I don't think these will be any such revolution because there's not that much benefit to a consumer. Even the BS AI you get in "Google Answers" will still be capable of way more than what you can do running an LLM locally, even with one of these gadgets. Also, people already HAVE this ability. It's called a GPU, or graphics card. I have Ollama running on my GPU right now, locally, unconnected to the internet.

If they put these in people's computers, I'd guess they'll afterwards tell you that what you REALLY should use them for is as part of an "AI Net", so that your computer can be leveraged as just part of one big answer machine, rather than really being a local copy that's just for your use. The AI companies will offer you small discounts in your subscription if you use them this way, that type of thing.

Nothing would be better for their business model that getting people to make their own homes and PC's an actual part of the AI network. Distributed data centers, that the CONSUMER pays the bills on.

RTX Spark: A desktop/laptop system-on-a-chip that pairs Blackwell graphics with an Arm CPU. It is highly efficient, drawing just single-digit wattage at idle and scaling up to 80W at full load.

80 watts is probably less than a current laptop CPU with attached laptop GPU.

much less than a deskside CPU... and a deskside CPU paired with a GeForce gamer GPU... try well over 1000 watts.

This isn't for LLM training, but AI inference... and for GPU graphics / CPU general math.

Running AI learning at the edge makes no sense... latency will kill you. So forget doing training operations with a million laptops burning electricity that you pay for.

The chip in my S25+ already does some AI stuff locally (like photo editing w/o phoning home).

Also, my desktop system doesn't draw anywhere near 1000W even at full load (I've never seen it hit 400W on my Kill-A-Watt even while stress-testing), with a higher-end GPU and 8 core/16 thread CPU, but I take your point there. And I'd add even with a lot more power than 80W (and 12GB of VRAM) locally-run Ollama is still pretty dang limited.

information. How can we possibly use such a distorted fantasy to make reliable judgements?

It is potentially as dangerous as Crypto with as little basis in functional reality.

Talk about a fact free universe?????????????

Overpromise and underdeliver.

AI training and experience can't always control them, making them a huge security risk.