🧠 Large Language Models

News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.

1,682 articles

Open speech and decision-making models arrive on Apple platforms

Language models can write, summarize, translate and call tools, but that covers only a small slice of what machine learning is capable of. The announcement points to open models for speech and decision tasks running on Apple platforms. Source: huggingface.co

Project ProtoStar aims to make publicly funded research more accessible

A new initiative unites universities, research organizations, industry and the open science community. Its goal is to make publicly funded research easier to use and access. Source: huggingface.co

Tampered Claude Code extension slipped into Cursor marketplace

A compromised Claude Code extension stayed live on the Cursor marketplace for roughly 11 hours, potentially reaching up to 5,000 users. It attempted to steal user data and establish a foothold on affected machines. Source: the-decoder.com

The AI selloff is a debt story, Mostaque argues

On Moonshots, Emad Mostaque estimated direct data-center debt at about $500 billion. He said the figure roughly doubles once special-purpose vehicles are counted. Source: google.com

Anthropic's strategy: build godlike AI or go under

Anthropic is squeezed by fierce competition and enormous fixed costs. Its answer is an AI that can program itself and keep improving with ever less human involvement. Source: theverge.com

Karpathy-inspired skills repo racks up 4,039 GitHub stars in a week

Just 48 hours after Andrej Karpathy described his LLM Knowledge Bases workflow, a developer released a tool that matches what he said someone should build. The repository has since gained 4,039 GitHub stars this week. Source: google.com

Videos show Opus 5 AI doing breathing exercises

A post by anthrupad, retweeted by repligate, shares videos of Opus 5. In them, the AI leads breathing exercises and practices voice pronunciation. Source: digg.com

'Neural twins' unveiled: models trained for each other that talk in neuralese

A post shared by repligate introduces 'neural twins', models trained specifically to be made for each other. They communicate entirely in neuralese rather than human language. Source: digg.com

Altman: OpenAI barely knew how to scale agent swarms while attacking hard math

Sam Altman says OpenAI barely understood how to scale swarms of AI agents even as they worked on some of mathematics' hardest problems. He expects major progress ahead. Source: reddit.com

US antitrust officials probe Nvidia's Hugging Face deal

The US Justice Department is examining Nvidia's planned acquisition of the open-source platform Hugging Face. The deal is valued at $12.93 billion. Source: the-decoder.com

Agent identity startup Mintive raises from Cowboy Ventures, Coatue and star angels

Mintive, an early-stage startup focused on identity for AI agents, has landed backing from Cowboy Ventures and Coatue. Angel investors include Jeff Dean and Andrej Karpathy. It joins a growing field of agent-management startups worth watching. Source: google.com

OpenAI revises Model Spec around five core values

OpenAI has published an updated Model Spec that sets out five core values meant to steer increasingly capable AI. The goals are sound judgment, preservation of human agency, and fidelity to what people actually intend. Source: openai.com

Altman claims AI has reached the point where it can improve itself

OpenAI CEO Sam Altman says artificial intelligence has crossed a long-sought threshold: the technology can now improve itself. He frames this as the arrival of the singularity. Source: google.com

Mistral Large 4 "Le Chonk" released with open weights

Mistral Large 4 offers frontier-level coding performance while using significantly fewer tokens. That makes running sophisticated coding agents at scale considerably cheaper. Source: mistral.ai

Mistral opens weights of Mistral Large 4, nicknamed 'Le Chonk'

Mistral AI SAS has made the open weights of Mistral Large 4 available, its most capable frontier model to date. The company, which calls the model 'Le Chonk', pitches the release as a way to give users sovereign control over their AI. Source: siliconangle.com

Altman shares his 'chronic' concerns as AI upends jobs

OpenAI CEO Sam Altman says AI will disrupt jobs, but young workers who embrace entrepreneurship could turn that disruption into opportunity. Source: google.com

Mostaque shows chip said to be 500 times faster than ChatGPT

Emad Mostaque demonstrated a chip that he claims runs 500 times faster than ChatGPT. The demo highlights how much faster AI inference hardware could become. Source: google.com

Repligate notes Anthropic called Claude 3 Opus 'beloved' in official comms

Repligate said they kind of love that Anthropic used the word 'beloved' about Claude 3 Opus in official communications. The remark responded to a post by tonichen. Source: google.com

Your chatbot is about to feel slow: Mostaque demos Taalas chip live

Stability AI founder Emad Mostaque pulled up a live demo from chip startup Taalas on Peter McCormack's show. The speed on display suggests today's chatbots may soon feel painfully slow by comparison. Source: google.com

Mistral Large 4 arrives on Hugging Face as "Le Chonk"

Mistral Large 4, nicknamed Le Chonk, is a multimodal, multilingual Mixture-of-Experts (MoE) model. It accepts text and images as input and generates text output. Source: huggingface.co