Passionate about making tech more responsible, notably via my v own https://bettertech.blog. And sharing memes ofc
James Martin
Loading...
More essential work from the incredible @ketanjoshi.co. 20GW in just one country is an obscene amount of energy (= 20 nuclear power stations). Why, once again, can’t they make smaller, less energy-sucking models? Scaling at all costs is totally incompatible with a viable planet and has to stop now
James Martin
“We're using $300m of Anthropic [Claude Code] this year (…) The vast majority of those tokens don't need to go to Anthropic. There needs to be some intermediary layer that's saying, oh, that one has to go to Anthropic, but these ones can be handled by smaller models…" - Marc Benioff, Salesforce
Stupid is as stupid does flip.it/IwX7Aj
This means that Doctolib, the app practically everyone in France uses to book medical appointments - and has received lots of state money as the founder is BFFs with Macron - is sending appointment notes taken by its AI assistant to Meta, Google & co to train their f’ing future models. Nice, eh?
AI now costs so much even the hyperscalers can’t afford it.
- $121 billion in bonds in 2025, versus a $28 billion annual average 2020-24 (x4)
- Tech now = 10% of credit market, vs 2% in 2005 (x5)
flip.it/FT7SHE
(Interestingly, those multiples are similar to AI’s electricity & water increases 🤭)
Brilliant. Apple, disappointed about having to *follow EU law*, blames EU law for Siri AI being unreleasable in Europe. So it’s not their fault that imposing one model (Gemini) on all of its clients, with no guarantee of data safety, is illegal. That’s all clear then 🙄 www.apple.com/newsroom/202...
Benioff adds there should be a systems which works out which model is "the most affordable for the job."
TL;DR: even the boss of Salesforce says tokenmaxxing’s days are numbered 💡
Source: podcasts.apple.com/fr/podcast/t...
"Destroy AI!", Ronny Chieng tells Harvard grads
youtube.com/shorts/0z7Q0... via @404media.co
Fable 5, Anthropic’s new model, is at least c. 2-3x bigger than its previous biggest, Opus 4.8. This means it could consume as many more resources. And yet it only scores 11% better on coding benchmark SWE-bench. Crazy, right? To #QuitBigAI, walk this way: bettertech.blog/frugal-ai-sw...
One French person in SIX consulted an AI service before deciding whom to vote for in the recent local elections. FIVE percent changed their minds. Do they know about bias, transparency, how these things are trained (by well off US white dudes)? 🤯 www.publicsenat.fr/actualites/p...
