tara tan

The Download for this week: • NVIDIA's $20B Groq acquisition debuts at GTC with the Groq 3 LPU, its first non-GPU data center chip. Huang projects $1T in AI infrastructure orders through 2027. • DoorDash turns 8M couriers into physical AI data collectors with a new Tasks app for filming household chores and real-world environments. • Google upgrades vibe coding and vibe design in the same week. Figma shares drop 4% on the Stitch news. • LeCun's first technical output since his new co AMI Labs' $1B raise: V-JEPA 2.1 learns robot grasping from 62 hours of unlabeled video. • Rivian spins out Mind Robotics, which raises $500M Series A to train robots on Rivian's factory sensor data. Plus: 66% of tech CEOs say they won't backfill departing roles. Block, Atlassian, and Meta are cutting a combined 20,000+ jobs citing AI. Hiring plans at their lowest since 2009. Read the full brief ↓ thereview.strangevc.com/publish/post/1…

NVIDIA's latest AI rack, Vera Rubin, produces the heat of 160 homes. The next generation will double that. The one after will likely double it again. The industry is racing to solve the heat problem, from subsea data centers to launching servers into orbit. But the most likely next step is the least exotic: liquid cooling. The catch? The hardware is the easy part. The real cost is operational. It rewires how facilities are built, staffed, diagnosed, and run. Our latest by Strange Research Fellow Rahul Narula explores what changes, and where the opportunity sits. open.substack.com/pub/strangerev…

I am looking for a fractional chief of staff! Please, must work well with humans and agents. More details here: thereview.strangevc.com/p/were-hiring-…

Humanoid robots can walk, run, even do backflips. What they can't do reliably: pick up a screwdriver. The hand is now the gating constraint for the entire industry. New piece on humanoid robots by Strange Research Fellow Joy Yang: thereview.strangevc.com/p/hands-down-t…

Good marketing people can move markets. Most marketing people just spin in circles. Choose wisely.

Always so fun to spot a @formlabs machine at a portfolio company 🧡

Claude is indispensable. That’s the only conclusion you can draw from what the Pentagon is doing and why they are standing ground. When you have every frontier AI lab in the country OpenAI, Google, Meta, xAI, at al bending over backwards to give the Department of Defense whatever it wants, with no restrictions, no red lines, and the Pentagon is still chasing the one company that won’t fully comply, it becomes obvious. They need Claude more than Claude needs them. What Anthropic refuses to comply with are on two specific things: First: fully autonomous weapons systems. Those that select and engage targets without meaningful human supervision. Second: mass domestic surveillance of American citizens. That’s it. Two red lines. Everything else is on the table. What I keep coming back to is the structural irony at the center of this. The safety culture that the Pentagon is trying to override is the same culture that produced the model they can’t easily replace. The careful thinking about failure modes, the investment in reliability, the institutional discipline around constraints is the very reason why Claude is on classified networks. I wrote about the full confrontation, the legal threats, the historical parallels, and what Friday means at the link below.

The first major milestones in AI for science may not be a breakthrough model or a novel reasoning architecture. It may be the integration layer (workflow orchestration, retrieval, synthesis) that helps researchers find non-obvious connections across the existing body of knowledge. New piece on AI co-scientists by Strange Research Fellow @mason_rand63492 open.substack.com/pub/strangerev…

interesting play from @anthropic. Wrote about the effects on data centers and the power grid this week, and how Virgina, for instance, is projecting a 2.6x increase in electricity price to cover new infra costs. open.substack.com/pub/strangerev…

@onderyazici lol dude

@onderyazici Yes for coding, but not for other tasks like writing and research

What are folks using as a shared database when working with Claude code / Claude projects / Gemini / Openclaw? I want to access the same projects across different platforms / agents and finding it a little frustrating for work beyond coding.

Excited to introduce the first cohort of Strange Research Fellows 2026! Welcome Joy Yang, Mason Rodriguez Rand, Valeriia Verkhovykh to the Strange Ventures family. They will be focused on beats we care about: robotics, advanced materials and manufacturing, and the future of compute. You'll hear from them very soon if you're subscribed to The Strange Review (thereview.strangevc.com). Welcome!

@samhcarter gotta be economic incentives, or if china, centralized policy

great breakdown of the private grid it begs the question: at which layer(s) can you actually create a self-enforcing coordination mechanism rather than relying on voluntary alignment & is such enforcement necessary or optimal? physical infrastructure, control plane, orchestration layer, or something else?

@clairevo @garrytan I made that dashboard! Glad it's helpful

As a SFUSD parent we’re obviously not happy with the strike, but the overwhelming sentiment within parent groups is supportive of teachers (a 3% raise is NOT a crazy demand.) I’ve appreciated this parent-created data tracker to understand some of the mechanics sfusd-dollars.vercel.app/fact-finding-2… No one is looking forward to a strike, and from my vantage, it seems both sides really want to get to a workable agreement. It’s not simple, but I was heartened by our kid’s teacher leaving the class Friday saying “I really hope I see you Monday.” Just one mom here who has been really helped by SF schools and hopes to see them open and thriving.

30,000 SF parents are preparing for the first SFUSD strike on Monday in almost 50 years Unions say they’ll accept the existing offer after a few days of strike: not because of budget cuts, but because union bosses chose political theater over your kids garryslist.org/posts/union-wa…

Interestingly, this exact thing happened in the 1880s. Every factory that wanted electricity had to generate its own. It took 40 years before we built a shared grid. → open.substack.com/pub/strangerev…

America's AI industry is quietly building its own parallel power system. It started in a Memphis parking lot. xAI needed hundreds of megawatts. The grid said: wait 5 years. So they trucked in jet-engine turbines and started generating their own electricity. 120 days later, 100,000 GPUs were running. Now everyone's doing it. 56 GW of private generation is planned across 46 sites, 30% of all U.S. datacenter capacity. I mapped every site, every machine, and the three strategies companies are using to outrun the grid.

agents *need* to be multi-channel - anything less is sub par user experience. i want to be able to access what i'm working on from anywhere