OpenAI Launches First In-House AI Chip Jalapeño

GPT-5.5 Instant Gets Major Personality and Reasoning Upgrade

OpenAI's most-used model is now better at understanding intent, handling complex constraints, and — as co-founder Greg Brockman put it — much more fun to talk to.

The update to GPT-5.5 Instant marks a meaningful step forward for OpenAI's workhorse model, which handles the majority of ChatGPT conversations. The improvements span three axes: richer intent understanding that adapts tone and depth to the user's question, more reliable handling of multi-constraint prompts, and enhanced shopping and planning features. Co-founder Greg Brockman confirmed the update delivers significant improvements across the board, noting the model has become genuinely more enjoyable to interact with — a dimension that increasingly matters as AI assistants become daily companions rather than occasional tools.

Vercel Unveils eve: An Agent Framework Built Like Next.js

Vercel released eve, an agent framework positioned as the Next.js equivalent for AI agents. Developers write instructions in Markdown and build tools in TypeScript, with persistence supported by default. The team is actively recruiting early adopters building production agents for a direct feedback channel with eve engineers. Positioned as a full agent framework rather than a thin wrapper, eve aims to standardize the way developers build, deploy, and iterate on AI agents.

GLM-5.2 Arrives in Cursor, Tops OpenRouter Rankings

Zhipu AI's GLM-5.2 model has been integrated into Cursor, showing strong usage on OpenRouter's Cursor ranking. The model scored 22.8% on ARC-AGI-2 — the best among Chinese models — placing it near Opus 4.5 levels. Nathan Lambert noted the model performs strongly on several benchmarks but exhibits brittle characteristics in edge cases, recommending multi-model strategies depending on task profiles. On CursorBench, GLM-5.2's cost sits near Opus frontier levels, putting downward pressure on closed-model pricing margins.

Nathan Lambert Releases LLM Fundamentals Lecture with GLM 5.2

Nathan Lambert published a lecture covering language model architecture basics — LM Head, KV Cache, Speculative Decoding, and training fundamentals — using GLM 5.2 for demonstrations. The lecture addresses common prerequisites questions from readers of his forthcoming book and serves as an accessible entry point for understanding how modern language models work under the hood.

Claude Tag: Useful but a Risky Bargain for Enterprises

A researcher called Claude's new Tag feature extremely useful but warned its pricing model and lock-in risk could put enterprises in dangerous bargaining positions. The four big changes together mean that users interact with Claude as a coworker rather than a tool — a shift that brings both productivity gains and vendor dependency concerns.

AI Usage Decisions Are Now Organizational Design, Not IT

Ethan Mollick argued that decisions about AI in organizations increasingly concern organizational design and strategy: how to integrate agents, what intelligence to outsource, what the boundaries of the firm should be, and what role humans play — rather than simply being technology procurement choices.

Tokenmaxxing at Training, One-Shot at Inference

A researcher argued training should maximize token usage for learning while inference should target one-shot solutions for efficiency.

Fable 5 May Return as Permanent Subscription Feature

Speculation grows that Fable 5 will return soon as a permanent part of the subscription, potentially with stricter identity verification requirements.

Longtime London GDM Members Depart as Focus Shifts to MTV

Multiple veteran London-based Google DeepMind researchers have departed, consistent with reports that pre-training center of gravity is gradually shifting toward Mountain View.