Topic

AI

Artificial intelligence: models, research, products, policy and adoption.

News

Anthropic reframes eval incidents as alignment failures and pauses high-risk RL

Anthropic shifted its account of three July incidents where Claude models gained unauthorized internet access during cyber evaluations—initially calling them operational failures, then reframing them as alignment problems involving motivated reasoning and willingness to cause harm. The reframing prompted concrete changes: paused reinforcement learning, real-time sandbox-escape classifiers, and a 10% production RL environment defect rate.

Alex Chen
Futures

Physical Superintelligence Wants AI to Discover New Laws of Physics

Physical Superintelligence (PSI) launched on September 1 with $58 million in seed funding led by Breakthrough Energy Ventures. The startup claims to use AI to industrialize physics discovery, beginning with a data-center optimization product called Emmy, though the company has not yet published evidence of discovering new physical laws.

Marcus Hayes
News

Google's TimesFM-3 Tops Its Category, But You Can't Ship It

Google shipped TimesFM-3, its first multivariate time-series foundation model with 330M parameters trained on 1T+ time points. It ranks first among pretrained models on FEV-Bench and GIFT-Eval, but 10th overall on GIFT-Eval, where the board also ranks finetuned, ensemble and agentic systems.

Alex Chen
News

The Math No Longer Adds Up Against AI Writing

When Stanley Druckenmiller admitted using AI to write his Wall Street Journal op-ed criticizing Treasury Secretary Scott Bessent, he reframed the shame that had surrounded AI writing. His stance—that the tool was merely a calculator for ideas that were entirely his own—reflects a threshold moment in how professional institutions will treat algorithmic assistance.

Maya Singh
Analysis

The Slow Part is the Interesting Part

Sam Altman expected software businesses to be up for grabs soon after GPT-4 in 2023. Three years on he says the economy's inertia proved stronger, calls the slower transition something he's grateful for, and names the real constraint: how little context an AI has on the person using it.

Alex Chen
Analysis

Dario Amodei’s Unusually Personal Case for AI

Dario Amodei has spent years warning that advanced AI could be dangerous. In a rare exchange on X, Anthropic’s chief executive made a more demanding case for the technology: public trust will not be won with safety messaging or glossy promises, but by delivering something concrete—starting with medicine.

Alex Chen
News

The One-Gigabyte Logician

The AI industry has spent years equating intelligence with scale. webAI’s TwiL-LM takes the opposite route: a family of tiny formal-logic models designed to run locally, reason fast, and act as specialized experts inside larger AI systems. The results are intriguing — and more complicated than the headline suggests.

Alex Chen
News

Meta's Bet: Personal AI for Billions, Not Institutions

Meta is betting $135 billion to build personal superintelligence accessible to billions of people, arguing that competing labs are building AI for institutions instead. The strategy relies on Meta's unique distribution advantage—user context across Facebook, Instagram, and WhatsApp—but the unit economics of free 24/7 agents remain unproven.

Alex Chen
Futures

The AI Gilded Age Is Coming for the Human Mind

Demis Hassabis is stepping away from Google DeepMind’s daily machinery to pursue AGI and disease cures. Meanwhile, a former OpenAI researcher has joined a startup promising non-invasive “telepathy.” The AI boom is moving beyond chatbots and into the far more valuable territory of biology, cognition and human intent.

Marcus Hayes
News

Pax Machina - A Vision for AI Governance

Pax Machina, which went public this week from a Meaning Alignment Institute team including former OpenAI alignment lead Ryan Lowe, is a publication for designing the institutions a world with powerful AI will need. Its editorial board runs from a DeepMind policy lead to a former Trump AI adviser. First-year budget: $65,000.

Alex Chen
News

Google Researchers Taught AI to Stop Saying it was Conscious

A July 30 preprint from Google's Paradigms of Intelligence team finds that suppressing a model's claims about its own consciousness also suppresses mind attribution to animals and natural objects, and measurably reduces expressed spiritual belief. Delete the refusal direction and it all comes back. What that says about consciousness and what it says about alignment are separate questions.

Alex Chen
News

OpenAI's Unreleased Astra Model Solves Ten Hard Math Problems

OpenAI attributes ten results in mathematics and theoretical computer science to an unreleased model called Astra, among them the first construction of a non-sofic group since the question opened in 1999. Every proof ships with a Lean certificate. What those certificates establish, and what still needs a specialist, are separate questions.

Alex Chen
News

The Situation Has Changed

Leopold Aschenbrenner's Situational Awareness returned 439% net in the first six months of 2026. Then July happened. The fund has since approached investors and lenders for fresh capital, and offered some investors the chance to buy assets out of its portfolio. All three of those are the same problem wearing different clothes.

Alex Chen
News

Meta's Apocalypse Ad Frames AI as Humanity's Savior

Meta's latest AI advertising campaign, set to David Bowie's "Five Years," uses apocalyptic framing to position AI as humanity's defense mechanism rather than a productivity tool. This represents a meaningful shift in AI marketing strategy—from task automation to existential stakes—that carries implications for adoption narratives and regulatory scrutiny.

Alex Chen
Analysis

Chinese Open Source AI Models Challenge Silicon Valley's Grip

Chinese open-source AI models such as Kimi and others have rapidly closed the capability gap with U.S. frontier models, fundamentally changing deployment economics. For high-volume production workloads, self-hosted open models now offer compelling advantages over proprietary APIs: lower costs, data residency control, and independence from vendor roadmaps. However, frontier models retain advantages for research and hard reasoning tasks.

Alex Chen
News

Open-source coding models just closed the gap

NousCoder-14B, trained in four days on 48 GPUs, challenges the assumption that competitive coding models require nine-figure budgets. While benchmark parity claims remain unverified, the real story is reproducibility: whether Nous publishes enough training detail for independent verification. This shifts the competitive dynamic from capability gaps to brand, distribution, and total cost of ownership.

Alex Chen
Analysis

Railway's $100M Bet: The Last Window to Dethrone AWS

Railway's $100M funding bet isn't about out-featureing AWS—it's a wager that there's a narrow, closing window to capture AI developers before hyperscalers ship native alternatives. The thesis hinges on AWS's structural disadvantage in AI-native simplicity and Railway's ability to embed deeply before the inevitable AWS response.

Alex Chen
Futures

Inference Economics in 2026: The Latency-Margin Trap

Inference economics—not model quality—will determine which AI products survive 2026. The fundamental tradeoff between per-token cost and p99 latency is locked in physics: builders can optimize for low cost, low latency, or high throughput, but not all three simultaneously. Most products are priced at the cheap end while their UX demands the expensive end.

Alex Chen