> if Anthropic models run in a Google cage, does that not also mean Google can distill Anthropic models even easier (both fast and unseen)?
Not really in the interest of the cloud org to allow that, they don't really care how well deepmind is doing as long as they can sell TPU and API (incl. reselling openai and anthropic).
There's primary, final, and useful energy. Easy to confuse them. Final is what is delivered for consumption. Useful energy is what is actually delivered in a useful form (motion, light, etc.) The rest is rejected energy (heat). This terminology simplifies a complex chain of events that is about moving and transforming fossil fuels from well (primary energy) via refineries, shipping, trucking, pumping, etc. to gas tank in your car (final energy). After which you burn it to actually get useful energy. There are losses throughout this chain.
The fallacy here is counting barrels of oil, shipments of coal, or whatever and assuming it needs to be replaced 1:1 by what used to be called alternative energy to get the same useful energy. But of course with electrified systems, the losses from generation to use are a lot less. Things like wires and transformers have their own inefficiencies of course. But the numbers are pretty good. You end up with 80-90% useful energy. As opposed to 30-40% for typical fossil fuel plants. And far less for any combustion engine vehicles.
It's called the primary energy fallacy. As Michael Liebreich suggests, there are people using this term that genuinely don't get it. And there are those using it intentionally that do. But the bottom line is that their math tends to be off by about 3-5x.
A lot of energy is used for heating, heating water is more efficient with fossil fuel than with electricity - because for electricity you have transmission loss.
That used to be the case, but with the efficient heat pumps that have come out in the last few years, that is no longer true. Since you’re using electricity to move heat rather than produce it, it’s two to three times more efficient than burning fossil fuels for heating.
Fossil fuels have transmission losses too. You just don't see them.
40% of all global shipping is used for fossil fuels. [1]
10% of the energy in a barrel of crude oil is used to refine it. [2]
I'm not even counting the energy cost of crude oil extraction, pipeline leaks, hauling gasoline, diesel and LPG to its final destination, or the million other places where energy is lost.
Does it, though? Assuming 80% of energy use is fossil and fossil has a 50% loss, if all that fossil were to transition to ~100% efficient electricity[0] we'd already be at 33% of our renewable electricity needs. That's quite a bit more than the 20% the "80% fossil" seems to suggest!
"One-third done" sounds an awful lot like we're going to start seeing optimistic "we'll be halfway there at the end of the decade" headlines soon...
[0]: Things like the grid and car batteries don't operate at 100% but closer to 95%, but things like heat pumps operate at 300% efficiency, so I reckon that'll more-or-less cancel out for back-of-the-napkin math.
Yes we're not done yet, but renewable only started gaining share again around 2010, and how quickly solar/battery is being deployed is really impressive.
If they thought it to be backdoor, why not sell for some other believer who pays more than the bounty? But anyway, not everyone is after money and fame.
That said if we're counting bridge/ferry in other countries why not also count it in Switzerland? (Ferry: Horgen and Beckenried, for trains: Lotschberg, Furka, Simplon, Vereina)
Do compilers use profile information to allocate compilation time (such as by using an extremely expensive but slightly better register allocator for a hot loop)?
I got the impression that expensive register allocation is not an option. A 1% improvement in register allocation that results in a 0.1% local performance improvement and a 0.002% program performance improvement may be considered not to justify 1000x local compile time, but for frequently run stable-ish software the compile time and compiler development/maintenence time might be justified.
(I am not a programmer, but I do like thinking and learning about such things.)
Gemini is the gemini app (gemini.google.com), LLM output on search is AI mode / AI overview. Not surprising given it's google, but those are vastly different product / teams / infrastructure / web data availability.
China mined more coal in 2025 then ever, but burned less in coal plans. Coal is slowly shifting from electricity production to synthetic oil and chemical production.
Not really in the interest of the cloud org to allow that, they don't really care how well deepmind is doing as long as they can sell TPU and API (incl. reselling openai and anthropic).
reply