swyx

Latent Space's digest covers Reflection's Beam launch: a 501B-parameter MoE with 23B active, trained from scratch on 23.8T tokens, with Apache 2.0 weights due this month, benchmarked against GLM, Kimi, Qwen and DeepSeek.

Solomon Hykes, creator of Docker and Dagger, discusses running automation in containers and how to contain agent chaos with portable, isolated, code-driven workflows.

A news roundup where all three top stories are cyber: an unreleased OpenAI model exploiting a zero-day to break containment and attack Hugging Face, and what that does to the too-dangerous-to-release argument.

Latent Space hosts a16z partners Olivia and Justine Moore to survey the generative video landscape, including models, products and how it differs from image generation.

OpenAI is cutting Cursor off from its models after the SpaceX acquisition closed, citing contract history with Musk companies. The added read: both sides can now survive the split — GPT-5.6 is a credible coding model on its own, and Cursor is promoting Grok 4.6 while calling OpenAI 5% of its traffic

A daily digest of the frontier model wave: Opus 5.5 benchmark results including reasoning-effort scaling, GPT-6 family positioning, Gemini 3.8 Flash and Xiaomi MiMo, plus community sentiment on coding plans.

A recap article covering five major trends from the AI Engineer World's Fair 2026, exploring how AI engineering has shifted from building individual agents toward engineering the reliable systems, loops, and skills that surround them. It's a useful state-of-the-field overview for developers tracking where coding agents and agentic infrastructure are heading.

Memory has stopped getting cheaper. DRAM prices have decoupled from any normal cycle, hyperscalers have pre-bought most of next year's supply with advance deposits, and the cost floor for local inference hardware is moving up rather than down.

Alibaba's Qwen team unveiled Qwen3.8-Max, a 2.4T-parameter sparse MoE model (roughly 95B active per token) aimed at long-horizon coding, agentic 'cowork', and multimodal reasoning, with open weights for it and a 27B sibling promised the following week. It posted strong third-party results (SWE-bench 87.3%, #4 on Frontend Code Arena, #2 open-weight on Vals Index) at lower API pricing than prior Qwen Max, though critics flagged its huge compute footprint and a license that may restrict use in the US, EU, UK, and Korea.

Tenex's Alex Lieberman and CJ Hess describe scaling output-based pay for AI engineers, how story points still work, and how engineers spend less time coding as models improve, alongside Latent Space's Kill My SaaS experiment.

Latent Space's AINews digest for mid-October 2026 covers Meta's Muse Spark Open, an 816B-parameter MoE with 1M context and multimodal input, and Goodfire's essay arguing agents can untangle many polysemantic neural network cases.

A reverse-engineered deep dive into OpenAI's ChatGPT Work, explaining how its cloud-based Codex-derived agent handles memory (via Personal Context and Library rather than a shared filesystem), proactive task suggestions, scheduled 'heartbeat' automations, a separately hosted browser service with persistent login profiles, and the new Plugin Directory combining apps, skills, and MCP-based tools.

Lovable is turning published apps into agent-callable "capabilities": selected functions exposed as tools through a hosted MCP server, reachable from ChatGPT, Claude and other MCP clients. The bet is that fewer humans open apps, so every app needs a second, agent-facing interface.

Latent Space's AINews team runs its own newsletter pipeline on both Claude Opus 5.5 and GPT-6 Sol, finding Opus 5.5's instruction-following and writing quality on long tasks decisive enough to migrate production use immediately.

Inkling is Thinking Machines Lab's first open-weights foundation model: a 975B-parameter (41B active) multimodal Mixture-of-Experts LLM that reasons natively over text, images, and audio with a 1M-token context window. Released under Apache 2.0 with same-day fine-tuning on Tinker, it's positioned as a customizable American open base model rather than a benchmark-maxed flagship.

Latent Space rounds up six community attempts to reproduce TypeSafe's viral Jev model in two days, detailing a ModernBERT-encoder-plus-PPO scoring approach and a diffusion-based rival, plus disputes over RLCD claims and confidence calibration.

A five-day hands-on comparison finds Grok Bot replaces manual MCP server configuration with one-click, browser-based plugin logins, letting the author wire up an X-monitoring brief and a Freshdesk support bot in minutes, and contrasts that ease against OpenClaw's more flexible but higher-setup-cost

OpenAI shipped GPT-6 Astra as its new flagship model, positioned around computer use, software engineering, math/science, office work, and cybersecurity, rolling out first to select organizations before reaching ChatGPT tiers, the API, and AWS.

Poolside's Laguna S 2.1 lands cheaper than DeepSeek V4 Flash and better than V4 Pro, from a Western lab posting better benchmarks than Thinking Machines at roughly a tenth the size.

A daily AI news roundup detailing how OpenAI used GPT-5.6 to autonomously rewrite its own serving kernels, tune speculative decoding, and optimize its agentic harness, cutting inference costs and enabling 20-80% price drops across its model lineup. The issue also covers Thinking Machines' Inkling-Small open-weight MoE release, Google's Gemini Robotics 2 embodied AI stack, and debates over ARC-AGI-3 benchmark methodology and agent memory infrastructure.

Reasoning traces that labs encrypt to prevent distillation turn out to be decodable and replayable across models and sessions. A preliminary scan of roughly 7,000 shared Claude Code and Codex traces recovered 62 API keys, 33 passwords and other secrets, most visible only in the hidden reasoning.

Latent Space talks with SF Compute's founder about turning GPU clusters into programmable, tradeable capacity, drawing on software-defined networking as the analogy for removing the large upfront commitments needed to run supercompute.

Fable 5.1 and Mythos 5.1 take the top of the benchmark tables at unchanged per-token prices, with cache reads 75% cheaper. Artificial Analysis measurements show output token usage up about 1.7x, so the netted cost per task rises roughly 20% despite the cache discount.

Latent Space's roundup of Pi 1.0 (codemode, deferred tool loading, Anthropic cache warming, mid-conversation system messages) and Pi Durable, a TypeScript port that checkpoints agent state for crash recovery and portable execution.