Now, we could copy his work and make an analog CPU at microcolumn level, with weights stored in flash memory. It wouldn't be low inversion with 6 order lower power consumption, but perhaps 2 orders lower than today. Imagine ChatGPT 4 on a microSD-card in your smartphone.

SNUFA, darn spellchecker…

Have you ever watched presentations of research work? And I just said I don't have notes for that presentation. You could although try a search on Youtube, the SNAFU webinars have been going on for years.
As a rule I share what I know. If someone doubts whether it is correct, well, fine with me. 😋

If I got it right, then the guy that did the presentation didn't use a few simple optimizations, but given how much the network shrunk (num params), it should give about 20–25% compute for the same result.

If I got the numbers right from this years SNUFA presentations, that would be the consequence. They should be uploaded on youtube by now. Didn't make notes, but it is only 10 hours or so with presentations.
(If my own (unpublished) work is correct it should be possible to make additional speedups.)

Most likely it will take a year or more before it has any real impact on the industry. That is, unless something changes that counteracts this. Or I have misinterpreted the presentation of the work.

Picked up some news the other day. Seems like some quite simple changes might lower the necessary compute during inference to one quarter. That could have some real impact on valuation of data centers for AI, and for NVIDIA's sale of GPGPUs to those.

Å snakke med en #AI bot kan sammenlignes med å snakke med en psykopat. Det er ingen følelser der, og ingen empati.

The wealthiest in the US having worse outcomes than the poorest in western European countries 🤦

"in some cases, the wealthiest Americans have survival rates on par with the poorest Europeans in western parts of Europe such as Germany, France and the Netherlands."

The US spend by far the most on healthcare, and they get very little for it. Regardless how rich you are.

"Comparison data also showed that at every wealth level in the U.S., mortality rates were higher than those in the parts of Europe the researchers studied. "

www.brown.edu/news/2025-04...

Could be interesting to check the mortality rate through Trump's presidency. I suspect it will rise, and the expected lifespan will fall. Not sure how much, but I'm pretty confident it will be visible.

Could have something to do with the food. Food on the IS is notoriously unhealthy.

The number of necessary solar panels are dependent on the size and efficiency of the panels. Although, a ten times difference deems little too much.

"fordummende underholdning" er ikke et informativt begrep.