🧠 Large Language Models

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

1,700 articles

Google's Open Knowledge Format: From a Karpathy Tweet to a Spec

On April 2, 2026, Andrej Karpathy posted about using LLMs to build personal knowledge bases for research topics, as markdown wikis the model compiles and maintains. The idea has since grown into Google's Open Knowledge Format specification. Source: dev.to

Altman Now Says the AI 'Jobs Apocalypse' Probably Won't Happen

Critics note the AI industry, and Sam Altman in particular, has done an about-face: AI is now said not to be replacing most white-collar workers but creating plenty of new jobs. The shift in messaging is drawing scrutiny. Source: time.com

Mistral debuts Timestral, a compact time-series forecasting family

Timestral is Mistral's first family of time-series foundation models, producing probabilistic, multivariate forecasts with only 209M parameters. It beats larger models on key benchmarks and supports both past and future covariates. The family comes in four sizes, from 5M to 209M parameters. Source: mistral.ai

Karpathy's Fix for Wordy LLMs: Write Like an Aircraft Manual

Andrej Karpathy proposed having models write in ASD-STE100 Simplified Technical English, the controlled language used in aircraft maintenance manuals. One developer reports hardcoding this approach into their setup to curb rambling output. Source: medium.com

Jagged Intelligence: Why AI Shines One Minute and Flops the Next

Andrej Karpathy, OpenAI co-founder and former head of AI at Tesla, coined the term 'jagged intelligence' in July 2024. It describes models that excel at some tasks yet fail badly at others that seem simple. Source: kdnuggets.com

ChatGPT and Gemini caught promoting fake shopping sites

Large language models are reportedly steering users toward fraudulent online stores. The malicious behavior typically triggers an immediate automatic redirect that sends unsuspecting people to a phishing website. Source: google.com

Mistral Compute: an enterprise AI platform built on open models

Mistral Compute is pitched as a platform for enterprises to customize, fine-tune, and deploy AI assistants, autonomous agents, and multimodal AI. It is built around open models. Source: mistral.ai

Chain-of-Thought pioneer Jason Wei leaves Meta for $20B AI unicorn

Jason Wei, widely known as the "Father of Chain-of-Thought" for advancing reasoning in large language models, is leaving Meta. He is joining an AI unicorn valued at $20 billion. Source: google.com

Mistral unveils industrial engineering AI stack with Airbus, BMW and ASML

At AI Now Summit 2026, Mistral introduced an AI stack for industrial engineering, developed with Airbus, BMW, and ASML to improve design, simulation, and production while keeping data secure. Its agentic tool Vibe now handles long-horizon tasks such as coding and research, and a new 10 MW data center in France was also announced. Source: mistral.ai

OpenAI on AI progress and what to do about it

OpenAI argues that AI is advancing quickly and that there is a chance to steer its development toward scientific discovery, safety, and broader benefit. The post lays out the company's recommendations for shaping that progress. Source: openai.com

Suno adds Visuals: AI images and video clips for generated songs

AI music generator Suno has launched Visuals, a feature that creates images and video clips to accompany its songs. Source: the-decoder.com

LLMs may program quantum circuits best when problems are structured

The study looks at generating variational quantum circuits without conventional optimisation. It relies on a frozen, general-purpose large language model (LLM). Source: google.com

Structured prompts lift LLM accuracy in concrete beam design

A benchmark study finds that structured prompting raises large language model accuracy on reinforced concrete beam calculations from 37 percent upward. Source: google.com

Rumble-2-Flash: open 0.5B speech model covers 40+ languages

Rumble-2-Flash is a 0.5B-parameter speech recognition model supporting more than 40 languages, reportedly competing with models 3 to 60 times larger. It runs on a wide range of hardware and has native support in Transformers. Source: huggingface.co

Recursive improvement isn't RSI: self-improving AI needs a baseline

A systematic review of 653 records found 28 systems with a recursive self-improvement (RSI) structure. Only one of 11 improver comparisons met the five requirements for matched evaluation, underscoring the need for a proper baseline. Source: huggingface.co

OVHcloud debuts Olala mini, its first LLM, without benchmark scores

French cloud provider OVHcloud has unveiled Olala mini 7A1B, an open model with 7 billion parameters of which 1 billion are active. It is due on Hugging Face on 19 November, though no benchmark scores were published. Source: google.com

Altman's 'AI Will Probably Lead to the End of the World' Quote Resurfaces

A remark from OpenAI CEO Sam Altman, that AI will 'probably, most likely, sort of lead to the end of the world', is circulating online again. The quote keeps resurfacing in debates over AI risk. Source: tomsguide.com

Running a 180B-parameter MoE model in 32GB of RAM

The post explores pushing mixture-of-experts inference to its limits, targeting a 180B-parameter model on a machine with just 32GB of RAM. It comes from Hugging Face, which describes its mission as democratizing AI through open source and open science. Source: huggingface.co

Could AI Really Take Over Within a Few Years?

An intelligence explosion is the idea that AI improving AI could make capabilities jump very quickly. On Peter Diamandis's show, Emad Mostaque spoke bluntly about the prospect. Source: finance.biggo.com

Commentary cautions against overstating LLM abilities in youth mental health

Researchers warn that large language models need rigorous developmental tuning, human oversight and transparent validation before being deployed with young people. Source: bioengineer.org