moment to moment, the front page of AI.

Industry
Big Tech's Race for the Classroom Isn't About Education
Anthropic's Teach For All deal is the latest move in a strategic competition to capture the teacher relationship. Whoever trains the teachers shapes which AI the next generation defaults to.
Go Deeper

Self-Attention: The Engine Behind Every Frontier AI Model

Self-attention is the core mechanism powering every frontier AI model. This explainer covers how Q/K/V vectors work, why multi-head attention captures richer patterns, how attention's parallelizability killed RNNs, and why the O(n²) cost creates a real opening for sub-quadratic alternatives. At frontier scale, no pure alternative has matched full attention; the field is converging on hybrid architectures that keep full attention where it matters and use cheaper primitives everywhere else.

How LLM Inference Works (And Why It's So Expensive)

LLM inference is two fundamentally different workloads sharing one GPU: compute-bound prefill and memory-bandwidth-bound decode. The KV cache makes autoregressive generation fast but expensive, and understanding it explains why context length, concurrency, and cost are inextricably linked.

What RLHF Actually Does to a Model

A technical breakdown of the RLHF pipeline and why safety alignment in large language models is far more fragile than it appears. Princeton researchers won an ICLR 2025 Outstanding Paper Award for showing that alignment concentrates in the first few output tokens, and just four gradient steps of fine-tuning can strip it away.

Tokenization: Why Your Prompt Costs What It Costs

Tokenization converts text into integer sequences that LLMs compute over, using Byte Pair Encoding trained on specific corpora. Vocabulary size is a business decision that affects cost and model size. Because tokenizers are trained on English-dominant data, non-English languages produce dramatically more tokens, leading to higher costs and measurably worse model performance, a structural fairness problem baked into the infrastructure.

Learn more
Recent
Policy & Business
Anthropic Poaches Microsoft's India Chief for Claude
Anthropic hires Microsoft India's former MD Irina Ghose to run its first India office, raiding enterprise AI talent in Claude's second-largest market.
Policy & Regulation
Anthropic's Trust Gets an Establishment Upgrade
Anthropic's Long-Term Benefit Trust swaps two EA-aligned founders for a former California Supreme Court Justice. What can the Trust actually do?
AI Policy
OpenAI Gates Cyber Tools Behind Identity Checks
OpenAI's Trusted Access for Cyber program gates GPT-5.3-Codex's offensive security features behind identity verification and use-case screening.
Feed
AI Research

GPT-5 Cuts Protein Synthesis Costs 40% in Lab Trial

Enterprise AI

OpenAI Frontier Takes Aim at Enterprise Agent Chaos

AI Models

Claude Opus 4.6: Anthropic Bets Big on Multi-Agent Teams

AI Development Tools

Claude Code Teams: What the Architecture Reveals

AI Research

Anthropic Rewrites Claude's Constitution as Philosophy

Business

Alphabet's Silence on Apple AI Deal Says Everything

Policy & Business

Anthropic Bets Against AI Ads, Altman Takes the Bait

AI Research

GPT-2 Training Now Costs Less Than a Pizza

Policy & Regulation

The UK Is Putting Claude Between Citizens and Their Benefits. The Stakes Are Not Theoretical.

AI Research

Anthropic Wants Claude in the Wet Lab. The Allen Institute and HHMI Will Test Whether That's Real.

AI Development Tools

Apple Made Claude the Default Coding Agent for 30 Million Developers

AI Infrastructure

OpenAI Says GPT-5.2 Is 40% Faster. The Interesting Part Is How They're Proving It.