What we're reading, in the order we're reading it.
Anthony and Harry's working stream — links, charts, tweets, and short takes. The unprocessed inputs behind the briefings.
There are 32 different ways AI can go rogue, scientists say — from hallucinating answers to a complete misalignment with humanity
New research has created the first comprehensive effort to categorize all the ways AI can go wrong, with many of those behaviors resembling human psychiatric disorders.
Why Anthropic’s Copyright Settlement Changes the Rules for AI Training
Anthropic’s settlement came after a mixed ruling on the “fair use” where it potentially faced massive piracy damages for downloading millions of books illegally. The settlement seems to clarify an important principle:...
Researchers Unlock 210% Performance Gains In Machine Learning With Spin Glass Feature Mapping
By harnessing the complex dynamics of simulated magnetic materials, researchers have created a new data processing technique that boosts the performance of machine learning models by up to 210% on...
Understanding LLMs: Insights from Mechanistic Interpretability
Introduction Since the release of ChatGPT in 2022, large language models (LLMs) based on the transformer architecture like ChatGPT, Gemini and Claude…
Lawsuit Links CA Teen’s Suicide To Artificial Intelligence
The parents of a California teenager who committed suicide sued OpenAI, alleging that ChatGPT taught him how to harm himself.
ChatGPT reportedly linked to first murder — here’s what we know and what OpenAI is saying
More bad news linked to AI
New Paper Finds Evidence That AI Is Already Killing the Job Market
A recent paper uncovered some of the first comprehensive evidence that the AI industry is throwing the job market into flux.
Doctors develop AI stethoscope that can detect major heart conditions in 15 seconds
Upgraded medical tool can diagnose heart failure, heart valve disease and abnormal heart rhythms
How Sakana AI’s new evolutionary algorithm builds powerful AI models without expensive retraining
M2N2 is a model merging technique that creates powerful multi-skilled agents without the high cost and data needs of retraining.
Are AI and Democracy Compatible?
Avoiding digital dictatorship may mean rethinking our relationship with machines.
What is nano-banana? Google’s new image editing model
Plenty of AI image models have made things a lot more impressive, and Google’s groundbreaking image model, Gemini 2.5 Flash (also known as nano-banana because AI people), is now making...
Microsoft’s signals shift from OpenAI with launch of first in-house AI models for Copilot
MAI-Voice-1 and MAI-1-preview models will be used to power consumer applications.
AI could already be stealing jobs from young US workers
Entry-level workers look to be most affected
With new in-house models, Microsoft lays the groundwork for independence from OpenAI
Microsoft is still deeply tied to OpenAI, but who knows what the future holds.
This Team Is Rethinking AI’s Core: Perforated AI Bets On Dendrites
For more than seventy years, deep learning has relied on a simplified model of brain function. Now, a Pittsburgh startup thinks the AI field is due for an update.
An artificial intelligence cloud platform for OCT-based retinal anomalies screening system in real clinical environments
npj Digital Medicine – An artificial intelligence cloud platform for OCT-based retinal anomalies screening system in real clinical environments
Anthropic to Collect Your Chats, Coding Sessions to Train Claude AI
In search of more data, Anthropic is asking users to ‘help improve Claude’ by providing the AI with access to their chatbot activity. It will also retain data for five...
The White House reportedly ordered xAI’s Grok to be approved for government use
The White House ordered the GSA to add Grok to its list of approved vendors after some back and forth regarding the governments use of the Elon Musk owned AI...
60 U.K. Lawmakers Accuse Google of Breaking AI Safety Pledge
PauseAI organised an open letter from UK lawmakers and civil society organisations to Demis Hassabis, CEO of Google DeepMind. PauseAI UK members emai…
The DeepSeek Hype Was All Made Up
Our disinformation research team found thousands of fake accounts fueling a coordinated campaign to sway markets during the DeepSeek media frenzy.
How AI Is Solving America’s $1 Trillion Manufacturing Labor Crisis
America faces 2.1 million unfilled manufacturing jobs by 2030, costing $1 trillion annually. See how AI startups like Laborup are solving the crisis.
Deploying DeepSeek with PD Disaggregation and Large-Scale Expert Parallelism on 96 H100 GPUs
DeepSeek is a popular open-source large language model (LLM) praised for its strong performance. However, its large size and unique architecture, which us…
Microsoft Introduces 2 In-House AI Models Amid Rising Competition
Microsoft is reducing its reliance on OpenAI for Copilot. One of its two new models powers text-to-audio generation, while the other improves text-based tasks.
Subliminal Learning Lets Student AI Models Learn Unexpected (and Sometimes Misaligned) Traits from Their Teachers
AI can transfer strange qualities through seemingly unrelated training—from a love of owls to something more dangerous
Alibaba Creates AI Chip to Help China Fill Nvidia Void
Chinese chip companies and artificial-intelligence developers are building up their arsenal of homegrown technology, backed by a government determined to win the AI race.
AudioStory: Generating Long-Form Narrative Audio with Large Language Models
Recent advances in text-to-audio (TTA) generation excel at synthesizing short audio clips but struggle with long-form narrative audio, which requires temporal coherence and compositional…
Healthcare AI today: Bad diet advice from LLMs, ‘Dawn’ working on cancer vaccine, Generative AI earns $50M, OpenAI’s healthcare push
OpenAI, the developer of ChatGPT, has made a couple of big hires as it prepares for a major expansion into healthcare.
Microsoft Launches Public Testing of First In-House Foundation Model
Microsoft AI (MAI) has begun public testing of the first foundation model it trained in-house. The model, dubbed MAI-1-preview, is being tested on
UCR pioneers way to remove private data from AI models
A team of computer scientists at UC Riverside has developed a method to erase private and copyrighted data from artificial intelligence models—without needing access to the original training data.
Study shows AI can hone in on genetic risks for 10 inherited diseases
Researchers say they’ve trained artificial intelligence models to assess genetic risks for 10 inherited diseases on a spectrum rather than simply “yes/no.”
Lawsuit links CA teen’s suicide to artificial intelligence
(The Center Square) – The parents of a California teenager who committed suicide sued OpenAI, alleging that ChatGPT taught him how to harm himself, according to a lawsuit the parents
In crowded voice AI market, OpenAI bets on instruction-following and expressive speech to win enterprise adoption
OpenAI’s new speech model, gpt-realtime, hopes that its more naturalistic voices would make enterprises use more AI generated voices in applications.
Lawsuit links CA teen’s suicide to artificial intelligence
The parents of a California teenager who committed suicide sued OpenAI, alleging that ChatGPT taught him how to harm himself, according to a lawsuit the parents filed Aug. 26. Matthew...
Spending on AI data centers could run into the trillions of dollars by the end of the decade
Building them could soon be a bigger business than building traditional office space.
AI really is cutting out entry-level jobs for human workers, study claims
New research paints a bleak picture for younger workers
Elon Musk accuses of Apple and OpenAI of stifling AI competition in antitrust lawsuit
Elon Musk on Monday targeted Apple and OpenAI in an antitrust lawsuit alleging that the iPhone maker and the ChatGPT maker are teaming up to thwart competition in artificial intelligence.