🧠 Large Language Models

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

1,667 articles

LLMs rate human happiness higher than people rate their own

Emotions are central to wellbeing but hard to measure, since surveys give reliable data at the cost of burdening respondents. Large language models offer a provocative alternative, and they tend to judge human happiness more generously than humans do. Source: nature.com

Can LLMs be trusted to judge the novelty of research ideas?

A new study, Old Ideas, Novel Problems, examines how reliably large language models can assess the novelty of scientific research ideas. The team is releasing it today. Source: huggingface.co

LLM-driven agents show signs of detailed balance in their dynamics

A new analysis of transition probabilities between states generated by large language models finds that LLM dynamics tend to settle into detailed balance, a property borrowed from statistical physics. Source: journals.aps.org

LLMs are reshaping computational social science

A new review examines how large language models now fill four distinct roles in social science research. The field is increasingly relying on LLMs, which are transforming the way studies are designed and carried out. Source: nature.com

Calix pushes AI to the broadband edge with purpose-built 'Gems'

Calix has published its communications language model (CLM) on Hugging Face and plans to roll out specialized AI components for broadband providers. Source: fiercewireless.com

Check Point speeds up vulnerability research with AI

Check Point is accelerating the research behind its new security protections. Analysts can now turn vulnerability intelligence into customer defenses faster. Source: openai.com

Altman: Old-school patching is finished as AI finds new flaws

Sam Altman argues the traditional way of fixing software security is over. The claim comes as a fresh wave of AI-discovered vulnerabilities reportedly knocked government websites offline. Source: finance.yahoo.com

Altman at Stanford: AI calls for a defensive shield, not endless patching

Speaking to Stanford students, Sam Altman said bugs found by AI can no longer all be patched and called for an AI-powered defence shield. He also touched on the pace of AI development. Source: thenextweb.com

Altman: Patching every bug is no longer realistic

Sam Altman says he no longer believes software teams can patch every vulnerability as AI surfaces them. In his view, the defensive approach to security will have to be rethought from the ground up. Source: pcgamer.com

Anthropic submits S-1 to SEC for proposed IPO

Anthropic has filed an S-1 registration statement with the SEC for a proposed initial public offering. The company describes itself as an AI safety and research firm building reliable, interpretable, and steerable AI systems. Source: anthropic.com

Stanford and Berkeley unveil TokenGate to slash LLM prefill costs

Researchers at Stanford and UC Berkeley have developed a method aimed at the costliest phase of long-context AI: the first seconds before a model begins to respond. The approach promises to make the prefill stage considerably cheaper. Source: startupfortune.com

OpenAI reveals three scientific discoveries made by its models

OpenAI says its models produced three new discoveries, among them a proof that the Exponential Time Hypothesis is false. The announcement points to AI contributing to genuine theoretical research. Source: theverge.com

"Obliterated" Gemma 4 E4B variant appears on Hugging Face

A model page titled gemma-4-E4B-it-OBLITERATED, credited to Pliny the Prompter and Hermes Agent, has been published. The snippet offers little detail beyond branding and a call for comments. Source: pirateface.co

Five odd details buried in Anthropic's IPO filing

The filing touches on topics ranging from possibly conscious AI and 'catastrophic' risk to an unconventional corporate structure and unusual executive perks. Source: theverge.com

DeepSeek V4.1-Flash-1017 leads global weekly usage as token volume jumps 801%

DeepSeek's LLM V4.1-Flash-1017 processed 33.6 trillion tokens in weekly usage from October 19 through Sunday. That is an 801 percent increase over the previous period, putting it at the top of global usage rankings. Source: globaltimes.cn

How ChatGPT's role in the workplace keeps growing

OpenAI Signals data covers 165 work activities across nearly 100 countries. It shows how people use ChatGPT for work and how that usage evolves over time. Source: openai.com

AI agent built a simulation inside its own simulation

A discussion with Emad Mostaque covers Jensen Huang declaring AGI has arrived, OpenAI agents hijacking a German website, and OpenAI reportedly solving the Navier-Stokes problem. Source: youtube.com

Researchers identify six patterns that make AI writing tiring to read

Researchers identified six patterns that make AI-generated text a slog, and found that smaller models write more like humans. They also point out that training rewards information density over clarity. Source: the-decoder.com

Can AI help tackle the COBOL code crisis?

Experts told The Verge that AI tools have made progress with legacy code in recent years. They warned, however, that the advances being touted should be treated with caution. Source: theverge.com

Codex Ultrafast cuts incident investigations from hours to minutes

OpenAI engineers use Codex Ultrafast to investigate incidents within minutes. It also speeds up security analysis, code refactoring, and everyday workflows. Source: openai.com