|
📖 Read In Depth
|
Microsoft and OpenAI end their exclusive and revenue-sharing deal
Microsoft and OpenAI have ended their exclusive partnership and revenue-sharing arrangement — a structural shift with major downstream implications. OpenAI gains freedom to work with other cloud providers (Google TPUs now plausible), while Microsoft loses its moat on frontier model access. This reshapes the competitive dynamics for every enterprise AI platform.
hn/Best Stories
|
Is OpenAI Falling Further Behind in the A.I. Race?
OpenAI has reportedly missed its own internal user and revenue targets, raising questions about its path to IPO and the sustainability of its massive compute commitments (including the Oracle deal). With Gemini and Anthropic eating into its lead, the piece examines whether OpenAI's monetization model is structurally sound or whether the hype is running ahead of the fundamentals.
nyt/Business
|
Mozilla Used Anthropic’s Mythos to Find and Fix 271 Bugs in Firefox
Mozilla used Anthropic's Mythos agent to find and fix 271 bugs in Firefox — a concrete, high-credibility case study of AI-assisted software engineering at scale on a real production codebase. Worth reading for the engineering details: what kinds of bugs were caught, how the agent was deployed, and what still required humans.
reddit/r/singularity
|
Talkie: a 13B vintage language model from 1930
Talkie is a 13B language model trained exclusively on pre-1931 text (260B tokens), released by Nick Levine, David Duvenaud, and Alec Radford. The scientific motivation is sharp: it's a controlled experiment to study whether LLMs generalize knowledge or merely memorize it, by testing whether the model can reason about events it was never trained on. The methodology and findings are worth engaging with seriously.
hn/Best Stories
|
Alignment Makes Models More Decisive Without Making Them More Truthful
A research finding that RLHF-style alignment makes models more decisive (lower entropy outputs) without making them more truthful — essentially confidence without calibration. This is a substantive critique of the alignment-as-safety narrative and has direct implications for anyone thinking about when to trust model outputs.
reddit/r/singularity
|
How South Korea Uses A.I. to Check on Its Elderly
South Korea — the world's fastest-aging society — is deploying AI systems to make welfare check calls to elderly people living alone and to flag early signs of dementia. This is a substantive case study in AI applied to a real social infrastructure problem, with details on how the system works and what the measurable outcomes have been.
nyt/Top Stories
|
|
⚡ FYI
|
DeepMind's David Silver just raised $1.1B to build an AI that learns without human data
David Silver, the AlphaGo/AlphaZero architect at DeepMind, has raised $1.1B to build an AI that learns without human data — essentially scaling the self-play paradigm beyond games. If this works at general reasoning tasks, it's a significant architectural bet against the RLHF/SFT paradigm that dominates current frontier models.
reddit/r/singularity
|
GitHub Copilot is moving to usage-based billing
GitHub Copilot is dropping its flat subscription model for usage-based billing, with model multipliers ranging from 1x to 27x depending on which model you use (Opus is 27x). The era of subsidized inference is ending — this will change how teams budget for AI coding tools and likely accelerates migration to direct API access via OpenRouter or similar.
hn/Best Stories
|
China blocks Meta's acquisition of AI startup Manus
China has moved to block Meta's acquisition of Manus, the AI agent startup, even though Manus already moved operations to Singapore after closing its China offices. The legal grounds are murky — Manus has no China operations — but this signals Beijing is willing to use regulatory tools to block Western firms from acquiring Chinese-origin AI talent and IP regardless of where the company is formally based.
hn/Best Stories
|
White collar employment is sharply declining: The number of the S&P 500 employees fell -400,000 in 2025, to 28.1 million, posting its first annual decline since 2016.
S&P 500 total headcount fell by ~400,000 in 2025 — the first annual decline since 2016. The data point is striking in context: it's happening concurrent with record AI capex and productivity claims. Whether this is AI-driven displacement, post-pandemic normalization, or macro effects from the Iran war is the key interpretive question.
reddit/r/singularity
|
Iran War Shakes Global Economy, but the U.S. Has Mostly Been Spared
Eight weeks into the U.S.-Israeli strikes against Iran, the global economy has taken significant damage — energy supply shocks, airline bailout requests, China's manufacturing slowing — while the U.S. has been relatively insulated so far. Good orientation piece for understanding the macroeconomic backdrop that's running underneath all the tech news.
nyt/Business
|
United Arab Emirates Says It Will Leave OPEC in Blow to Oil Cartel
The UAE is leaving OPEC, a blow to the cartel's cohesion at precisely the moment the Iran war is creating the largest energy supply shock on record. The UAE has long chafed at production quotas it viewed as underselling its capacity; this exit weakens OPEC's ability to coordinate output and could accelerate oil market fragmentation.
nyt/Top Stories
|
FDA approves first gene therapy for treatment of genetic hearing loss
The FDA has approved the first gene therapy for genetic hearing loss — a meaningful milestone in the broader arc of gene therapy moving from rare disease proof-of-concepts toward more common conditions. The underlying science (likely OTOF mutation repair) is worth a brief look for anyone tracking the genetics-as-medicine transition.
hn/Best Stories
|
4TB of voice samples just stolen from 40k AI contractors at Mercor
4TB of voice samples from ~40,000 AI data contractors were stolen from Mercor — one of the largest breaches of AI training data to date. The irony noted in HN comments is sharp: the suggested mitigation is to send your voice to another AI company to check if it's already in circulation. Raises real questions about data governance in the AI training supply chain.
hn/Best Stories
|
5 Tall Tasks for John Ternus, Apple’s Next C.E.O.
Tim Cook is out, and hardware chief John Ternus is taking over as Apple CEO. The five challenges laid out — AI strategy, China dependence, post-iPhone growth, regulatory pressure, and maintaining the hardware premium — are a useful framework for thinking about whether Apple can remain a platform or risks becoming a premium device brand without a sticky ecosystem story.
nyt/Technology
|