Anthropic launches Claude Code Review

#1 𝕏

Claude launched Code Review as a research-preview beta for Team and Enterprise, bringing AI-powered code feedback and analysis to development workflows.

#2 𝕏

DeepLearning.AI Google introduced Nano Banana 2 (Gemini 3.1 Flash Image), a faster, lower-cost image generator built on Gemini 3 Flash that produces images in seconds with iterative editing and cuts generation costs by roughly half.

#3 📝 OpenAI News

OpenAI to acquire Promptfoo - OpenAI announced on March 9, 2026 that it will acquire Promptfoo, an AI-powered prompt testing and validation framework designed to help developers ensure prompt reliability and performance.

#4 𝕏

Andrew Ng launched Context Hub, an open CLI tool that gives coding agents up-to-date API docs to prevent outdated calls and hallucinated parameters. It also lets agents annotate and save workarounds, with plans for agents to share these learnings over time.

#5 in

Udi Menkes open-sourced 28 AI-powered PM skills and used Claude Code to prepare and execute a real product launch in one session.

#6 𝕏

Claude reports that after running their AI review tool on most Anthropic PRs, substantive review comments jumped from 16% to 54%, with under 1% false positives. On large PRs (1,000+ lines), it flagged findings in 84% of cases, averaging 7.5 issues each.

#7 𝕏

Boris Cherny launched Claude Code’s new Code Review feature, where AI agents deep-review every PR, boosting Anthropic engineers’ code output by 200% this year and catching bugs he’d otherwise miss.

#8 𝕏

Andrej Karpathy ran a two-day autoresearch tuning on a depth-12 nanochat model, uncovering ~20 additive changes that, when applied to depth-24, cut “Time to GPT-2” from 2.02 h to 1.80 h (~11% faster) for a new leaderboard record.

#9 𝕏

Cognition launched Devin Review, a free no-signup PR review tool with autofix, smart diff organization, copy/move detection, and codebase-aware chat—just swap “github” with “devinreview” on any PR to try it.

#10 𝕏

Philipp Schmid shows how the Gemini Interactions API can process minutes to hours of YouTube video content in seconds with a single API call, highlighting a major leap in video understanding.

#11 𝕏

LlamaIndex 🦙 launched Surreal Slides, a tool that uses LlamaParse to convert slide decks into structured data stored in SurrealDB, letting you query your entire presentation library in natural language without manual scanning.

#12 in

Greg Isenberg discovered a GitHub repo that spins up a 43-agent AI agency—spanning engineering, design, marketing, product, PM, testing, and support—to coordinate end-to-end idea shipping, and it racked up 10k+ stars in under 7 days.

#13 ▶️

My OpenClaw is the only marketer I need

Greg Isenberg

#14 📝 Doug Turnbull

Ugly hack to force BM25 to 0-1 - Describes a practical technique to normalize BM25 lexical scores into a 0–1 range. Addresses the problem that BM25 scores are inconsistent and proposes an ad-hoc workaround to make them comparable.

#15 📝 Simon Willison

Perhaps not Boring Technology after all - Reflection on the concern that LLMs might bias technology choices toward well-represented tools in their training data, making it harder for new tools to gain traction. The piece explores the implications of that effect for tooling and innovation.

#16 𝕏

Peter Yang distills five takeaways from Pencil CEO Tom Krcha on how humanized AI agents—armed with context, personality, and iterative creativity—will revolutionize design without sacrificing craft and delight.

#17 𝕏

claire vo 🖤 @jameygannon spent thousands of hours in Midjourney using Pinterest mood boards and a custom personalization code to generate consistent AI-driven brand imagery without complex prompts.

Also covered by: @Claire Vo

#18 𝕏

Shreyas Doshi urges PM Builders to carefully read AI’s initial output, evaluate its logic against their specific context, and then craft precise follow-up prompts to drive more critical, accurate reasoning.

#19 𝕏

Santiago argues that most ideas aren’t worth implementing and he uses AI to optimize and accelerate his existing work rather than shipping new features.

#20 ▶️

5 steps to generate consistent brand images with Midjourney

How I AI Podcast

Also covered by: @Claire Vo

#21 𝕏

Guillermo Rauch spotlights OpenBlock’s AI coding agents with built-in Vercel CLI integration—demonstrating in a video how these “real agents” can auto-generate, deploy, and manage production code seamlessly.

#22 𝕏

Josh Woodward unveils NotebookLM’s new ePub upload feature, letting students import full e-books directly for AI-driven note-taking, summaries, and Q&A.

#23 𝕏

Andrej Karpathy clarifies that this version of autoresearch lacked time-controlled evaluation, so runs with lower validation loss but longer training times were rejected. Changes are only accepted if they achieve better-or-equal loss and better-or-equal training time.

#24 𝕏

Cognition launched a Datadog integration in the MCP Marketplace, letting your AI agent Devin pull monitoring data via the MCP Server. They’ve published full step-by-step setup instructions.

#25 𝕏

Boris Cherny shows that ramping up “test-time compute” by feeding more tokens into a coding prompt—and using multiple uncorrelated context windows—yields significantly better code results.