This week in deep learning, we bring you Introducing Claude Fable 5.1 and Claude Mythos 5.1, Evaluating LLMs Under Production Parity and a paper on Automated Researchers Can Reliably Mitigate Alignment Failures.
You may also enjoy OpenAI launches GPT-6 Astra, The ECI frontier has advanced by 14 points per year since reasoning models, a paper on LLMs Are Not (Consistently) Bayesian: Quantifying Internal (In)consistencies of LLMs’ Probabilistic Beliefs, and more!
As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.
Until next week!
Industry
Introducing Claude Fable 5.1 and Claude Mythos 5.1
Anthropic released Claude Fable 5.1 and Mythos 5.1, cutting cache reads 75% to $0.25 per million tokens and more than doubling Terminal-Bench-Science scores to 52.6%.
Welcome to the AGI era: OpenAI launches GPT-6 Astra
OpenAI launched GPT-6 Astra, a computer-use model scoring 98.6% on ARC-AGI-3 and 100% on ExploitBench, priced at $10/$50 per million tokens.
OpenAI’s new reasoning technique alarms AI safety experts
Astra’s “opaque recurrence” loops computation instead of visible chain-of-thought, which Redwood researchers warn destroys the monitorability that oversight currently depends on.
Nvidia confirms it will buy Hugging Face for $12.9 billion
Nvidia confirmed a $12.93 billion acquisition of Hugging Face, whose platform hosts 3 million models and 18 million developers, pledging the hub stays open.
Introducing Gemini 3.8 Flash and 3.8 Flash Cyber
Google released Gemini 3.8 Flash and a cyber-specialized variant that Chrome Security reports produces 2.6x more correct vulnerability patches than competing commercial models.
Google DeepMind’s WeatherNext 3 delivers 5km hourly forecasts, five times sharper than its predecessor, improving precipitation accuracy up to 60% against NASA satellite data.
MLOps/LLMOps/AgentOps
The Economics of Agent Optimization
A detailed blog post about context engineering for enterprise agents, reporting 54% better evidence recall, 34% lower retrieval token costs, and roughly 97% input-token reduction from tool search.
UX Without a UI: Notes from Researching Terminal UX for Opik MCP
A thoughtful blog post about designing terminal UX for LLM observability, covering six rendering constraints and “numbered spans” for referencing trace rows without copying.
What’s new in AI infrastructure and orchestration in August
A monthly roundup about AI infrastructure updates, including gVisor sandboxes on distributed Ray clusters, fully stateless MCP, and a customer cutting pipeline costs 90%.
A blog post about running Cursor Cloud Agents on customer-controlled machine pools with hibernation and snapshots; cloud agents now open over 60% of Cursor’s merged pull requests.
Learning
Evaluating LLMs Under Production Parity
A methodical article about a replay pipeline for safely swapping models, where hallucination rate acted as an eliminatory gate rejecting three otherwise-passing candidates.
The ECI frontier has advanced by 14 points per year since reasoning models
A data analysis about capability growth accelerating to 14 index points per year since reasoning models arrived, more than double the prior six-point rate.
AI didn’t write our SDK. It changed how we built it.
A candid engineering blog post about spec-first delegation to AI agents, which produced a working SDK prototype in two days versus a typical two-to-three-week effort.
From 1 Bloated Context Window to 6 Scoped Subagents
A hands-on course lesson about replacing one bloated context window with six scoped subagents, using semaphore limits, shared byte budgets, and per-child request caps.
How to extract meaning from charts and tables in PDFs
A practical guide about late-interaction multi-vector retrieval over chart-heavy PDFs, skipping OCR and chunking entirely in roughly 50 lines of Python.
What Is Math For in the Age of AI?
A panel discussion about what mathematics is for in the age of AI, covering AlphaProof’s olympiad results and concerns about vanishing training problems for graduate students.
How small AI models can make a big impact for enterprises
A blog post about enterprise model portfolios, citing a 30B/3B-active coding model that outperforms some 123B models on Artificial Analysis’ coding index.
Libraries & Code
An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.
TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.
Papers & Publications
Automated Researchers Can Mitigate Well-Characterized Alignment Failures
Abstract:
Automating alignment research may accelerate progress toward aligned AI, but whether it does is hard to measure. Luckily, many alignment failures, such as deception, sycophancy, and jailbreaks, are already measurable by public benchmarks. We study whether automated alignment researchers (AARs) can post-train to mitigate alignment failures by proposing training methods and data to simultaneously optimize multiple safety benchmarks, while largely preserving general capability. Across 10 alignment failures, the strongest AAR methods significantly reduce the targeted alignment failures and generalize to a held-out benchmark, multi-turn behavioral audits, and models up to 4.7× larger than the target model. As a human baseline, 28 experienced researchers receive up to eight hours to develop methods for the same benchmarks, but their methods underperform the best AAR methods. Using human ideas as the AARs’ initial research direction does not improve performance, suggesting current AARs may not need guidance from experienced researchers. These results suggest that automating alignment research on well-characterized failures may be practical in the near term.
Abstract:
Modern AI systems are being deployed in complex domains such as medicine, science, and law, where there is often not a single correct answer given the observed evidence. Such systems must be able to represent and update uncertain beliefs about the world as new evidence arrives to make rational decisions. We introduce the novel technique of studying LLMs as information processing rules and utilize the information processing gap—the deviation from Bayes updates—to study the internal (in)consistencies of how LLMs update their probabilistic beliefs from evidence. Our extensive experiments evaluate multiple approaches in which LLMs can incorporate evidence into their beliefs. Some of these approaches produce (nearly) Bayesian updates, thus optimally processing evidence; others use a learned heuristic. Surprisingly, the non-Bayesian heuristic updates often outperform exact Bayesian updates (optimal information processing) in terms of downstream task performance—indicating the LLMs’ probabilistic models of the world are misspecified. Lastly, we show how our measure can provide diagnostics to identify issues with LLM-powered inferential s
