That video is very poor, for focusing on a red herring of trying to build a monolith 1gw datacenter in space. There is no plan for such. Instead a mesh of smaller (fits in 1 rocket) datacenters that each have their own solar/radiators auto deployed with the datacluster. Inference does not need fast networking between clusters as models fit in a single cluster easily.

It’s still completely uneconomic because even GPUs that survive 5 years, are worthless after 5 years. Plan is to trash the datacenters every 5 years including the solar and radiators.

AI: help me get straight A+s in school wink wink

Solar on deserts, or even Nebraska farmland, is far more profitable than corn for ethanol and produces far more energy. Even with 2-5 hours of sunlight/day, battery systems, would be 5x cheaper in Nebraska than space for 24/7/365 power production, with perhaps minor throtling in winter. A datacenter in the middle of a solar farm also eliminates noise complaints from cooling systems.

large radiators argument is extremely poor. Plan is a mesh of small datacenter modules. The inter nvl72 networking speed is irrelevant for inference/skynet. However, there is no hot semiconductor technology theory that competes with silicon for feature size/speed. There are much better designs than circulating fluid radiators, but all of them, add complexity and pump/nozzle failures. Supercritical compression (heat pump) systems is vastly more prone to failure, including in pipe/connection network. It also means thick, less deployable, radiators.

The core reason space datacenters are uneconomical, is the 5 year replacement cycle. Includes trashing the solar and radiators every 5 years even if they could last 30+ because datacenter modules need to auto-deploy their own power and radiators. Ocean deployments would be far better. Wind/solar/batteries doesn’t need frequent change. Just drop in a new module every 5 years, while taking out the old one, and rebuilding its insides.

They could be configured such that no off switch exists

Power grid dependence can be a weak point. Democracy can clash with CIA/oligarchy. A big enough mob can overwhelm security forces not particularly ideologically aligned to defending skynet.

It’s a good point that providing the security systems of nuclear power plants is cheaper and more job creating than space datacenters. You can also get more uppity when international adversaries nuke your datacenter on US soil as opposed to space.

But more money to a guy that recycles $100m+ into politicians every election cycle, provides decision maker influence to subsidize space skynet. Continuity of nazi/zionazi ideology may require being beyond any temporary loss of control over law.

BS is fair, 90% of it. There needed to be an excuse for xAI merger is origin of the BS.

There is zero economic reason. Most optimistic cost estimates are 5x more expensive than earth systems powered by solar and batteries. With near term optimism 10x more expensive.

The only reason is skynet that is immune for anyone in the US pulling the plug from skynet. Russia/China can already hit sattelites or terrestrial datacenters with missiles, and space ones are more vulnerable because Americans are more likely to cheer destruction of skynet assets used to oppress them.

The other factor is international law permits squatters rights to the most desirable orbits. 300 years from now, we may need power/data beemed to/from space. Privatizing it for oligarchy is a great long term investment if it is subsidized by US government.

There is white hat potential for this “creativity”. Find more/“all” vulnerabilities and get them patched. US empire allegiance will promote other paths, though.

It’s a stupid study designed to answer the question “if we give money to just car owners or just rich people… will they be happy/comply”. One of the problems is it costs money to measure/police compliance for any conditional aid.

Another approach to discouraging undesirable behaviour is to tax it. Revenue instead of expense, including just funding such a pointless study. Parking fees/taxes where you don’t want people to go is pretty easy.

So incredibly stupid/wasteful. Just put a $3/gallon additional carbon tax on gasoline/diesel. use revenue to at least cut other taxes/rent reductions if property tax cuts, or give every resident a dividend cheque based on the carbon tax revenue. Car owners will adjust and traffic/smog improved, without just paying car owners with free money. FFS, there are cheaper than $600/mo cars I can use to scam for free money.

Huawei is far ahead of Nvidia in AI clusters despite sanctions made against it, and 5 year behind tech stack. 50% cost/performance advantage over GB300 nvl72, (including building costs) by building outdoor siteable pods with 3 week lead times. CXMT is now most valuable company in China, but because of Huawei sanctions, they sell 64gb ddr5 modules at higher price than Samsung just to Huawei. Huawei needs its own RAM fab because it is harassed by west, and Chinese suppliers price to them with “captured customer” awareness.

We/all need AI for medical breakthrough potential… though.

I’m somewhat ok with catering to a political filter in an instance/frontend, even if I prefer open inclusion. But DDOSing anyone in fediverse is bad/unacceptable.

Open models are still most cheaply hosted on a cloud, with batching and 24/7 use. API rates from developer lab are generally fair. Self hosting does have some significant tangible benefits though: Fine tuning for domain specific to organization, and not letting LLM provider train from your prompts/answers, followed by competing with your organization in the future as a result of “distilling your IP”.

It is burying old news in headline. The consequences/pitfalls of their “strategic shift” are fairly new. The layoffs were explicitly justified for pivot to datacenters that OpenAI will “surely” be able to rent. The new extra problems in that strategic shift just makes them look worse for going all in on the bubble.

There is zero proof of distillation. Minimax 2.7 development was surrounded by moderate use of Claude. M3 is their latest generation, and pretty solid, but its performance cannot be attributed solely (or even 5%) to distillation, and that is only lab that has been accused of significant API use. These claims are all 3-4 months old by now, and Anthropic blocked China access after publishing the accusations. Repeated BS is BS from losers trying to lobby for support.

gherkin tests?

Also Cursor (recently bought by spacex for $60B) “stole” not only the entirety of kimi 2.5, but all of their users data to train its composer models ontop of kimi 2.5.