🧠 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

Nasdaq slides as AI stocks sell off; OpenAI revenue and hedge fund rumors in focus

The TBPN show covers a Nasdaq drop driven by an AI stock selloff, alongside talk of OpenAI's revenue inflection and rumors of a hedge fund blowup. The discussion also asks why hyperscalers are holding steady. Source: youtube.com

Anthropic's $1.5 trillion IPO meets investor pushback

Anthropic's planned IPO is hitting trouble as investors resist the $1.5 trillion valuation. A broader sell-off in tech stocks is adding to the pressure. Source: the-decoder.com

Can LLMs judge how smartly autonomous vehicles behave in traffic?

A new framework turns autonomous-vehicle interaction data into structured language so an LLM can give expert-like assessments. The results are then mapped onto a five-dimensional evaluation. Source: eurekalert.org

Using large language models to talk to dolphins

Denise Herzing, founder and director of the Wild Dolphin Project, began studying dolphin communication more than 40 years ago. The video explores how LLMs might be used to understand and speak with dolphins. Source: google.com

Study evaluates an LLM tool that makes breast cancer notes clearer for patients

LLMs could make clinical notes easier for patients to understand, but such tools need to be tested with patients themselves. This mixed-methods evaluation examines an LLM-powered note augmentation tool for breast cancer patients. Source: nature.com

Knowledge graph framework aims to secure LLMs in smart grids

A knowledge-grounded security assessment framework improves risk quantification and cross-domain access verification for LLMs used in smart-grid systems. Source: google.com

Paper proposes framework for generative AI anomalies using Jungian projection

An academic analysis offers a framework for understanding odd generative AI behavior through Jungian projection. It references the cyborgism community and independent researchers such as Janus/repligate, who use tree-based methods to explore models. Source: philarchive.org

Cult.fit builds operational muscle with Codex

Cult.fit uses Codex to help teams across functions build their own tools and move faster. The tool is changing how work gets done at the company. Source: openai.com

Harness adapters bring safe RL to agentic development

Harness adapters can make local agentic development more efficient and accurate. In smoke training tests on held-out eval tasks, they made fewer tool calls. Source: huggingface.co

How training methods shaped the evolution of large language models

The piece traces the progression from GPT-1 in 2018, a 117M-parameter transformer trained to predict the next token across roughly 7,000 books. That simple objective proved effective and set the pattern for later models. Source: sociable.co

ElevenLabs lets you post AI music straight to SoundCloud

SoundCloud is partnering with another generative AI music service, ElevenLabs. Users can publish their generated tracks directly to the platform. Source: theverge.com

YouTube video urges viewers to try Herdr, with little detail in the snippet

A video titled 'you need to use Herdr RIGHT NOW!!' promotes a tool called Herdr. The snippet contains mostly unrelated sidebar content and offers no details about the product. Source: youtube.com

Phonon TTS: high-quality text-to-speech that runs on-device

Phonon is a small text-to-speech model that runs entirely on your own device. It needs no server or network connection, so what you type stays private. Source: huggingface.co

Stochastic thermodynamics for autoregressive models from a non-Markovian view

A new general framework defines entropy production for autoregressive generative models. It builds a bridge between these models and stochastic thermodynamics. Source: google.com

Why general-purpose LLMs only take brokerages part of the way

Large language models are trained on broad internet-scale data, but day-to-day brokerage work calls for more focused AI tools. Source: google.com

OpenAI unveils GPT-6 Cyber for vulnerability hunting and fixes

OpenAI has introduced GPT-6 Cyber, a cybersecurity-focused model built to find exploitable vulnerabilities and develop patches. It is trained to operate strictly within its authorized scope. Source: openai.com

Gemini Omni 1.2 Flash adds 15-second video and cuts prices by 30%

Gemini Omni 1.2 Flash now generates and edits video natively at up to 15 seconds, with improved consistency. Google is also lowering pricing by 30%. Source: blog.google

Karpathy reportedly builds a website in two hours with Claude

A video circulating online shows Andrej Karpathy using Claude to put together a website in roughly two hours. The snippet offers little detail beyond the video title and its comments section. Source: youtube.com

Tencent Launches Hy4, Its New Open-Source LLM Aimed at the Global Top Tier

Tencent has announced Hy4, its latest-generation large language model, describing it as a significant jump in performance and capabilities. The company positions the open-source release among the world's leading open AI models. Source: google.com

Karpathy's 14-minute talk pitched as better than most AI courses

A post promotes a 14-minute session with OpenAI co-founder Andrej Karpathy, claiming it explains how ChatGPT and Claude work better than most AI courses. The claim is promotional and unverified. Source: google.com