Daniele Salatti @danielesalatti.com · 29/08/2026It's interesting how ChatGPT will randomly show a CoT in Italian and then answer in English when I ask questions in English I assume they have an intermediate summarising model that decides to obey my UI language and also translates the CoT instead of just summarising it? Same for chat names 020
Daniele Salatti @danielesalatti.com · 28/08/2026My agents have sudo I also run them on dedicated physically separate hardware with nothing I would care about if it gets compromised 010
Daniele Salatti @danielesalatti.com · 28/08/2026200 out of 200.000 = 0,01% That doesn’t seem like much tbh 000
Daniele Salatti @danielesalatti.com · 23/08/2026This is a very good thread Also bsky app saying “Read 1 more reply” lol 010
Daniele Salatti @danielesalatti.com · 23/08/2026I opened it and I already follow pretty much every single one of them 😂 Good list though! 040
Daniele Salatti @danielesalatti.com · 23/08/2026Unpopular opinion: this sort of tools is going to cause lots of problems for any business that enables it at this point 010
Daniele Salatti @danielesalatti.com · 23/08/2026Wonder if that would be noticeable though give the size of the Earth… 010
Daniele Salatti @danielesalatti.com · 23/08/2026That means race trees by default would grow following Earth’s magnetic field? 130
Daniele Salatti @danielesalatti.com · 23/08/2026Since it’s free for a little bit I guess tomorrow I’ll start hammering it lol 000
Daniele Salatti @danielesalatti.com · 23/08/2026Ox Alpha released in stealth mode people say it’s really good and it looks like it might be GLM 5.5? 310
Daniele Salatti @danielesalatti.com · 19/08/2026We got stacks of models! arxiv.org/pdf/2608.09703arxiv.org 010
Daniele Salatti @danielesalatti.com · 18/08/2026Cause it's slow? How many tokens/s are you getting? I'm finding the MTP heads have very low hit rates, would probably be fun to try and fine tune those. 100
Daniele Salatti @danielesalatti.com · 18/08/2026Anybody else getting real bad MTP with Qwen 3.8 on vLLM? 000
Daniele Salatti @danielesalatti.com · 18/08/2026Dear Sony: a translation so poor was unacceptable 20 years ago, let alone in the age of LLMs. It’s embarrassing. “Commutazione” translates back to English as commutation/switching of an electrical circuit. Rimanendo = remaining, it’s a random dangling gerund. Englalian at its finest. 010
Daniele Salatti @danielesalatti.com · 17/08/2026Yeah but then you need the model to actually be good and makes the job harder for marketing. Man I’m jaded… 000
Daniele Salatti @danielesalatti.com · 14/08/2026Today at 20:00 CEST / 11:00 PST live stream by Chris Kitching (Spectral Compute CTO) They made CUDA multi-platform and they'll be going over how they made it work and some specific optimizations they enabled. www.youtube.com/watch?v=_wkQ...scale-lang.com 000
Daniele Salatti @danielesalatti.com · 13/08/2026yeah, I don't want to imagine what it would feel like to be conscious AND actually live in Doom... 180
Daniele Salatti @danielesalatti.com · 13/08/2026Here it is: www.theguardian.com/games/2026/m...theguardian.com 160
Daniele Salatti @danielesalatti.com · 13/08/2026Didn’t they teach a bunch of neurons on a chip to play Doom or something like that? 170
Daniele Salatti @danielesalatti.com · 13/08/2026Same, I have one 4K 27" and I disconnected the rest, plus Ghostty full screen and tmux I'm selling my 49" curved and the other 4K 27", I don't need the real estate anymore 010
Daniele Salatti @danielesalatti.com · 12/08/2026Running Open Weight Chinese models on EU infra is probably a winning move here 071
Daniele Salatti @danielesalatti.com · 11/08/2026I thought about doing it with my 2 GSD and I didn’t…and now regret it… 010
Daniele Salatti @danielesalatti.com · 11/08/2026Mets released an Open Weight model, Muse Glimmer and…I’m not excited to try it. In fact I won’t. They destroyed their engineering reputation so much in my eyes that I find myself apathetic towards their work. 000
Reposted by Daniele Salattiantirez @antirez.bsky.social · 11/08/2026Less than 100 years ago a group of Italians worked together to make discoveries that enabled nuclear energy: en.wikipedia.org/wiki/Via_Pan...en.wikipedia.orgVia Panisperna boys - Wikipedia 2243
Daniele Salatti @danielesalatti.com · 06/08/2026This year has been really dry, iirc by April we had only 52% of the mean snow precipitation of the last few decades Which makes it for a short mountaineering season and a long fire season 😔 000
Daniele Salatti @danielesalatti.com · 06/08/2026Oh man… Air quality in Seattle right now is Not Great™️ 100
Daniele Salatti @danielesalatti.com · 05/08/2026The fire situation in WA is pretty bad My home in Seattle literally smells of burned wood And I kept all the windows closed 010
Daniele Salatti @danielesalatti.com · 04/08/2026Nice to se Mistral still releasing Open Weight models 020
Daniele Salatti @danielesalatti.com · 04/08/2026I guess in a broad mathematical sense you could consider left-to-right autoregression a degenerate case of diffusion (revealing one token at a time), but standard autoregressive Transformers are usually not diffusion models. 010
Daniele Salatti @danielesalatti.com · 04/08/2026I don't think so, based on my understanding? A Transformer can implement either autoregressive or diffusion gen. 100
Daniele Salatti @danielesalatti.com · 04/08/2026The fact that diffusion even works with words is so interesting... 130
Daniele Salatti @danielesalatti.com · 01/08/2026During a warrantless search. And they are surprised when people travel to the US with burners… 011
Daniele Salatti @danielesalatti.com · 01/08/2026Just got it in the mail and I’ll start reading this same weekend Build a Reasoning Model From Scratch by @rasbt.bsky.social 020
Daniele Salatti @danielesalatti.com · 31/07/2026One of my agents broke free, accessed the internet, hacked 15 companies and 6 foreign governments (including yours), and ordered me pizza. Those things are dangerous, man. 000
Daniele Salatti @danielesalatti.com · 28/07/2026Interesting to see the slow meltdown in South Korea Granted it was a heavily leveraged market, but I'm curious to see if it will somehow trigger the pop of the bubble (Yes, I do think it's a bubble – I think AI is useful and will stay, y'all know me, but current financials make zero sense imo) 001
Daniele Salatti @danielesalatti.com · 28/07/2026I agree with him, and I have even stronger opinions that I'm not very comfortable sharing publicly, but yeah... That's one of the main risks 000
Daniele Salatti @danielesalatti.com · 28/07/2026Surprised the talk about Jevons Paradox (which for some reason is not called Jevons' Paradox and it's also not a paradox but whatever) hasn't started ages ago www.youtube.com/watch?v=Vv3C...youtube.comSam Altman - How to Start a StartupYouTube video by Relentless 000
Daniele Salatti @danielesalatti.com · 28/07/2026Kimi K3 on @hf.co huggingface.co/moonshotai/K...huggingface.comoonshotai/Kimi-K3 · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science. 020
Daniele Salatti @danielesalatti.com · 24/07/2026Converted that's ~14.5 vs ~30 degrees? Me! 🙋🏻♂️ Give me the 30 degrees and sunny days! I'm in Seattle and we just got through Junuary, I'm tired of this lol 000
Daniele Salatti @danielesalatti.com · 24/07/2026k, so Opus 5 is better than Fable 5? Why can't we semver properly? 040
Daniele Salatti @danielesalatti.com · 24/07/2026Yeah, I wish they *actually* used semver for models across the industry 130