Scott Alexander reframes hallucinations. Fowler coins 'supervisory engineering.' Meta cuts 20%.
Jensen Huang doubled Nvidia's chip forecast to $1 trillion through 2027 at GTC yesterday, with 'inference' as the new keyword. Stripe revealed their AI agents ship 1,300 PRs a week on infrastructure built for humans, not models. Also: Scott Alexander has the sharpest take on hallucinations in months, Martin Fowler named the engineering role that's quietly replacing traditional dev, and Ben Thompson argues agents are why this isn't a bubble.
Jensen Huang at GTC yesterday: Nvidia now expects over $1 trillion in AI chip sales through 2027, up from a $500 billion forecast. The driver is what Huang calls an 'inference inflection,' with agentic workloads creating sustained demand. Nvidia also showed off the Groq 3 LPX inference server rack for trillion-parameter models.
Read full story→Reuters reports Meta plans layoffs affecting 20% or more of staff, driven by $600 billion in planned data center spending through 2028. Zuckerberg has told employees AI will let individuals do what previously required large teams. The cuts would be Meta's biggest since the 2022-23 'year of efficiency' round.
Several xAI co-founders including Zihang Dai and Guodong Zhang have left after Musk brought in 'fixers' from SpaceX and Tesla to audit the company. The trigger: Grok's coding product falling behind Anthropic and OpenAI. xAI is still recruiting and merging data with X, but morale is reportedly low.
ByteDance suspended the worldwide rollout of its video-generation model Seedance 2.0 after Disney, Warner Bros., Paramount Skydance, and Netflix sent legal demands over unauthorized use of copyrighted characters and celebrity likenesses. Viral clips of Brad Pitt fighting Tom Cruise and a Darth Vader vs. Deadpool lightsaber duel forced the issue. Engineers are now adding guardrails, but the new filters are already rejecting legitimate prompts from paying Chinese customers, and enterprise access requires a 10 million yuan ($1.45M) minimum spend.
Stripe's internal AI agents ('Minions') ship 1,300 pull requests weekly. The real story isn't the models. Stripe already had isolated, parallelizable dev infrastructure: pre-warmed 'devboxes' and a 'blueprints' orchestration system. That foundation is what makes unattended agents reliable. Human engineers have shifted from writing code to reviewing it.
Simon Willison breaks down what coding agents actually do under the hood: an LLM processes chat-templated prompts, extracts tool calls from its response, executes them, feeds results back, and loops. The model is stateless between turns; the system prompt and tool definitions are what give it continuity.
Scott Alexander argues 'hallucinations' misframes the problem. LLMs aren't broken when they confabulate; they're doing exactly what training optimized for: finding the most probable next token, even when that means guessing confidently. The better analogy is a student who always fills in an answer rather than leaving it blank.
Ben Thompson argues the compute boom isn't a bubble because agents represent a genuine third inflection after ChatGPT and reasoning models. Agents integrate models with 'harnesses' to perform complex tasks autonomously, demanding more compute and making integrated providers like Anthropic and OpenAI more durable than the bubble narrative suggests.
Martin Fowler introduces 'supervisory engineering,' where developers direct, evaluate, and correct AI agent output in a 'middle loop.' The shift isn't from writing code to not writing code. It's from writing code to defining acceptance criteria, evaluation filters, and knowing when to regenerate a component rather than patch it.
Geometric analysis of model internals shows LLMs don't just fail to retrieve the right answer when hallucinating. They actively suppress the correct token when factual accuracy conflicts with contextual coherence. The implication: hallucination detection likely needs to be domain-specific, since the suppression signature varies.
Geoffrey Huntley argues China releasing advanced models for free or near-free isn't generosity. It's economic warfare designed to undermine massive US investment in proprietary AI, drawing parallels to Linux undercutting Microsoft. Whether intentional or not, the cost advantage of open models is real and the dependency questions are just starting.
MCP tool definitions eat a large chunk of your agent's context window before it does anything useful. Apideck CLI takes a different approach: progressive disclosure, loading tool info on demand instead of upfront. Runs locally, enforces permissions structurally rather than via prompts, and skips the protocol overhead entirely.
Garry Tan's gstack adds ten specialized 'brains' to Claude Code, each accessible via a slash command: code review, planning, shipping, QA testing, and more. The idea is switching cognitive modes on demand, from founder-level product taste to paranoid security review. Runs multiple sessions in parallel.
Godogen chains Claude Code skills to build complete Godot 4 games: architecture design, art generation via Gemini and Tripo3D, GDScript coding, and visual QA through screenshots. Handles 2D and 3D projects, runs on commodity hardware.
TradingAgents coordinates multiple LLM agents for financial trading, with each agent handling a different aspect: analysis, risk, execution. Open-source Python framework for building and testing AI-driven trading strategies, not a production bot.
Andrej Karpathy released a tool that visualizes US job data from the Bureau of Labor Statistics: employment size, growth projections, pay, education requirements. The twist is an LLM that scores each occupation on user-defined criteria like 'digital AI exposure.' Rough estimates, not displacement predictions.
LangChain's new langgraph deploy CLI builds Docker images for agent projects and provisions Postgres/Redis locally. Also adds commands for listing, logging, and deleting deployments, plus templates for scaffolding new agents.