Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Large language models have reached a hard limit in their ability to handle multi-step, coherent tasks due to their stateless, pattern-matching design, forcing industries to build external components for planning and memo…
OpenAI transformed ChatGPT into an enterprise-focused superapp with coding tools and AI agents, ending the chatbot-as-product era. New research quantified token usage in AI coding workflows, revealing code review as the …
OpenAI introduced Lockdown Mode to block outbound network requests that could exfiltrate sensitive data during the final stage of a prompt injection attack. The feature does not prevent prompt injections from appearing i…
OpenAI has introduced a Lockdown Mode for ChatGPT that restricts connected tools and integrations to prevent data exfiltration, responding to evidence that LLM-connected tooling serves as a viable vector for leaking sens…
AI providers are quietly reducing model quality and throttling capacity as demand outpaces infrastructure supply, according to an analysis by a senior director at AMD's AI group. The study of 6,852 Claude Code sessions f…
A speculative proposal argues that training massively overparameterized neural networks with extremely high learning rates on small, diverse datasets could trigger a "catapulting" effect, producing human-like generalizat…
Microsoft Azure CTO Mark Russinovich gave Anthropic's Claude Opus 4.6 a raw 6502 machine-language binary he wrote as a teenager in 1986 for the Apple II, with no source code or documentation. The AI decompiled the binary…
A new study from Teia Studio demonstrates that AI safety detection methods are fundamentally insufficient to prevent catastrophic LLM failures, with creator José Enrique Vásquez Valenzuela showing through documented stre…
Three new research papers provide converging evidence that AI use can degrade human cognitive skills through a phenomenon researchers call "cognitive surrender," where users adopt AI outputs without scrutiny even when th…
A senior AI reporter at Ars Technica was fired after publishing an article containing fabricated quotes generated by a language model, which the journalist failed to verify before publication. The incident highlights a c…
Ahmad Al-Dahle, former Apple executive and Meta AI leader, now serves as CTO of Airbnb where he is deploying AI to encourage real-world human experiences. In a new interview, Al-Dahle discusses his belief that artificial…
A software developer has proposed a new AI architecture called "context sculpting" that would allow language models to edit their own context window rather than treating it as an immutable, append-only conversation log. …
GitHub deprecated the GPT-5.2 and GPT-5.2-Codex models across most Copilot experiences on June 5, 2026, including Copilot Chat, inline edits, and code completions. The company recommended GPT-5.5 and GPT-5.3-Codex as alt…
A developer has compiled a library of 100 reusable system prompt patterns that work in production AI applications, organized by their function. The patterns include identity definitions with explicit exclusion clauses to…
A developer built Quincy, a local-first PDF reader for Mac that includes AI features for studying dense texts, after struggling to focus while reading "C++ Primer." The tool allows users to chat with pages, generate quiz…
Developers experienced outages across Claude AI, Cursor, and Amazon Q on Tuesday, with the downtime traced to Anthropic's API infrastructure buckling under high traffic loads and regional server failures. The disruption …
A developer built a three-tier content quality ladder for programmatic directory sites, using a single `model_used` column to track content origin across three sites launched in April. The system prioritizes Claude Haiku…
LinkedIn engineers Karthik Ramgopal and Prince Valluri detailed the company's use of multi-agent orchestration platforms to operationalize AI as a new execution layer across the enterprise. A separate developer demonstra…
Google's LiteRT-LM optimization technique delivers a 2.2x performance boost for Gemma 4 local inference on edge devices and consumer hardware, enabling faster multi-token prediction without cloud API calls. LinkedIn's pl…
CJ Zafir teased Mac-1, a 6.6 billion parameter AI model designed to run locally on Macs and operate native macOS tools, in a thread on X on Saturday. Zafir claims the model requires 7GB of RAM, runs best with 12GB, and p…