Well, the Mac Studio 512GB is planned for this month (but I guess Meta already has some), but it should be possible to link 4 256GB macs. That would be a $40k cluster, requiring only 4 x 500W of power. In comparison, a single Nvidia GB200 has 384GB of memory, needs 2700W and costs $70k. So maybe Meta is planned to only use macs for its next AI datacenter: just a shed with 2 million macs connected to get a 1GW for only $20B. Or maybe Meta will just lease them for $200/mo each, so for only $400M...
Comments & Reviews
"Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence." "Two models often have similar benchmark scores, yet one seems to handle a much wider range of unfamiliar tasks." "That’s what complaints of "benchmaxxing" and praise for "big model smell" both get at: a breadth of capability that popular benchmarks in a few domains can’t fully capture." "To quantify this mismatch, we investigate how performance on tasks targeted during post-training...
30 years old and still as kool as ever. It definitely aged better than most people I know...
"In response to user feedback, we’ve removed built-in AI from the open-source versions. AI tools are installed separately." Sane choice. Let's hope it'll stay like this as long as possible, because users should never have to complain for AI to be removed by default.
Enforcing UTF-8 as the default encoding for I/O operations typically shows how technical debt can grow in the Python world (and it can still be disabled in many different ways for interoperability purpose). As for the JIT (in CPython), some reports estimate speed improvements of 10%, which is far from PyPy and Pyston performances but the work has just very recently started.
Intelligence has always been what was missing in Artificial Intelligence. Great to see it's finally been added now.
The actual reality is that tokens burn is steadily growing but spending is plateauing for months now because most companies are mostly using cheap models since the return of investment is still not measurable yet. Now, we can say that LLMs hit a peak about one year ago with thinking models like GPT-5, and since then the bubble slowly deflated, with no real breakthrough but only been marketing stunts (new benchmarks, empty promises and ludicrous financial projections). But when you're a company...
It's already been existing for many years now. It's called "Steam search", where you can find an infinite number of sloppy, broken and ugly games pretending to be AAA games, by entering only few words — no coding required.
Good news (even for non-Chrome users) because JXL gives much better results for photographies than Webp (that is still average but still much better than JPEG) and could even be better than AVIF (that is already very alright, and even supports animations and transparency). And given how much Google contributed to develop JXL, become coming around for whatever reason to not really support it (citing slow adoption), it makes sense to enable it, even if Firefox enabled it only recently. So it...
Well, if this 1T params model is a "chonk", it would be Garfield trying to get few lasagna meals a day, when other frontier models are cheetahs eating data as much as their weight for breakfast. It's priced much more than most Chinese models, and just as much as medium-sized closed US models (that are permanently on sale; actually, Le Chonk is also on a 50% sale), Mistral will certainly hardly earn much money with it. But once more, EU tries to catch up with China and US, so let's not complain.
Chrome OS is based on Gentoo that is utterly patched to run Android apps, when Googlebook OS is based on Android. But the real strategy behind it is to sell premium ($900+) computers instead of very cheap ones ($150-$400), to integrate Gemini as much as possible (since a decent NPU is required). It's no longer an operating system but an "intelligence system", said Google that is slowly loosing its mind.
TypeSafe is this tool/API where you ask "what's the color of the sky?" with a set a possible answers "blue, white, gray, red" and the model responds: "blue: 81%, white: 12%, gray: 4%, red: 2%". It's some kind of basic fuzzy logic, and companies like that: "how we fire Bob?", "yes, no", the model responds "yes: 51%, no: 49%", and Bob is out of job.