Latent Space

The Inference Inflection is HEATING up.

The end of an era.

Cursor product launch and megakernel GPU engineering tradeoffs discussed in Latent Space newsletter roundup.

Technical reconstruction of ChatGPT Work's agent architecture: memory, scheduling, browser automation, plugins, and tool composition.

Alibaba releases Qwen 3.8 Max (2.4T params) and 27B open-weight models optimized for coding and collaboration tasks.

Baseten Series F funding and technical deep-dive on autoregressive and diffusion inference optimization strategies.

DeepSeek V4-Flash 0731 released; minimal other activity.

GPT 5.6 pricing drops 20–80% via recursive self-optimization; equivalent GPT 5.4 intelligence now 13x cheaper in 4 months.

AI engineers adopt ontologies to constrain probabilistic agents within deterministic logical boundaries, reviving semantic web techniques.

Latent Space observes AI adoption accelerating in financial services as the next major vertical after software engineering.

The Big Pause is coming.

OpenAI's Akshay Nathan details ChatGPT Work product strategy: Sites, memory, subagents, finance, no-code tools scaling from 0 to 10M users.

Kimi K3 open-weights model released amid broader industry discussion on open model availability and strategy.

Anthropic releases Claude Opus 5 matching Fable performance at half the cost, demonstrating efficiency gains in model distillation.

Black Forest Labs releases FLUX 3 multimodal model with reported improvements over Gemini 2.0, Grok Imagine, and includes video-action robotics variant.

Laguna S 2.1, a 118B MoE model from Poolside AI, achieves Deepseek v4 Pro performance at lower cost than v4 Flash.

Poolside AI co-CEO Eiso Kant describes building a model factory enabling efficient training of 118B MoE models competitive with 1T open-weight alternatives.

Latent Space observes emerging trend in AI cybersecurity coverage without detailing specific breakthroughs or novel attacks.

Xaira Therapeutics builds causal models for drug discovery using synthetic data generation; Bo Wang and Ci Chu discuss data requirements for model training.

No significant AI developments reported.

No significant AI industry announcements or developments reported.

Kimi K3 2.8T-A50B released as largest open-weight model with Opus 4.8-class performance at Sonnet 5 pricing.

Lila Sciences argues scientific labs as data sources for AI training, positioning robotics and experimental workflows as frontier training data beyond internet corpora.

Thinky releases Inkling, a 975B multimodal open-weights model under Apache 2.0, with a smaller 276B variant.

Codex user base growing at 1M users daily; limited concrete details on adoption or product impact.

AIE World's Fair 2026 identified shift from agent-centric tooling to systems-level architecture design patterns.

Codex usage grew 10x to 7M users in 6 months; article questions whether it has outpaced Claude Code amid sparse adoption metrics.

Commentary noting an absence of major announcements following a week of model releases.

A big day for OpenAI.

SpaceXAI continues to move faster than any other frontier lab on earth.

Modal CTO Akshat Bubna discusses infrastructure requirements for agentic AI systems, covering lessons from building agent-native cloud platform.

Lilian Weng summarizes 35 papers on harness engineering for RSI; meta-analysis of recent research without new findings.

Commentary on a recent model launch framed as historically significant, lacks specifics on model capabilities, architecture, or performance.

AI Engineer World's Fair concludes with debate on agentic loops and report on engineering practices; keynotes address development priorities.

Vercel's Andrew Qu discusses eve agent framework, emphasizing skills, sandboxes, and agent-readable web design as architectural primitives.

Adobe experiments with agentic sites that dynamically generate pages based on individual user intent, signaling shift toward personalized web experiences.

Paul Bakaus on Impeccable discusses skill engineering, human-in-the-loop agent design, and limitations of one-shot prompting approaches.

Newsletter post noting absence of significant AI industry announcements on a given day.

AIEWF speakers debate autoresearch and software factory vision, raising concerns about human agency and control in AI-driven development.

Introspection co-founder explains autoresearch loops, agent recipes, and self-improving systems while arguing humans remain essential to AI software development.

Cursor's Forward Deployed Engineers help enterprises implement AI agents as software factories, per Pauline Brunet.

Evan Feinberg and Sergey Edunov discuss diffusion models for drug discovery at Genesis Molecular AI, including PEARL's OpenBind performance and protein co-folding advances.

Warp CEO Zach Lloyd argues automated software factories will become standard for major projects, outlining preparation strategies for engineers.

AI Engineer World's Fair coverage: agent loops, software factories, forward-deployed engineering, and open model adoption emerging as key themes.

Latent Space newsletter teases upcoming model releases (Sonnet 5, Fable 5) with minimal detail; appears to be placeholder or speculative content.

Sierra's Natalie Meurer discusses convergence of product and forward-deployed engineers in software development.

Ahmad Osman argues local AI inference is rapidly closing performance/cost gap vs. cloud APIs across consumer and enterprise deployments.

Commentary on a slow news day in AI; no substantive developments or announcements.

Oddly tiered releases to both OAI and ANT on the same day.

OpenAI internal Codex usage surged 13–56x across departments since Nov 2025, with Research leading adoption.