1ChatGPT Images 2.0 Ships With a Brain

OpenAI just dropped the biggest leap in AI image generation since DALL-E 3. ChatGPT Images 2.0 brings thinking capabilities to image generation — the model reasons about composition, renders precise text, and follows complex layout instructions before a single pixel is placed. This isn't incremental; it's a new architecture that treats image generation as a reasoning problem. Available now in ChatGPT — go break it. Read more →

gpt-image-2 hits the API and Codex on day one. The production-grade version landed simultaneously for developers — stronger text rendering, editing, and resolution control ready for integration. If you've been waiting to add image generation to your pipeline, the wait is over. (1,117 likes | 86 RTs) Read more →

2Google Ships Deep Research Max With Full Agent Tooling

Google DeepMind fires back with Deep Research Max, their SOTA autonomous research agent powered by Gemini 3.1 Pro. It can safely navigate both the open web and your custom data — internal docs, specialized databases, whatever you point it at. With MCP support, native charts, and full tool integration baked in, this is Google's clearest play yet at the agentic research space. (1,445 likes | 143 RTs) Read more →

Deep Research API gets MCP, streaming, and multi-modal input. The biggest API upgrade yet from Google's Deep Research team: MCP support lets you wire it into existing tool chains, real-time streaming means you can watch research unfold, and multi-modal input opens the door to image-and-document analysis pipelines. If you're building autonomous research workflows, this is the integration point. (1,530 likes | 123 RTs) Read more →

Claude Cowork builds live dashboards connected to your apps. Claude can now create live artifacts — dashboards and trackers wired directly to your apps and files. Boris Cherny (Claude Code creator) calls it the most practical new Claude feature this week, and he's right: this turns Claude from a code generator into an internal-tools builder. (18,188 likes | 1,429 RTs) Read more →

Brex open-sources CrabTrap: an LLM judge for agent safety. An HTTP proxy that interposes an LLM judge between your agents and production APIs — every action gets validated before it executes. If you're deploying autonomous agents and losing sleep over what they might do unsupervised, this is the safety layer you've been building internally. (53 likes | 8 RTs) Read more →

Inside GPT-Image-2: how thinking makes image generation smarter. The research team behind the model explains the reasoning capabilities in a detailed thread — this is the first image model that plans composition before generating. The architecture treats prompts as problems to solve, not just instructions to follow. Essential reading if you want to understand why this model handles complex multi-element scenes that stump everything else. (1,935 likes | 127 RTs) Read more →

A dead-simple trick for getting diverse LLM outputs. Getting LLMs to produce genuinely varied responses is a known hard problem. This technique has the model generate and manipulate a random seed before producing output — it's trivial to implement and solves the "everything sounds the same" failure mode in creative and sampling tasks. (281 likes | 29 RTs) Read more →

Simon Willison stress-tests ChatGPT Images 2.0 with adversarial prompts. The practical companion to OpenAI's official announcement — Willison pushes the model with complex composition requests, hidden elements, and edge-case text rendering. If you want to know where the model actually breaks rather than where it shines, start here. Read more →

GPT-Image-2 sweeps every Image Arena leaderboard. First clean sweep in Arena history — GPT-Image-2 claimed #1 across all Image Arena categories on launch day. Third-party validation that OpenAI's claims aren't just marketing: the model genuinely outperforms everything else in head-to-head human preference testing. (2,904 likes | 311 RTs) Read more →

3AI Agents Match 146 Economist Teams — With Tighter Variance

A classic economics study gave 146 human teams the same dataset and got wildly different answers. A new replication paper reruns the experiment with agentic AI: Claude Code and Codex land near the human median, but with far tighter variance. The implications are staggering — AI agents aren't just matching human researchers, they're doing it more consistently. When reproducibility is the metric, the machines are already winning. (706 likes | 121 RTs) Read more →

MegaStyle claims the ImageNet moment for style transfer. Full open-source release of code, training data, and models — arriving the same week as ChatGPT Images 2.0 to offer an open alternative for style-specific generation. If you need fine-grained control over artistic style rather than general-purpose generation, this is your starting point. (273 likes | 37 RTs) Read more →

Mollick: we just crossed a quality threshold nobody expected. Ethan Mollick has been testing GPT-Image-2 for weeks and reports a capability inflection point — AI-generated text in slides, papers, and visual materials is now practically usable without manual cleanup. This isn't about "better images." It's about an entire category of professional output becoming automatable. (1,073 likes | 77 RTs) Read more →

Meta starts capturing employee keystrokes for AI training. Mouse movements, keystrokes, screen interactions — Meta is collecting it all from employees to train AI models. Following Atlassian's quiet data collection move last week, workplace surveillance for AI training is hardening into an industry pattern. If you haven't reviewed your company's AI data policies recently, now's the time. (262 likes | 242 RTs) Read more →

Kimi K2.6 positions as the open-source front-runner ahead of DeepSeek v4. Latent Space's deep analysis makes the case that Moonshot's Kimi K2.6 is the strongest open model available right now — refreshed to compete with Claude Opus 4.6 and potentially ahead of the anticipated DeepSeek v4. Essential context for teams deciding between open and closed models this quarter. (169 likes | 93 RTs) Read more →

4HuggingFace Open-Sources ml-intern: The Agent That Automates Post-Training

HuggingFace releases ml-intern, an agent that automates the post-training pipeline — evaluation, fine-tuning decisions, and model comparison handled by an AI teammate. This isn't a demo; it's a working implementation of the "AI colleague" pattern that actually ships production ML work. If you maintain models and your post-training workflow involves repetitive eval-tweak-repeat cycles, clone this today. (2,901 likes | 360 RTs) Read more →

Agent-Simulator streams iOS into your browser with MCP. An open-source tool that pipes your iOS simulator into a browser window with full MCP integration — AI agents can interact with, inspect, and jump directly to React Native/Expo source code. Mobile development just got a lot more agent-friendly. (148 likes | 13 RTs) Read more →

Kimi 2.6 Code adds a Claude Code-style terminal. One of the strongest open-source models now has a proper terminal-based coding interface. This closes the gap between "powerful model you access through an API" and "usable coding agent workflow you'd actually reach for daily." (113 likes | 8 RTs) Read more →

Autoregressive Image Generation: GPT-Image-2 generates images token-by-token, the same way language models generate text — fundamentally different from the diffusion approach used by Midjourney and DALL-E 3. Diffusion models start with noise and iteratively refine it into an image, which is great for visual quality but struggles with precise text and spatial reasoning. Autoregressive generation lets the model "think" sequentially about composition, render accurate text, and follow complex layout instructions because each token is conditioned on everything that came before. It's why GPT-Image-2 can reliably put the right words on a sign in the right place — the model plans before it paints.

• Claude Code's /btw side chat: Quick questions without breaking your main workflow — CMD+; on desktop. (84 likes) Link
• Official guide to prompt caching in the Claude API: If you're running Claude in loops or agents, caching can cut your bill dramatically. (103 likes) Link
• GoModel: Lightweight open-source Go gateway for multi-provider AI routing. (155 likes | 61 RTs) Link
• Google's Nano Banana Pro: Strict brand consistency across logos, typography, and assets — production brand work, not just pretty pictures. (748 likes | 63 RTs) Link
• Claude Code exits the Pro tier: Coding agents are now a premium segment, not a bundled feature. (169 likes | 93 RTs) Link

When AI agents produce tighter variance than 146 human economist teams, the benchmark question flips. A landmark economics replication study gave identical datasets to 146 human research teams and got wildly divergent results — a well-known problem in social science called the "garden of forking paths." The new paper reruns the experiment with Claude Code and Codex. Both land near the human median, but with dramatically tighter variance. This isn't about AI being smarter than economists — it's about AI being more consistent. When reproducibility is the core challenge of empirical research, and AI agents deliver more reproducible results than large teams of trained humans on identical data, the question flips from "can agents do research" to "should humans still be the reproducibility benchmark." The implications extend far beyond economics: any field where researcher degrees of freedom produce conflicting findings — medicine, psychology, policy analysis — just got a new tool for methodological discipline. (706 likes | 121 RTs) Read more →

Until next time ✌️