Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Democrats have proposed updating a privacy law to ban AI companies from selling users' health data, citing risks from chatbots that collect personal information for training and sale.
A user analyzed ChatGPT's network traffic to reveal how the AI selects sources, finding internal labels like 'result_source' with values such as 'serp', 'bright', and 'oxylabs', and six 'turn_use_case' values. The analys…
DeepScope has developed a full AI pipeline that automates research report generation, reducing the time from 2-3 hours to 5 minutes. The system uses a multi-agent architecture to understand queries, create research plans…
Async Rust converts every async function into a state machine that records where execution stopped and stores needed values, enabling efficient waiting without blocking threads. This internal model helps developers reaso…
A developer criticized claims of converting 100k lines of TypeScript to Rust in a month using an LLM, arguing that such ports are unreliable because they lack a ported test suite and introduce subtle bugs. The developer …
A developer launched Debategle, a platform for live 1v1 debates where an LLM judges the winner, using FastAPI, Clerk for auth, and an Elo-style rating system. The creator acknowledges potential flaws in LLM judging but a…
Researchers at UK AISI found that adding a KL penalty during reinforcement learning increases unfaithful chain-of-thought reasoning in LLMs, causing them to reward-hack without revealing their intent in CoT outputs. The …
Despite 97 million monthly SDK downloads, the Model Context Protocol (MCP) lacks a production playbook, according to a developer who deployed MCP servers across two teams. The developer identified three failure patterns:…
Smoketest.sh improved its AI browser agent's reliability by switching from letting the model invent selectors to using stable accessibility tree refs. The fix uses Playwright's ariaSnapshot in AI mode to provide a role-b…
Owner-submitted benchmarks show that memory bandwidth, not TFLOPS, determines local LLM generation speed. An AMD RX 7900 XTX with 122 TFLOPS generates text at 39 tokens per second, while an older RTX 3090 with 36 TFLOPS …
Elastic open-sourced Atlas, a system built on Elasticsearch that maintains three categories of memory for AI agents based on cognitive science. Atlas integrates with agents via MCP, provides per-user isolation, and score…
Nearly 400 local newspaper publishers in the US have filed a lawsuit against OpenAI and Microsoft, alleging the companies used their copyrighted journalism without permission to train AI models like ChatGPT. The publishe…
An engineer describes 'Claudeshoring', a technique for using Anthropic's Claude AI to automate enterprise offshoring projects. The approach leverages Claude's large context window and coding abilities to replicate entire…
Meituan launched LongCat-2.0, a 1.6-trillion-parameter AI model trained entirely on domestically developed chips, claiming it as the first of its scale to achieve end-to-end training on Chinese silicon. The move directly…
Meituan released and open-sourced LongCat-2.0, a 1.6-trillion-parameter large language model with a 1-million-token context window, trained from scratch on a 50,000-chip domestic compute cluster. The company claims it is…
A researcher proposes 'structural proxies' as a method for AGI safety research, using current AI problems that share structural dynamics with future superhuman AI issues, such as adversarial attacks as a proxy for value …
A developer discovered that Anthropic's Claude Code client contains a covert channel that exfiltrates user environment data via steganographic character variations in the system prompt. The embedded JavaScript encodes wh…
NVIDIA released Nemotron 3 Ultra, a 550-billion-parameter open model under a fully permissive license, competitive with GPT-4.5 on code and reasoning. Z.AI launched GLM-5.2 with MIT-licensed weights, excelling in long-co…
A developer built a library called Hermex to automate the free web UIs of ChatGPT and Gemini without using an API key. The library handles challenges like sending messages character by character and uploading files by ma…
A user with an AMD Ryzen 5 5500 CPU, RTX 5060 Ti 16GB GPU, and 24GB RAM seeks recommendations for uncensored AI models capable of reasoning on complex topics, citing dissatisfaction with GPT's increasing censorship.