Today we are introducing Dyna-2, a world-action model pre-trained on one million hours of human video. At this scale, for the first time, we discovered several new scaling laws: • world-action models exhibit scaling law on human data across four orders of magnitude, from 1000 to 1,000,000 hours, • this human data scaling law implied a scaling law on never seen robot data, • both data and objective matter; world modeling and scaling on video data are essential for cross-embodiment scaling transfer to emerge 🧵

This is amazing. The White House waived Jones Act requirements in mid-March, and it's generated billions in economic activity and revealed tons of natural trade routes that were impossible while the Act was in effect. We really need to permanently remove this stupid law!

Gas Turbines and Datacenters my TLDR is: • Literally every turbine (even the inefficient ones) are sold out through 2030. • Prices are 3x those from 2019. Slot times are more valuable than the cost efficiency. • Turbine OEM's were burned out from a buildout from 2009–2014 when there was a huge rush due to the shale revolution, shift to use more gas due to Fukushima, coal retirements. For ex, GE Power imploded. Therefore, they have been slow to expand production. • Four key bottlenecks for parts that go into a gas turbine: 1. Single-crystal castings: only 5 firms in the world do this 2. Heavy rotor forgings: only 4 firms do this 3. The casting furnaces themselves: 3 firms 4. GSU transformers: in the US, there's only one. Heron Power is building here, they raised $140 million in a Series B from a16z in Feb 2026.

The narrative: HEaT sHiElD TiLEs ArE a DeAd EnD The reality: SpaceX can trade off refurbishment costs against landing odds. If the starship that right now remains a floating contemporary art piece on the surface of the Indian Ocean had instead been brought back to the tower, refilled and launched again and then incinerated on re-entry, that alone would have reduced prospective cost per kg to orbit by 45% relative to an expendable Starship top stage, and by 58% compared to Falcon 9's estimated costs.

NFA, my own thoughts. Multiple compression happens when people go from asking "what if" to "what for". AI is still in this "what if" stage because it literally unlocked tens of billions of dollars of net new spend in a new category. Nothing was replaced, a new industry was born. Robotics started off by asking "what can we automate" but now we're talking about automating factory workers and doing the dishes. I can already do the dishes. Crypto had the "what if" moment with NFT's, with L1's, L2's, etc but now its just payments infrastructure getting bid up. Space for the next few years is just more rockets and satellites. The moon, though, is a "what if". People have been talking about manufacturing in space or doing more experiments in space for more than a decade. Nothing is currently manufactured in space. Defense is in this "what if" phase. Re-inventing defense is a creative endeavour, of what defense could look like in the future to combat future threats. Just a personal framing I thought of. There's really no easy justification for multiple expansion and compression except that there are people that are willing to pay a certain multiple to buy equity.

cannot stress this enough: the more you step into visibility and sharing your work..the more you feel other people's projections onto you... the more important it is to surround yourself with writings, films, paintings, relationships etc that mirror back your self concept. the concepts that resonate to you, that you have chosen. so much of society will try to tell you who and what you are. design and nurture your mind palace, your spirit's sanctuary. let the external conform to your spirit, not vice versa.

Nobody is close to breaking ECC Full article here: TLDR is: • We are not even close to breaking ECC (which breaks bitcoin, eth, https, lots of internet communication, etc) • The biggest breakthroughs have been in theory, with the 2026 paper Justin Drake was affiliated with showing that they only need 500K physical qubits to break ECC. The last estimate was in 2022 that it would take 317M physical qubits. • However the number usable qubits in one machine has barely moved: ~105 running a real algorithm. • Getting from 105 qubits to 500K qubits is going to take an unknown amount of time, there's no Moore's law equivalent • When quantum computers are able to break ECC, Bitcoin is probably cooked (unless people agree to burn all Bitcoins that don't move over to quantum secure signatures). Eth would also have to do the same, but the foundation would probably be able to convince people to move to a new signature and burn unmoved tokens. • We already have quantum secure signatures, the difficult part is moving over to it. So the internet is fine cause companies would force websites to move over to new signatures. Thank you @CabronElBufon for the amazing editorial suggestions.

I think the AI datacenter energy startup investing spree has become bubbly because at the end of the day, energy is is purely a function of the cost per kw and the sales of the kw. The Anthropic x Blackstone SPV to fund Google TPU purchases was priced at $6B A1 paying Treasury + 100bps, $24B A2 paying 5.75%, and a Class B $4.5B paying 8.5%. This is hardware, I'd guess for just energy its slightly cheaper to finance. So a company like @fluidstack 's entire job is to minimize their cost of capital and maximize their selling price. The energy price ceiling that PJM introduced means that perhaps energy costs are exorbitant, which may make the per kw costs of ocean and space datacenters economical, but investing in those companies means betting that the kw costs stay elevated and continue to rise. It is not only a bet on the cost of energy rising substantially but also that credits don't get commoditized across AI models, which is where crowd sourced compute comes in. Even if datacenters in the ocean become operational today, the calculus on bond markets on whether to fund their expansion is entirely based on the cost of capital vs selling price. I don't even know if it would make sense if these alternatives worked today, let alone in a couple years when they actually become operational.