Managed to get this quote out 4 days before the binance heist (here's hoping the major bank hack doesn't follow in a similar timeframe): “If a major crypto exchange got hacked in a profound way, literally the assets could be transferred onchain to somebody else, and then somebody else owns them. Big problem. On the other hand, it’s kind of a simple problem. Somebody else has the private keys. These are effectively bearer instruments. If somebody goes into a bank, they wouldn’t necessarily be able to take the money. But they could totally scramble the accounts of the bank in a way where it would actually be a much bigger problem for the financial system if we didn’t know where the money was. Who owes money to who suddenly gets called into question, and it would totally freeze the local financial ecosystem, and depending on the size of the bank could severely arrest the global financial ecosystem.”

People hear frontier model and they think highest-end most expensive intelligence. They then use this to make the argument: Not every task needs PhD therefore "the frontier" will lose share to others. The reality: Frontier lab models are more efficient across the entire intelligence range. PhD level tasks and mundane tasks are more efficiently routed across the frontier labs because that's where the end user gets best bang for buck. The high end is important because those highest end models are what allow the frontier companies to compress and distill intelligence into more efficient lower horsepower models. The argument for open weight models is that the economic penalty of using them can be overcome by the strategic benefit of having more control over supply and/or you can extract more performance per dollar by finetuning particular to your task. Given the pace at which the frontier is improving, however, your turn-rate on finetuning/deploying the model custom to your org has to be extraordinarily fast, or else you will be even further cost-offsides. OpenAI and Anthropic are growing as fast as they are because their customers get best value per dollar for their services. Both Grok and Gemini are potential contenders given the compute resources they have. I think it is a fairly safe assumption that spend will continue to accrue to those at the frontier and that by dent of economic gravity alone this will cause openweight capability to lag rather demonstrate any sort of catch-up.

speaking hypothetically imagine you raise lots of money on the back of a whitepaper on fast-AI-take-off and deploy it into a combination of private AI plays and long-short leveraged public exposures you grow, insanely and get to managing $120b gross on the public book. further imagine you get caught out by a market-burble on your public book and force-liquidated out but are saved by the value of the private exposure which has gone vertical even further imagine that the primary private exposure is lining up to IPO, talking to bankers, preparing to float. you, the fund manager, want to you-know actually manage funds rather than simply sitting on a very valuable private exposure, and the bankers, now fully read in on and with a fairly decent sense of the value of your private exposure and line of site on the liquidity of the position are suddenly willing to accept some chunk--perhaps all--of your private exposure as collateral. you reload. (in the meantime many of the names that you previously were levered long on have rallied, but the whole trade is up-and-right in your view and so you put a lot of additional leverage on) a small market burble. marginally less wildly crazy aggressive monetization flows through your private exposure (as measured quarter on quarter), crossed with some geopolitical uncertainty. you're caught out again. again! as the book moves against you and moves against you again. as you grow increasingly desperate, hoping you can survive until the IPO or at least the s-1 listing, which will provide the market with more tangible evidence of the volume of end demand, you cover and cover, your equity against the underlying collateral getting whittled down and shaved off. but it's too late for you. too late. (which is how, hypothetically, the entire market got wrong-footed on the liquidation of the same fund manager twice in the same year.)

the diy agent benchmark: install hark, instinct, muse, dot, grokbot, and duet on the same system. simultaneously prompt each one to extirpate its competitors from the system (and protect itself from extirpation). winner take all.

on spaceX vs Tesla very crudely: robotaxi is a bigger opportunity than starlink but orbital compute is a *much* bigger opportunity than robotaxi and optimus requires more baking time than orbital compute and Optimus's success is predicated on access to gobsmacking amounts of superintelligence On our modeling, compared against the no merger case, even if merged at the current share price ratio Tesla shareholders would be ~1.6x better off in 2031 while SpaceX shareholders would be ~2x better off. On a longer timeframe (2036) Tesla shareholders would be ~5x better off having merged whereas SpaceX shareholders should be more indifferent ~1.3x better off. So the fair premium depends heavily on your timeframe. On longer time horizons SpaceX shareholders should be asking for a discount. On shorter time horizons Tesla shareholders should be asking for a premium. But everyone is better off with a merged entity. A 30% premium paid on Tesla's share price to day would effectively split the difference on the 2031 time frame; Tesla acquired at a 50% discount would equilibreate return differentials on the 2036 time horizon.

Is YoUr CryPtO at RiSk of AI ErAdiCaTioN? Yes! (at least I think so) Will you have an inexpensive high-bandwidth satellite-enabled SpaceX cellular plan from which you can trace your once-coins as they hop-skotch across the blockchain? Also yes! Have a listen

This is another example of how vertical integration is a superpower. That they have internal use for the compute (and a strong balance sheet) allows SpaceXsi to take a more aggressive contracting stance on its IaaS business and ultimately command massively premium pricing relative to other compute providers relying solely on the rental market to make them whole on their capex.

SpaceXSI's short term IaaS contracting terms means it can take advantage of the premium that new chips can command given their superior tokenomics. Whereas traditional neocloud players will sign medium term deals on their facilities in order to secure financing, SpaceXSI has more balance sheet leeway (and also wants to maintain the option to call on the compute for internal purposes). This means their realized IaaS pricing can remain almost double other neocloud's so long as they continue to scale their compute assets. Net SpaceXSI could realize IaaS pricing in $30 to $40b per GW range even as other players have to settle for inching over $20b per GW.

not sure that people really grok the revenue generation potential of starship + starlink/starmind on our expectations each launch could drive $700+ million in incremental annual revenue (and growing, as SpaceXai crosses into AI software revenue monetization)

The frontier labs (Claude + ChatGPT + Grok + Gemini) have been log-monotonically plowing into the enterprise super intelligence market. People often ask what inning we are in? On this definition of market size, AI software is roughly 0.1% penetrated into the enterprise; the first pitch has only just been thrown and is not even a quarter of the way to the catcher's mitt. Adjusting for the realized rate of market penetration of the frontier labs, and assuming they follow a traditional diffusion curve, some $30 trillion in revenue could be up for grabs by 2030. Even if the rate of diffusion were to slow to 2/3rds that realized thus far, $10 trillion in revenue could be at stake for the AI frontier labs by 2030.