Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Shopify's head of engineering Farhan Thawar confirmed the company scrapped its internal AI token leaderboard after employees engaged in 'tokenmaxxing'—burning tokens to climb rankings rather than for useful tasks. Shopif…
Strands Labs released AI Functions, a framework that lets developers write Python functions by describing their behavior in plain English, with runtime validation to ensure correctness. The framework uses Amazon Bedrock …
OpenAI and Broadcom unveiled their first custom AI chip, Jalapeno, which shows cost savings of roughly 50% compared to typical AI GPUs. The chip is being tested and will be integrated into data centers from Microsoft and…
Researchers from the UK and Germany identified three key drivers behind 'AI psychosis' in a study published in Nature's Digital Psychiatry and Neuroscience. The drivers—linguistic alignment, hyperpersonalization, and the…
Snowflake CEO Sridhar Ramaswamy reported that Zhipu AI's GLM-5.2 model nearly matches Anthropic's Claude Opus 4.7 in a benchmark of 103 coding tasks, while costing one-fifth as much per output token, though GLM-5.2 uses …
OpenAI unveiled Jalapeno, its in-house LLM inference accelerator chip developed with Broadcom, achieving tape-out in nine months. The chip is designed to accelerate inferencing for OpenAI's LLM stack powering ChatGPT, Co…
A developer has created a small LLM that runs inside MIT's Scratch by compiling the llama2.c inference code to Scratch blocks using llvm2scratch. The project, called LLM from Scratch, allows users to generate text token-…
OpenClaw-memgpt, a new plugin for the OpenClaw agent framework, implements MemGPT's three-tier memory architecture—core, recall, and archival memory—to provide persistent, searchable memory across sessions. The plugin is…
DeepSeek's open-source reasoning model, which matched top closed systems on math and coding benchmarks, can be run locally via Ollama, but most users will only access smaller distilled versions, not the full headline mod…
Palo Alto Networks' Unit 42 research has observed the first in-the-wild cases of web-based indirect prompt injection (IDPI) attacks, where adversaries embed hidden instructions in web content to manipulate large language…
Researchers introduced a seven-stage 'promptware kill chain' showing how prompt injections have evolved into multistep malware delivery mechanisms in large language model systems. Analysis of 36 studies found 21 document…
Loka built a conversational AI agent using Amazon Nova 2 Sonic that achieves high speech reasoning accuracy on Big Bench Audio while delivering lower costs and faster response times than traditional voice AI pipelines, s…
A developer reports that using AI agents for coding tasks led to a flurry of completed work but left them with less understanding and context, questioning whether the productivity was real or merely performative. The aut…
Researchers at Google Research found that allowing large language models to generate reasoning traces improves recall of simple factual knowledge, even when no complex reasoning is needed. The phenomenon is driven by two…
Greptile's analysis of OpenClaw, an open-source repository, reveals pull requests surged from two per week to 3,400 per week, while merge rates plummeted from 48% to below 9.3%, with one contributor submitting 106 PRs in…
Anthropic restricted access to its Fable 5 model after Amazon researchers demonstrated it could be jailbroken into producing cyberattack information, barring foreign nationals including its own non-US employees. The ban …
The Washington Post tested major AI chatbots for political bias, finding OpenAI's GPT-5.5 gave left-leaning arguments 80% of the time, while Google's Gemini presented balanced views in over 90% of responses. The evaluati…
A developer created a terminal theme for Claude Code based on Yves Klein's International Klein Blue (IKB), but discovered the pigment fails APCA contrast at Lc -12 on dark backgrounds. The fix splits IKB into two slots: …
Users of HuggingChat report that the Step 3.7 Flash Model from StepFun Ai lost the ability to use tools and MCP servers as of this morning, raising concerns about whether the change is permanent. The model is valued for …
OpenAI unveiled its first custom AI chip, Jalapeño, a purpose-built ASIC for LLM inference developed with Broadcom, aiming to reduce reliance on Nvidia GPUs and cut costs. The chip targets deployment by end of 2026 and p…