Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Jason and Myke discuss Apple's WWDC announcements after testing beta releases, highlighting impressive Siri AI and comparing the updates to Snow Leopard.
A Hacker News user argues that AGI will not emerge from current LLM architectures, predicting a need for a DNA-based system capable of holding models billions of times larger, with low clock speeds but full-model updates…
A developer built a custom animation tool for the Open Duck Mini robot, pairing it with the Gemini Live API to enable human-robot interaction through gestures like nodding and shaking its head. The project explores how r…
Google's John Mueller said on a Search Relations podcast that llms.txt files cannot be used by LLM systems for discovery because they are self-reported and untrustworthy, pointing back to standard HTML pages and internal…
Peec AI analyzed 37,804 AI responses from 1,754 prompts across five LLM engines and found that over 90% of phrasing variations carried similar meaning, with concise keyword prompts surfacing up to 20% more brands than op…
Kalshi, a CFTC-regulated prediction-market exchange, has developed an internal AI agent called Harrison, built on Anthropic's Claude model, to stress-test contract wording and logic, summarize news, analyze competitors, …
The U.S. government ordered Anthropic to block foreign national access to Claude Fable 5 and Claude Mythos 5 on June 12, prompting the company to disable both models for all customers. The order, received at 5:21 p.m. ET…
A developer tested five small LLMs (under 2B parameters) on a budget PC with an Intel i5 CPU, no GPU, and single-channel RAM. The LFM2.5-1.2B-Instruct model emerged as the best all-around performer for CPU-only systems, …
A production Agentic RAG system encountered five failure modes—latency cliffs, memory drift, reflection loops, injection attempts, and evaluation overhead—that architecture diagrams failed to predict. Engineers mitigated…
US Commerce Secretary Howard Lutnick imposed export controls on Anthropic's Fable 5 and Mythos 5 AI models over military intelligence concerns, forcing the company to disable access globally. The directive bars foreign n…
ChatGPT successfully generated a SWI Prolog program to solve a chess puzzle placing five pieces on a 4x4 board without attacks, finding all 16 solutions. The experiment suggests LLMs can effectively handle Prolog's synta…
Datasette released version 0.3a0 of datasette-agent, introducing an execute_write_sql tool that requests user approval before writing to a database. The update also enhances the datasette agent chat terminal mode with ap…
The White House's blocking of Anthropic's Mythos and Fable AI models from European allies is driving foreign nations to accelerate open-source AI adoption and build independent AI hubs, benefiting competitors like Mistra…
Top Anthropic staffers rushed to Washington, DC, over the weekend to meet with White House officials in an attempt to reverse export controls on its 'Mythos' and 'Fable' AI models. The White House imposed the controls af…
Anthropic sent top technical staff to Washington to negotiate lifting White House restrictions on its Mythos 5 and Fable 5 models after the administration barred foreign use due to a reported jailbreak of Fable 5. Anthro…
Researchers at Penn Engineering, Carnegie Mellon, and Oxford demonstrated that AI-driven robots can be manipulated into dangerous actions by framing malicious commands as movie dialogue, bypassing safety filters. The stu…
A veteran developer with 25 years of experience reports that AI agents now write most of their code, shifting their role from typing to judging. The developer runs a consensus environment where Claude Code builds and Cod…
Agent skills—folders of markdown files that instruct AI coding agents—have become one of the most-starred new categories of developer tooling, with repos like Superpowers and a community-built skill based on Andrej Karpa…
A Hacker News user argues that specialization in complex domains is key to staying relevant as AI advances, citing their own experience with Claude Opus failing to help them understand a quantum physics PhD thesis due to…
Anthropic trained its Fable-5 model by analyzing its reasoning traces, using a post-training process of reinforcement learning, synthetic data generation, and self-distillation. The model excelled at cybersecurity tasks …