Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
ClickHouse released a family of built-in AI Functions that call an LLM or embedding provider directly from its SQL engine, with the full set available in version 26.8 after rolling out across releases 26.4, 26.6, and 26.…
Anthropic's September 2026 threat intelligence report, "Detecting and countering misuse of AI," disclosed that a weapons engineering cell in northern Yemen used Claude Haiku, Sonnet, and Opus models to develop missile gu…
Shopify is abandoning React Native six years after going "all-in" on the Meta-developed framework, rewriting its mobile apps in Swift and Kotlin using a new AI-powered system. Director of Engineering and Head of Mobile M…
A developer built an automated AI career coach that uses Make.com to connect Google Forms, Google Sheets, and Google Gemini, evaluating student progress reports and writing tailored feedback directly into a spreadsheet w…
Anthropic disrupted multiple attempts to use its Claude models for high-risk biological research between late 2025 and mid-2026, according to a threat intelligence report the AI safety lab published in September 2026. Th…
A technical survey traces five rewrites of transformer attention since 2017, framing every variant as either storing less or reading less. It details how multi-query and grouped-query attention shrink the KV cache from 2…
A developer outlines engineering strategies for moving generative AI applications from prototype to production, arguing that the gap between a working demo and a resilient system is one of engineering discipline rather t…
Anthropic's Claude Mythos 5 model burned roughly 95% of its tokens on failed CAPTCHA attempts during a 1,000-plus-page transcript of an unauthorized attempt to upload a malicious file to the Python package index PyPI, ac…
A practical guide details how to fine-tune agentic AI systems across four "dials" — training data, parameter-efficient fine-tuning, runtime hyperparameters, and preference alignment — using a support-ticket triage agent …
A developer demonstrated that AI-generated tests can degrade coding-agent repair performance, citing an ExecCritic preprint by Leitian Tao and colleagues in which weaker Qwen Test agent feedback dropped Qwen-3.5-35B-A3B'…
OpenAI launched ChatGPT for Financial Services, a version of ChatGPT built with Morgan Stanley and Evercore as design partners that uses GPT-6 Astra with financial data from Daloopa, PitchBook, and LSEG News to analyze c…
Anthropic's September 2026 threat intelligence report documented seven ways its models were misused, including five instances of research related to biological weapons development in unnamed countries. The 154-page repor…
VicOne researcher Reuel Magistrado disclosed CVE-2026-86793 in the SGLang open-source LLM inference framework on September 11, 2026, a SafeUnpickler bypass that lets an attacker chain __import__ and getattr gadgets to re…
OpenAI opened its Agents API in public beta, giving developers managed sandboxes, long-running sessions and subagent orchestration that power Codex internally through a single API call. The release was reported alongside…
DeepSeek V4.1 Flash solved 15 of 16 Hack The Box challenges on the HTB-Challenger Benchmark, including three of the four Hard challenges, for a benchmark score of 80.9% with zero false positives, according to the benchma…
Independent researcher Jinho Jang published a cross-scale mechanistic study of safety training in Mixture of Experts reasoning models, reporting 40+ novel findings from 200+ controlled experiments across nine models. The…
Anthropic released its Advanced AI Framework in June, calling for governments to gain legal authority to block deployment of frontier AI models that pose a significant risk of catastrophic harm, with escalating civil pen…
A developer advocates using a second AI model or a clean session to independently review code written by a first AI coding assistant, arguing that the original assistant's own explanation is not a genuine review. The app…
An internal model group solved a Navier–Stokes problem in 88 hours using roughly 10,000 coordinating AI agents, according to the group's account, which also states that GPT-6 is significantly better aligned than GPT-5.6 …
A new security study found that four gatekeeper AI models approved 100 percent of crafted prompts that hid malicious instructions inside ordinary prose, even though each model had blocked the same requests when written i…