Marcel Salathé

Incredible. No SaaS business would survive this, unless they offered something absolutely everybody wants to have. And everybody wants Claude.

@albertwenger The reason why I'm thinking it could be important is that it has all the hallmarks of a microbiome change - idiopathic, it just begins and sometimes goes away, strongly linked to diet. And that although diet changes seem to help, there is no one size fits all

@marcelsalathe Definitely looking into that also.

Personal health blog post: GERD WTF? continuations.com/gerd-wtf

🎙️From task-specific AI to persuasive, are we witnessing a fundamental rift in technology? On #InsideAI, @cervisiarius and @marcelsalathe discuss LLM persuasion, language bias in AI, and why “machine psychology” may be the next frontier. 🎧open.spotify.com/episode/6j9BLc…

We’re excited to announce the speakers for “AI & the Molecular World” at #AMLD2026 in Lausanne 🇨🇭 on Feb 10! 🧬🤖 The track explores generative modelling, new molecular representations, and more across scales from small molecules to proteins. See below 👇

Seeing Vance's comments, I lived a similar moment 19 years ago in a Russian courtroom hearing a judge state that the police officer's testimony had to be trusted over our video evidence because "he was wearing the uniform". Total immunity beats reality in a police state.

You keep your life only at the regime's discretion is as clear as could be. I remember pointing out the difference between officially sanctioned police brutality in Russia versus George Floyd, where the officer had a fair trial and was jailed. The WH reaction here is Russian.

I’m not joking and this isn’t funny. We have been trying to have a baby since last year. There are various options, not everyone is aligned… I gave Claude Code a description of the problem, it impregnated my wife in an hour.

Our new podcast episode is live! @marcelsalathe welcomes Eamonn Maguire from @ProtonPrivacy for a rich discussion on privacy-focused AI and how interdisciplinary thinking shapes resilient tech ecosystems. 🔒 🎧 open.spotify.com/episode/3RiE0I…

Can a small Swiss lab beat trillion-parameter AI giants? In our new Inside AI episode, Aldo Podestà of Giotto.ai shares how his team hit #1 on the ARC-AGI reasoning benchmark — with a model 1,000× smaller than Grok 4. 🎧Listen here: podcasts.apple.com/ch/podcast/con…

@marcelsalathe plaide pour une régulation légère de l’IA en Europe. Trop de règles risquent de freiner l’innovation, mais un cadre souple peut propulser nos start-up et des projets comme Apertus au sommet du potentiel mondial. #IA #Souveraineté @letemps letemps.ch/cyber/marcel-s…

"European leaders outraged" here is the equivalent of Americans sharing their "thoughts and prayers" after a mass shooting. Does anyone have the political will and military force to do anything about it? The answer, sadly, is no, in both cases.

Ok, logging off. GPT-5 is not a major leap, but a consistent improvement, particularly in terms of reducing hallucinations (not hard compared to o3 though, whose hallucinations were often off the charts). Key development seems that OpenAI made everything much more seamless as an experience. Every user will automatically get the performance that only very experienced users could access before. It just goes to show how much we're already used to this level of intelligence. In every other universe, GPT-5 would likely be mind-blowing. In the universe of today, it's simply a solid advancement.

Hm, did GPT-5 check this graph? Something's off 😅

Wait what? All previous openAI models will be deprecated?

OpenAI just announced GPT-5! 700 million people now use ChatGPT every week. The most interesting new feature: GPT-5 can figure out by itself whether to give you an immediate response or whether it should “reason” for a while before answering. And GPT-5 will be available even to free users 😮 Performance on academic benchmarks doesn’t look dramatically better than o3. But hallucination rates are way down, which means fewer confidently wrong answers. So this seems to be mostly about reliability, which is great.

@francoisfleuret Agreed. Key improvement seems to be rather massive reduction in hallucinations

Ok, no leap.

@bfalsafi joins @marcelsalathe to break down the true energy cost of AI ⚡️ They talk GPUs, data centers VS cloud computing VS supercomputers, and why AI’s environmental impact is bigger than you think but also solvable. Tune in shorturl.at/qs3OQ

In the new episode of Inside AI, @marcelsalathe chats with @antonosika , co-founder of @lovable They talk about LLMs & product acceleration, security in AI coding and scaling in Europe’s startup scene. 🎧Listen on Spotify: podcasts.apple.com/ch/podcast/con…