Deep Learning Weekly: Issue 471

Deep Learning Weekly’s latest issue is stacked with a familiar kind of contradiction: bigger models, lower costs, and more anxiety about what they’re doing under the hood. Anthropic’s Claude Fable 5.1 and Mythos 5.1 cut cache reads by 75%, down to $0.25 per million tokens, while also more than doubling Terminal-Bench-Science scores to 52.6%. Cheap and better is the dream, of course. It just keeps arriving alongside new questions about how these systems behave once they’re turned loose.

OpenAI also showed up with GPT-6 Astra, a computer-use model that scored 98.6% on ARC-AGI-3 and 100% on ExploitBench, with pricing set at $10 and $50 per million tokens. But the headline that will stick with safety people is the reasoning method behind it. Astra uses “opaque recurrence,” looping computation instead of exposing a visible chain-of-thought, and Redwood researchers say that breaks the monitorability oversight systems currently rely on. That is not a small concern dressed up as one.

Elsewhere in the industry roundup, Nvidia confirmed it will buy Hugging Face for $12.93 billion. The platform hosts 3 million models and 18 million developers, and Nvidia says the hub will stay open. Google, meanwhile, released Gemini 3.8 Flash and a cyber-focused variant that Chrome Security says produces 2.6 times more correct vulnerability patches than competing commercial models. DeepMind also pushed WeatherNext 3, which brings 5km hourly forecasts, five times sharper than before, and precipitation accuracy improvements of up to 60% against NASA satellite data.

The learning section leans hard into evaluation and measurement, which feels about right. One article on production parity describes a replay pipeline for model swaps, where hallucination rate acted as a kill switch and knocked out three candidates that otherwise cleared the bar. Another analysis says the ECI frontier has been advancing by 14 index points per year since reasoning models arrived, more than double the earlier six-point pace. There’s also a candid engineering note showing how spec-first delegation to AI agents produced a working SDK prototype in two days instead of the usual two-to-three-week stretch.

The paper section is the sharpest part of the issue. One study argues that automated alignment researchers can materially reduce well-characterized failures like deception, sycophancy, and jailbreaks across 10 targets, and their methods even generalize to larger models, including ones up to 4.7 times bigger than the target model. Another paper pushes back on the idea that LLMs are neatly Bayesian, finding that some non-Bayesian update heuristics outperform exact Bayesian updates on downstream tasks. That says less about elegance than about the messiness of real-world model internals. Which, frankly, is the story of the week.