Claude gains interactive charts and diagrams across all plans

🧠 LAUNCH

Claude gains interactive charts and diagrams across all plans.

Claude can now build interactive charts and diagrams directly in the chat — and it's available today in beta on every plan, including free. This isn't a static image export; it's live, explorable visualization that turns Claude into a genuine data exploration tool. If you've been bouncing between ChatGPT and a Jupyter notebook just to eyeball a dataset, that workflow just collapsed into one window. (12,472 likes | 878 RTs) Read more →

OpenAI's Video API gets Sora 2 upgrade: custom characters, 20-second clips.

OpenAI just shipped major Video API upgrades powered by Sora 2 — custom characters and objects, 16:9 and 9:16 exports, clips up to 20 seconds, and video continuation. This is the moment Sora stops being a toy and starts being infrastructure: developers can now build real video features with consistent characters and flexible aspect ratios. If you're building anything media-adjacent, the API docs are worth a look today. (303 likes | 34 RTs) Read more →

Replit Agent 4 pivots from coding tool to full knowledge work agent. Slides, videos, spreadsheets — Replit's latest agent handles it all, completing the company's transformation from "IDE for students" to "AI productivity suite." The bet is that the same agent loop that works for code works for everything else. (Latent Space) Read more →

NVIDIA Nemotron 3 Super drops on HuggingFace: 120B total parameters, but only 12B active per token via Mixture of Experts. That active-parameter efficiency means this runs on hardware that would choke on a dense 70B model — the parameter count on the tin is increasingly misleading. Already trending. (126 likes | 2.8K downloads) Read more →

Meta open-sources Canopy Height Maps v2 (CHMv2), a high-resolution forest canopy mapping model built on their DINOv3 Sat-L foundation model in partnership with World Resources Institute. A concrete example of foundation models being specialized for environmental monitoring at global scale. (143 likes | 20 RTs) Read more →

💡 INSIGHT

Karpathy coins "intelligence brownouts" after OAuth outage kills his research agents.

Andrej Karpathy's autonomous research agents got wiped out by an OAuth outage, and his response landed like a grenade: "Intelligence brownouts will be interesting — the planet losing IQ points when frontier AI stutters." He's right. As more real workflows depend on frontier AI, we need to treat AI availability like we treat power grid reliability. If your agents don't have a failover plan, you don't have agents — you have a liability. (5,650 likes | 244 RTs) Read more →

Bold prediction: Amazon will ban AI-generated code changes within weeks. Whether or not it actually happens, the prediction from Bindu Reddy (2,379 likes) reflects a real tension enterprises face — AI-assisted coding ships faster but accumulates tech debt that's harder to audit. The counterargument: banning the tool instead of improving the review process is like banning calculators because accountants make mistakes. (2,379 likes | 239 RTs) Read more →

Anthropic's Alex Albert draws a direct line from last year's agentic coding revolution to this year's knowledge work transformation. His thesis: Claude for spreadsheets, presentations, and analysis is on the exact same trajectory that coding agents followed — early skepticism, then rapid adoption, then "how did we ever work without this." (439 likes | 16 RTs) Read more →

Ramp AI Index confirms Anthropic is now the default for business spending. This isn't vibes — it's hard transaction data from thousands of companies. The enterprise shift away from OpenAI dominance is showing up in credit card receipts. (1,461 likes | 121 RTs) Read more →

📝 TECHNIQUE

OpenAI shares engineering lessons from building computer-use agents.

Making long-running agent workflows practical required three things: tightening the execution loop so agents recover from failures faster, providing rich context via file systems instead of cramming everything into the prompt, and enabling network access with real security guardrails. If you're building anything that runs for more than 30 seconds autonomously, these are the architecture patterns to steal. (578 likes | 75 RTs) Read more →

How a CPO at a $2.6B company runs his entire day through Claude Code. A detailed walkthrough from Pendo's field CPO covering CLAUDE.md setup, Skills vs MCP vs Hooks, and why he switched from Cursor to terminal. The key insight: Claude Code isn't just for writing code — it's a general-purpose command center when you configure it right. (467 likes | 55 RTs) Read more →

🔧 TOOL

Claude in Excel and PowerPoint now shares context across apps and gains Skills support. If you live in Office, this is the update that makes Claude actually useful across your workflow — it remembers what it saw in your spreadsheet when you switch to your deck. (456 likes | 25 RTs) Read more →

One year of the Responses API: OpenAI reflects on how tool-use turned API responses into actions. When models can call tools, the line between "answer" and "execution" disappears — a useful retrospective on the current state of agentic API capabilities. (469 likes | 34 RTs) Read more →

HuggingFace PRO now offers 1TB private + 10TB public storage for $9/month. That's a 10x bump that makes the Hub a serious alternative to S3 for ML artifact storage. If you're paying more than $9/month to host model weights, do the math. (113 likes | 7 RTs) Read more →

🔬 RESEARCH

NVIDIA publishes how AI-Q reached #1 on both DeepResearch benchmarks. The detailed methodology reveals how agentic search and retrieval architectures can be systematically optimized — not just scaled. If you're building research agents, the architectural choices here are more valuable than the leaderboard position. Read more →

AutoResearch-RL: A new paper proposes reinforcement learning agents that perpetually self-evaluate and discover novel neural architectures autonomously. The agents don't just optimize hyperparameters — they invent new designs without human intervention. A step toward AI that improves its own infrastructure. (351 likes | 51 RTs) Read more →

🏗️ BUILD

OneCLI: A credential vault purpose-built for AI agents, written in Rust. Solves the messy problem of how autonomous agents securely access secrets, API keys, and credentials without hardcoding them. If your agents currently read from .env files, this is the upgrade. (44 likes | 16 RTs) Read more →

LogClaw: Open-source AI SRE that monitors logs and auto-creates tickets when issues are detected. Bridges the gap between observability and incident response without human triage — plug it into your staging environment and see what it catches. (13 likes | 7 RTs) Read more →

🎓 MODEL LITERACY

Mixture of Experts (MoE) — Active vs Total Parameters: NVIDIA's Nemotron 3 Super claims 120B parameters but only activates 12B per token. How? In a Mixture of Experts architecture, the model contains many specialized "expert" sub-networks, but a learned routing mechanism sends each input token to only a small subset of experts. The total parameter count reflects everything the model knows; the active parameter count reflects what it actually uses for any single prediction. This is why a "120B model" can run on hardware that chokes on a dense 70B model — the memory footprint is large, but the compute per token is small. When evaluating open models, the active parameter count tells you the inference cost; the total parameter count tells you the knowledge capacity. The number on the tin is increasingly misleading.

⚡ QUICK LINKS

Anthropic Institute: New effort to advance public conversation about powerful AI and governance. (3,861 likes | 444 RTs) Link
Anthropic opens Sydney office: Fourth APAC location after Tokyo, Bengaluru, and Seoul. (2,197 likes | 165 RTs) Link
DeepMind names London HQ 'Platform 37': After AlphaGo's famous Move 37. (1,438 likes | 129 RTs) Link
Mistral AI Now Summit: First flagship conference, Paris, May 28. (268 likes | 47 RTs) Link
Google AI heart health screening: Deployed in rural Australia. Link
DeepMind reflects on 10 years since AlphaGo: From games to scientific discovery. (87 likes | 13 RTs) Link

🎯 PICK OF THE DAY

Karpathy's lost research agents reveal we've already crossed the dependency threshold. When an OAuth outage killed Andrej Karpathy's autonomous research lab, it wasn't just an inconvenience — it was a preview of a world where frontier AI stuttering means real work stops. His term "intelligence brownouts" is going to stick because it names something we've all felt but haven't articulated: we've quietly built critical workflows on top of AI infrastructure that has no failover. Your coding agents, your research pipelines, your data analysis — when Claude or GPT goes down, what's your plan B? For most teams, the honest answer is "wait." That's the same answer utilities gave before we built redundant power grids. The difference is we're building this dependency in months, not decades, and nobody is planning the redundancy layer. If you're running agents in production, today's the day to build a degradation strategy — because the brownouts are only going to get more expensive. (5,650 likes | 244 RTs) Read more →

Until next time ✌️