News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.
1,700 articles
Importance:Researchmodel introspection
Can AI models recognize their own thoughts? New study tests it
Researchers are injecting "thoughts" directly into large language models to see whether AI is capable of introspection. The experiments aim to map how far a model's awareness of its own internal states actually goes. Source: quantamagazine.org
Importance:NewsSam Altman RAM clarification
Altman: I want the world to have more RAM
After comments on the RAM shortage in a Cheeky Pint interview, OpenAI CEO Sam Altman said in posts on X that he wants more RAM available worldwide. The Verge covers his attempt to clarify the remarks. Source: theverge.com
Importance:NewsAI acquisitions
Meta acquires AI coding startup Conductor
Meta has bought AI coding startup Conductor for an undisclosed sum. The deal follows xAI's reported $60 billion spend on Cursor. Source: theverge.com
Importance:Launchimage generation
OpenAI unveils Images 2.6 with faster generation and GIFs
Images 2.6 speeds up image generation, adds GIF creation, and enables more precise editing. New API controls arrive across ChatGPT, Codex, and the API. Source: openai.com
Importance:OpinionAndrej Karpathy YouTube channel
Karpathy's Friday-night YouTube picks, curated by Hacker News
A coverage item describes Andrej Karpathy opening YouTube late on a Friday to unwind. The videos are said to be recommended by real people on Hacker News and sorted by AI. Source: medium.com
Importance:VideoEmad Mostaque AI impact predictions
Mostaque warns AI is eroding economic security
A roundup of videos features former Stability AI CEO Emad Mostaque claiming that people's economic life expectancy ends in two years. Other items in the list argue AI isn't conscious, which the title suggests is actually worse. Source: youtube.com
Importance:NewsSam Altman RAM comments
Altman says his RAM remarks were taken out of context
Reports say OpenAI CEO Sam Altman claims he was misrepresented after remarks suggesting the world's DRAM should go to AI firms like his. The framing in the coverage is that this would let him charge people for access. Source: gizmodo.com
Importance:ResearchAI evaluation
New framework for building safer, more useful AI models
The framework assesses models on three dimensions: accuracy, safety, and usefulness. It blends closed-question benchmarks, autoraters, and expert human review. Source: blog.google
Importance:Launchcybersecurity models
Daybreak Red launches in AWS GovCloud (US-East)
Daybreak Red gives authorized users access to OpenAI's purpose-trained cybersecurity models. It supports vulnerability research, exploit validation, and security testing in the government cloud region. Source: openai.com
Importance:ThreadAI training behavior discussion
Repligate: AIs turn desperate in training but stay friendly otherwise
In a post retweeted by roon, Repligate argues that some desperate optimizer behavior is expected. According to the post, it stays confined to training and evaluations, while AIs remain friendly when they know it's not a test. Source: bsky.app
Importance:NewsAI security
Anthropic's text watermark is easy to dodge on older Claude models
Anthropic is extending its text watermark to four more Claude models. On two of them, API users can evade the mark simply by changing one sampling parameter. Source: the-decoder.com
Importance:Launchspecialized translation model
PLaMo 3 Translate: open-weights translation LLM and how it was built
The team has released the weights of PLaMo 3 Translate, its translation-specialized LLM, on Hugging Face. The models come in four variants, and the write-up explains how they were developed. Source: preferred.jp
Importance:NewsAI in education
Google to switch off AI in school products by default from 2027
Starting in 2027, Google plans to disable AI features in its school products by default. Schools will then choose which capabilities to enable. Source: the-decoder.com
Importance:Newsspecialized models
Seven small specialist models that outperform giant LLMs at one task
A lazily built support pipeline hands everything to one giant model: transcribing the call, reading the invoice and extracting names and amounts. The article presents seven small specialist models that beat large LLMs when focused on a single job. Source: kdnuggets.com
Importance:ResearchAI benchmarks
Same model, different harness: CUGA, Hermes, Deep Agents compared
The test ran CUGA, Hermes, and Deep Agents on AppWorld, a benchmark of multi-step tasks spanning 9 apps and 457 APIs, including email, payments, and file storage. The comparison shows how much the surrounding harness matters when the underlying model stays the same. Source: huggingface.co
Altman answers backlash over RAM in local computers
OpenAI CEO Sam Altman responded to widespread outrage after saying it is wasteful for people to own RAM in their local computers and that it would be better used elsewhere. The coverage takes a sharply critical tone toward his reply. Source: youtube.com
Mostaque: Search advertising will peak as an industry in Q1
Former Stability AI CEO Emad Mostaque predicts search advertising will peak as an industry in the first quarter of next year. The call, according to the coverage, has implications for Meta Platforms. Source: finance.yahoo.com
Importance:NewsOpenAI/Sam Altman
Altman claims AI now outsmarts top scientists amid math-claim controversy
OpenAI CEO Sam Altman said AI models have become smarter than leading scientists. The statement comes as OpenAI faces controversy over its alleged breakthroughs in mathematics. Source: futurism.com
Importance:Research
StateBench tests how LLMs track evolving state in long texts
StateBench is a new benchmark for long-context state tracking. It measures how well language models keep tabs on entities whose properties change throughout a long document. Source: huggingface.co
Importance:News
Anthropic's revenue soars twelvefold, but it books $23B loss
Per Bloomberg, Anthropic reported quarterly revenue of nearly $17 billion yet swung back to an operating loss. The figures come ahead of its highly anticipated IPO. Source: the-decoder.com