swyx

TypeSafe CEO Diogo Almeida explains System One Models, the Jev coding model, RLCD, RLHF mode collapse, calibration, refusals when AI sits inside software dependencies, and why the company avoids public benchmarks.

An argument that the agent capability jump around Christmas 2025 came from the model and the harness improving together, and that the pattern repeats: models absorb harness features into their weights, engineers delete what got absorbed, and what remains is a harness for human attention.

Recap of DeepSeek's quiet v4.1-Flash release: a new causal encoder-decoder architecture (763B-A8B, 16B decoder) with vision, retiring V4 Pro, continuing DeepSeek's history of hyperfocused architecture bets between major versions.

Stripe's ~$7B acquisition of OpenRouter closes 90 days after a $1.3B Series B, on roughly $140M annualized revenue and ~70% gross margins, with 250T tokens routed monthly — 5x growth since February. A read on where value accrues in AI infra: the routing layer, not the GPUs.

After DevDay, OpenAI's Computer Use lead and an API engineer explain faster, cheaper computer use, Dots, the Agents and Decisions APIs, async tool calling, mid-turn steering, UltraFast inference, and compaction.

A daily AI newsletter that aggregates and summarizes the most significant developments across AI Twitter, Reddit, and Discord — from model releases and research papers to agent tooling and benchmarks. Each issue distills hundreds of sources into a curated, skimmable recap.

An article arguing that AI engineers are rediscovering ontologies and Semantic Web technologies (RDF, RDFS, OWL, Schema.org) as logical guardrails to keep probabilistic LLM agents from going off the rails. It covers talks from AI Engineer World's Fair, including Frank Coyle's neurosymbolic AI framing and Neo4j's ontology-based semantic layer for enabling 'thin agents,' plus commentary from OpenLink Software's Kingsley Idehen on RDF-based agent memory.

Dyna Robotics' chief scientist explains Dyna-2, a world-action model trained on a million hours of human video, and how long-horizon reliability, not demos, drives deployment. Covers scaling laws, 45-minute task learning and a 24-hour napkin-folding run.

Latent Space interviews TypeSafe AI's CEO on Jev, a 'System One' decision-model category that returns typed probabilistic outputs instead of text, with linked cookbooks for coding-agent and game use cases.

An experimental Adobe concept where websites assemble themselves in real time around each visitor's intent. Instead of static pages, an LLM retrieves from a company's existing content and composes a personalized page on the fly for individual users and even AI agents.

Anthropic's Thariq Shihipar covers Claude Code power-user workflows, Mods for rewriting the harness, multiplayer agents, mutable software, and agent security incidents with defenses like sandboxing, classifiers, and probes.

Yi Tay of Google DeepMind discusses captaining the IMO gold effort, Deep Think, on-policy RL, and standing up the Gemini reasoning team in Singapore.

OpenRouter's CEO and an a16z partner discuss building a multi-model routing layer, why labs struggle with distribution, usage rankings as a market signal, model fusion experiments, and agent-driven token fraud.

Richard Socher discusses Recursive's bet on recursive self-improvement, including claimed results where its research system beat human teams on optimization tasks and found GPU kernel improvements without CUDA experts.

A daily AI-industry digest recapping a podcast debate on why hand-fused 'megakernel' inference designs are losing favor as NVIDIA's Rubin architecture addresses the launch-overhead problems they were built to solve, alongside Cursor's open-sourcing of its MoK training megakernel and a broad roundup of new model releases, agent-tooling launches, and frontier-model security incidents.

Covers Gemini 4 Argon reaching general availability across the Gemini API, Gemini app, Search AI Mode, and Antigravity. Includes introductory API pricing and its scheduled increase, independent benchmark confirmation, and highlights from the AI Engineer NYC keynotes.

Latent Space interviews Periodic Labs' Liam Fedus and Ekin Dogus Cubuk on building AI scientists: RL in noisy physical environments, autonomous high-throughput labs, and why experiments, including failures, are needed beyond published data.

Latent Space talks with the Command Code founder about adding a learned 'taste' layer to coding agents, so DeepSeek V4 can beat Opus 4.7 on a developer's own conventions, plus his path from open source and DevRel.

A biotech founder argues AI has cut the cost of reasoning in science but not the cost of running physical experiments, and outlines two industry responses: 'foundries' that industrialize measurement and 'navigators' that plan smarter experiments.

Matt Pocock of AI Hero discusses the 'grill me' prompt that has the model question you before building, and lessons from teaching Claude Code courses amid rapid tooling change.

Latent Space's hands-on review of GPT-6 Astra after burning 20B+ tokens: the model reportedly acts as an autonomous AI engineer, choosing and training models, labeling data, watching pipelines, debugging deployed systems, and commanding subagents across long threads.

Examines Stacklok's bet that cloud-native control-loop orchestration can make coding-agent harnesses reliable beyond the desktop, covering session reliability, isolated tool execution and context preservation.

Report from AI Engineer NYC on Anthropic's workshop for dynamic workflows, a multiagent orchestration layer for Claude Managed Agents. It covers the organizational model behind flows and what Anthropic says they are built for.

Alex Zhang, the MIT researcher behind Recursive Language Models, discusses harnesses as compositional generalizers, programmatic subagent calling, agent swarms, AI-written GPU kernels and capability overhang in current models.