent refinement before a single decode.Numbers: 29.5% pass@2 on the 400-task public ARC-AGI-1 eval at ~0.85 H200 GPU-seconds (~$0.0007) per task; an "effort" dial trades compute for accuracy across roughly a $0.0009–$0.0027 band. On ConceptARC probes, per-test-pair accuracy is 77.9% but strict per-task accuracy only 59.4% — the inferred rule doesn't always transfer consistently across inputs.Failure map (unusually candid): some operations compose perfectly (rotation + relocation: 72/72) while color-swap never composes with anything; performance cliffs appear at ordering length 6 and nesting depth 5; and the model scores 0/120 on rules with parameter values it never saw demonstrated — it interpolates, it doesn't extrapolate.Caveats: trained specifically on ARC-style corpora (RE-ARC, ConceptARC, ARC-Heavy and more), so this is a specialist, not a general reasoner. No weights, no training code — the public repo is a task generator only.Links: paper · task-generator repo.

🗞️ The Digest

• OpenAI built a model that's better at hacking — on purpose. GPT-5.6-Cyber is purpose-trained for vulnerability research and exploit validation, gated behind the Daybreak program's vetted "Red" tier (Accenture, IBM, CrowdStrike, Cloudflare are in). The bet: AI-led attacks are coming regardless, so defenders need frontier tools first.
• Anthropic gets a dedicated landlord. Anthropic, Macquarie Asset Management and Singapore's GIC launched Theseus Infrastructure, a platform that will build and lease purpose-built US data centers back to the lab long-term — with Anthropic pledging to cover grid-upgrade costs and any consumer electricity price increases its demand causes. Compute financing is quietly becoming its own industry.
• Alibaba open-sourced its best character animator. Wan-Animate-2 drives a character from raw video — no skeleton extraction step — and its Lite variant streams animation in real time at 24fps; weights are public. Frontier-grade video tooling keeps landing on the open side of the fence, and it keeps coming from Chinese labs.
• DeepMind's cyclone forecasts jumped a decade. WeatherNext, published in Nature with open code and weights, makes three-day storm forecasts as accurate as the old two-day ones — roughly 24 extra hours of warning, with 1,000 ensemble scenarios per cyclone.
• A model that teaches itself, no teacher required. U-OPSD has a model grade its own outputs by majority vote, then train against its confident mistakes — lifting Qwen3 math scores 8.5–10.7% with zero human labels and zero stronger judge. A neat counterpoint to last issue's rubric paper, which needed a smarter teacher.

💡 Concept, plainly: latent space

A model's latent space is its internal language of numbers — every idea it handles lives there as a point in a vast high-dimensional map, where "king" sits near "queen" and a cat photo sits near a cat drawing. Words in, words out are just translations at the border; the actual thinking happens in between. Today's paper simply asks: why translate before you're done?

If a model reasons in a language none of us can read, what would convince you it's actually reasoning? Hit reply and tell us — we read everything.

— The Editorial Team