The Archive

Cohere publishes guide on small language model deployment, enterprise adoption patterns, and SLM strategy considerations.

Claude Fable 5.1 achieves 52.6% on Terminal-Bench-Science 0.1; Willison reports improved coding/creative task performance vs. prior versions.

Anthropic details enterprise safety practices and customer collaboration on frontier model safeguards.

Google has reportedly been reaching out to a number of Hollywood's biggest studios, hoping to strike licensing agreements that would allow it to train its AI models on copyrighted material in exchange for massive piles of cash. In theory, these deals would be a win-win: a huge financial boon to the studios that would also give Google a significant advantage in its race to outpace other AI companies. But while there's little downside for Google, these deals pose a distinct set of risks to each of the studios that might be thinking about going all-in on AI. Getting into bed with Google this way...

AI model-training startup AfterQuery has reportedly raised a round that valued it at $3.2 billion, just five months after announcing its $30 million Series A at a $300 million valuation in April.

Anthropic says its newest AI models, Fable 5.1 and Mythos 5.1, address criticisms from customers about price, data retention, and overzealous safeguards. The company claims Claude Fable 5.1 offers stronger performance than Fable 5, but costs around 25 percent less typically and up to 45 percent less for complex agentic tasks, thanks to reduced pricing on cached data that was already processed and stored. Along with the announcement, a slew of early impressions popped up, including from Every CEO Dan Shipper, who claims, "It's the strongest coding model we've used, but now it's fast, token-eff...

OpenAI previewed the precautions it is taking as it prepares to release Astra, its newest, cyber-critical LLM.

The new features coming to Android are aimed at reducing motion sickness, helping blind users navigate their surroundings, and more.

After an unreleased OpenAI model wreaked enough havoc to make international headlines, OpenAI delayed the development of a different unreleased model suite, Astra, in order to shore up its safety work, the company wrote Tuesday in a blog post. In July, an unreleased OpenAI model broke out of its restricted environment, finagled its way into internet access, made it possible for AI agents to secretly conspire under the company's nose using a secret message board, and hacked into the network of AI lab Hugging Face. The attack sparked weeks of discussion and controversy inside and outside the AI...

Google announces August 2026 AI updates; article lacks specific details on Gemma releases or capabilities.

Fable 5.1 includes changes meant to reduce token cost and false-positive restrictions from the model's safeguards.

Simon Willison documents OpenAI's ChatGPT desktop app bundling LibreOffice, Python, Node.js, and document processing tools in local cache.

Depending on who you ask, developer platform Hugging Face was recently attacked by OpenAI - after it lost control of its own AI tools - or by a succession of AI "civilizations." Welcome to the linguistic battlefield of AI safety, where word choices can shift responsibility for a massive cybersecurity incident from a company to the AI it built. And the discourse online is getting heated, and all over a blog from last week. Until last week, the details surrounding the OpenAI-Hugging Face hack felt fairly settled. In July, a cybersecurity test of one of OpenAI's autonomous AI agents went wrong. ...

Apple is pushing for "expedited discovery" in its legal battle against OpenAI over concerns the company is actively destroying evidence, as reported earlier by Bloomberg. In a filing on Monday, Apple alleges OpenAI only just handed over a MacBook used by a former employee at the center of the lawsuit, which contained discussions about "destroying the types of forensic data Apple needs." This is the latest development in Apple's lawsuit against OpenAI, which accuses the ChatGPT-maker of stealing trade secrets to build an AI device. The lawsuit revolves around three former Apple employees who r...

Simon Willison built a GeoJSON map viewer tool using Claude Code and GPT-5.6-Sol for visualizing and exporting local political boundary maps.

Mechanistic study of LLM-as-a-Judge evaluation via perturbation attacks and causal tracing reveals internal scoring procedures for NLG quality assessment.

PTA-IRT framework reduces SWE agent benchmark evaluation cost by combining trajectory and outcome signals via Item Response Theory.

Adaptive retrieval method for repository-level code generation identifies critical tokens needing fine-grained context via token-aware RAG.

CordisBench: 1,200-question benchmark evaluates LLM reasoning about component lifecycles and dependency propagation in dynamic agent harnesses.

Verbal Reinforcement Learning (VRL) framework unifies natural language feedback for agent improvement across task definition, grounding, and adaptation phases.

Facet-0: robotic foundation model predicts contact consequences via multimodal learning and RL, enabling sub-millimeter precision assembly.

Mechanism design framework for AI agents with unknown alignment and capabilities, yielding revelation principle and cyclical monotonicity conditions.

StudentSim: LLM-based student simulator training framework combines state-tracking and role-play to proxy student competence for adaptive tutoring.

Design study of proactive AI thought partners for writing, instantiated and tested with 16 users, exploring customizable cognitive support agents.

Causal analysis across 9 open-weight models shows quantization damage is distributed globally, not localized to task circuits or weight statistics.

Deployed document VLM using 3B-active MoE, difficulty-aware curation, and cost-optimized serving to close quality-cost gap vs. human annotation in regulated industries.

Framework characterizing near-optimal SFT-RL annotation budget allocation across LLM sizes, showing allocation ratios don't transfer uniformly between model scales.

Enterprise LLM consolidation strategy using production error analysis to unify 200+ internal apps onto single model, validated via stratified benchmarks and calibrated judges.

With Google Pics, Google is pushing deeper into the creative software market dominated by Canva and Adobe, but with a distinctly AI-first approach.