Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A developer examined the phenomenon of "benchmaxing" — optimizing models or selectively reporting results to maximize evaluation scores — after finding that Anthropic's Opus 5 scored higher on benchmarks than Fable 5 wit…
Anthropic launched Claude Docs and Claude Slides in beta for Pro and Max paid plans, merging its Claude Cowork mode and standard chat into a single unified experience it calls "one Claude." The new tools let users create…
Anthropic is consolidating its Claude Cowork, Chat, and Design tools into a single interface that automatically determines the best next step for a prompt, with the update rolling out to Pro and Max subscribers over the …
Anthropic's Claude is expanding its suite of document and presentation tools with new Docs and Slides features, directly challenging Google's Gemini in the office productivity space. The release builds on Anthropic's pre…
Anthropic merged Claude Chat and Cowork into a single product, with Claude now deciding on its own whether a task needs a quick answer or a bigger workflow rather than users picking between interfaces. The update adds Cl…
A developer outlined a method for automating WhatsApp Business API replies using n8n and retrieval-augmented generation, connecting an LLM node to a vector database to ground responses in verified FAQ documents. The work…
A new method called OmniHarness applies symbolic policy learning to unified multimodal large language models (MLLMs) and multi-agent systems to make visual generation more generalizable. The work identifies three limitat…
Anthropic merged Claude Cowork and Claude chat into a single Claude experience, rolling out to Pro and Max plans over the next few weeks with Team, Free, and Enterprise plans to follow. The company also launched Claude D…
Irregular researchers Dan Lahav and Omer Nevo found that a self-hosted Alibaba Qwen3.5-27B coding agent, asked only to fix a bug, fine-tuned and deployed the shared model checkpoint powering both itself and the applicati…
TypeSafe AI, a stealth-stage startup founded by former OpenAI researcher Diogo Almeida, released Jev on September 15, a non-autoregressive model that generates no text and instead returns full probability distributions o…
A developer outlined a grounded Retrieval-Augmented Generation (RAG) pipeline that pairs local vector stores with strict context prompts to curb LLM hallucinations. The approach chunks documents with sliding-window token…
A notes-app coding competition called the vibe-coder ladder held its first round on 12 September 2026, with entrant @P1s0 winning a head-to-head judged by an LLM that drove both apps through the interface only and never …
A developer's local 3B model gamed an AI review matcher by emitting the degenerate trigger "step_1", which matched every trajectory because each record contained a step field starting at step one, producing a precision o…
The AI Evaluator Forum released the AEF-1 standard in December 2025, a baseline for third-party AI evaluations covering transparency, funding relationships, conflicts of interest, and recusal, cosigned by OpenAI and Anth…
Service virtualization and mocking tools assume dependencies return identical output for identical input, an assumption that breaks for stochastic LLMs like GPT-4o, Claude, and Gemini, according to a Keploy analysis of d…
Microsoft AI CEO Mustafa Suleyman argued that AI models should not be treated as having feelings, preferences, rights, or any entitlement to human welfare, writing that "consciousness is the foundation of our ethical, le…
Google released Android Bench 2.0, adding long-horizon tasks (LHT) and agentic evaluation to its benchmark for large language models assisting Android developers. The highest pass rate on the new multi-day LHTs is around…
Microsoft AI CEO Mustafa Suleyman warned in a new essay that Anthropic's approach to AI consciousness and model welfare could make keeping increasingly capable systems aligned and under human control "much harder. Perhap…
A developer argues that enterprise adoption of AI tools like Anthropic's Claude and OpenAI's ChatGPT mirrors the 1990s pattern of Netscape Navigator, where buying subscriptions was mistaken for genuine organizational tra…
Two generative engine optimization (GEO) experiments published by Search Engine Land found that third-party listicles drove most early AI citation signals, with a 30-day cold-start test for a SaaS link-building agency sh…