swyx

A panel of heavy AI users reports what works after OpenAI DevDay: agent loops and forks, subagents, context management, model speed and quality, local models, prompt injection risk, payments, and delegation.

Latent Space interviews Periodic Labs' founders one year after launch about building an AI scientist, autonomous physics labs, and applying frontier models to materials discovery from semiconductors to superconductors.

Radical Numerics' Eric Nguyen argues, citing the OpenAI-to-Hugging Face attack, that open-weight models are becoming essential defensive infrastructure against AI-enabled biological and cyber threats rather than just amplifying risk.

A newsletter on the open weights debate, where a small set of players will actually decide the outcome while everyone else grandstands, starting with the open models letter signed by NVIDIA and Microsoft.

A long-form technical interview with Baseten's Philip Kiely and Ali Taha covering how production LLM inference actually works: cache-aware routing, disaggregated prefill/decode, speculative decoding, quantization error cancellation, KV-cache management, model parallelism, and custom GPU kernels. The conversation also covers grafting a vision encoder from Kimi onto GLM-5.2, the tradeoffs between shared APIs and dedicated deployments, diffusion vs. autoregressive video generation, and the convergence of training and inference through continual learning.

Latent Space interviews MIT's Alex Zhang on Recursive Language Models, context offloading, subagent swarms, KernelBench and GPU Mode, and treating harnesses as compositional generalizers.

Modal is an AI-native cloud platform that provides serverless functions, elastic GPU inference, sandboxes, and multi-node training via a decorator-based Python SDK. It's built for bursty, compute-heavy AI workloads that traditional Kubernetes-based clouds weren't designed for, and is increasingly optimized for AI agents to operate directly.

AINews covers TypeSafe's launch of Jev, a 'System One' model built only to decide, classify, route, and score, not to reason or generate code. It's pitched as a fast, calibrated, hallucination-resistant complement sitting in front of slower LLMs.

AINews covers Anthropic's Claude Haiku 5.5 launch: pricing parity with GPT-6 Luna, roughly 75% lower cost than Haiku 4.5, availability in Claude Platform and Claude Code, and same-day price cuts on Sonnet 5.5 and subscriptions.

Modal's CEO explains why AI workloads do not match the fixed-size GPU contracts used to buy capacity, and how a serverless platform absorbs the gap. Covers inference economics, sandbox growth, and what rapid scaling means for compute availability.

Latent Space interviews Runway's CTO and research lead about WorldPrompt, a timestamped-event input format for controlling real-time interactive world generation, and situates it against Genie 3 and Odyssey-2 Pro.

Glean's Arvind Jain explains why frontier model cost and stronger open weights have made routing a core deployment layer, including the case for answering some queries without an LLM at all.

Good Start Labs' CEO explains turning games like Diplomacy into verifiable-reward training environments for AI agents, spun out of Every, and how o3 and Opus 4 diverged sharply when faced with a betrayal-for-advantage scenario.

Latent Space interviews AIUC cofounder Rune Kvist on AIUC-1, an emerging standard backed by real insurance for testing agent security, safety and reliability, following the company's $40M Series A used by Cursor, Harvey, Lovable and ElevenLabs.

Kimi K3 is a frontier-class open-weights AI model from Moonshot AI with 2.8T total parameters, 1M-token context, and native multimodal input. It targets long-horizon agentic coding and self-evolving workflows, claiming performance near top closed models at pricing comparable to mid-tier options.

Latent Space interviews Pliny the Liberator and John V on jailbreaking methods, adversarial machine learning roots, and the state of AI red teaming.

Eiso Kant of Poolside on how the model factory works, following Laguna S 2.1 beating a Thinking Machines release nearly ten times its size, and a technical report praised for the unusual level of detail.

Latent Space talks with Airbnb CTO Ahmad Al-Dahle, formerly of Meta's Llama effort, about an inside-out AI strategy: internal tools like Everest speed development, then the same capabilities power new guest-facing services.

Why parameter count alone no longer describes a model: data volume, compute allocation and deployment conditions each change what a given size buys. Walks through the limits of Chinchilla-era assumptions under inference scaling, and credits GLM-5.3's gains to reinforcement learning on long-horizon t

Runway's co-founder covers the path from Gen-1 to Gen-4.5, world models as physics simulators for robotics, neural-rendered interfaces, real-time generation, and a proposed test for world models.

Latent Space talks with Anthropic's Thariq Shihipar about Claude Code's next phase, alongside a recap of Anthropic's recent releases including Sonnet 5.5, Opus 5.5, Claude Mods, and the plugins portal.

AINews covers DeepSeek's open-sourced DeepInfer serving stack, its disclosed production figures for caching and per-GPU throughput, and an analysis of what they imply for inference margins and payback.

Vercel's framework for building AI agents, providing prescriptive primitives for context management, tools, resumability, and long-running work. Deploy to Vercel for built-in observability and evaluations, with support for filesystem agents, skills, compaction, and subagents.

AIUC cofounder Rune Kvist covers the AIUC-1 agent standard, jailbreak and data-leak testing, liability when agents fail, and why risk and trust may become the binding constraint on AI adoption.