Andrea Causio

my own personal AGI moment arrived last week: gpt 5.5 completed our mandatory HR training videos for me, driving chrome via devtools opus 4.7 was a huge wuss about the whole thing and refused while aggressively lecturing me. i can understand why pete hegseth banned it

Open source is vibe coding with humans instead of agents.

look at Germany go they are finally ready to compete with Llama-2 7B

@sharongoldman Lol

Me: I am very impressed with the new OpenAI image model. Wowza. My sister, professor of anatomy at a med school: "Ask it to create an anatomically correct labeled image of the human thorax." Me: ok! Here it is. How is it? Sister: Oy. Me: what's wrong with it? Sister: It has a giant extra set of veins going to the upper lobe of the left lung. The arch of the aorta is way too low. The costal cartilage label is pointing to the lungs. The labels are pointing all wrong. The label to the trachea is pointing to the thyroid gland. The heart label is pointing to the sternum and so on and so forth. I think the leaders on the labels all pointing to the wrong things are the weirdest thing Me: it's still cute tho Sister: Oh and there are some crazy bones coming out the side of the neck above the first rib. Me: 🙂

@ColtonOrtolf Been trying to do this for some time but without a lot of success. Can you share the code? Would love to replicate your experience

Can a knowledge graph make an LLM a better doctor? I encoded 3 cardiometabolic clinical guidelines from USPSTF, ACC/AHA, KDIGO into a graph where medications, conditions, and recommendations are structured nodes. Then I tested "does giving an LLM this structured context actually produce better clinical recommendations?" I ran 16 synthetic patient cases through 3 setups (1) Vanilla LLM with just patient data (winging it) (2) LLM + retrieved guideline text (standard RAG) (3) LLM + graph-derived context showing which guidelines apply, what they recommend, and where they agree. Graph context won on every dimension. The composite score: 4.14 vs 3.38 (flat RAG) vs 2.98 (vanilla). Interestingly, when a patient matches multiple guidelines, the graph can tell the LLM "hey, 3 independent guidelines all point at the same statin medications for this patient." RAG retrieves chunks of text and hopes the model connects the dots. My hypothesis: structure beats retrieval for multi-source clinical reasoning. Still lots of tuning to do, but super interesting early results. Built the whole thing with Claude Code in about 3 days

@AlecStapp God, I wish I saw this before paying $50 for filing state taxes on an income under $400

This may sound like a sponsored post but I promise it's not (I just hate TurboTax and their rent-seeking lobbyists with the passion of a thousand suns): Use FreeTaxUSA instead. It's >90% cheaper than TurboTax and easier to use/higher quality software.

@garrytan Having AI write code is literally less than half of the prompts I write at this point. Researching, conceptualizing, planning has become prominent.

This is 100% correct. I experience this 10 hours a day now

@vincenzo @niulinx Good job!! Leggevo proprio oggi che Tesla FSD é arrivata in EU (Olanda), ottimo sapere che la principale concorrente sará italiana!

Segnatevi questo nome: @niulinx. Il Fondo Artificial Intelligence di #CDP Venture Capital, assieme al Fondo Life Ventures di #A2A, ha appena costruito l'investimento in quella che diventera' la principale compagnia europea 🇪🇺 nella guida autonoma. So che e' un'affermazione audace, ma ne sono convintissimo. Ci tengo a ringraziare il caso o l'algoritmo di LinkedIn che lo scorso Luglio mi fece apparire un pezzo incredibile del prof Sergio Savaresi (lo copio nel commento sotto). La sua squadra AIDA - Artificial Intelligence Driving Autonomous del @polimi aveva vinto l'Indy Autonomous Challenge a Laguna Seca. In altre parole: la' dove Valentino Rossi fece un sorpasso leggendario su Casey Stoner al Cavatappi, ora trionfava una macchina da corsa Indy del Politecnico di Milano. Tecnologia italiana 🇮🇹 In quella che e' terra di caccia della Silicon Valley. Contatto il prof Savaresi per congratularmi. Ma soprattutto per capire se volesse scaricare a terra le competenze della sua squadra, e andare ben oltre l'accademia per creare una compagnia col potenziale di campione nazionale. Il prof Savaresi lo stava gia' facendo, e aveva coinvolto un imprenditore di razza come @lforesti per fare il salto di qualita'. Raro vedere un professore andare oltre i brevetti e pubblicazioni su carta, e che invece costruisce un prodotto che funziona nella pratica e vince nel mondo. Ancora piu' raro vedere un professore comprendere che una startup ha bisogno di molto piu' di accademici per competere a livello globale. Rispetto 👏 Questa e' una storia che oggi scrive la sua prima pagina. Di questo sono grato alla squadra di A2A Life Venture di Patrick Oungre e Sebastiano Silvestri, che con noi ha creduto sin dall'inizio nel potenziale di Niulinx e senza la quale questa impresa non avrebbe mai visto la luce. E di cuore grazie alla squadra Fondo #ArtificialIntelligence di CDP Venture Capital (Matteo La Naia, Marco Allegretti, Roberto Ceraolo, Carolina Rutta, Stefano Tomei, Alessandro Cordova) per l'impegno costante in questi mesi nel creare basi solide per il futuro di Niulinx. Senza ovviamente contare l'insieme di partner industriali strategici (@Pirelli, @fsitaliane), venture capital (VC Partners), family office (Falck) e consorzi nel settore automobilistico (MOST) che hanno seguito l'investimento. "Smart money", soldi intelligenti in quanto portano esperienze fondamentali per la crescita di Niulinx in svariati campi di applicazione. In bocca al lupo a Luca Foresti e a tutta la squadra di #Niulinx, di cui sono onorato di entrare a far parte come membro del Consiglio di Amministrazione. Enormi aspettative, tantissimo lavoro davanti. Ma sono certo che, col supporto dell'Unione Europea, questa sara' un sfida che possiamo vincere.

The Chinese model of keeping prices low cannot go on forever. This is a screenshot of the recent Zhipu/z Ai price updates. Really wonder who’d pay the same amount as for Claude for a model that performs well in benchmarks but severely underperforms in real life applications

@IterIntellectus No idea how they measure that but that absolutely doesn’t reflect my experience with the models

1) what

@BoWang87 @UHN I talked to an AI researcher last year, and they were skeptical about local models. They suggested that I work on differential privacy. Guess time solves all doubts

Absolutely agree: the future of healthcare AI is on-device intelligence. Small embedded LLMs are getting good enough to bring real medical reasoning directly to the edge — faster, cheaper, more private, and far more scalable. My lab at @UHN has been pushing in this direction too. We recently fine-tuned Qwen3.5 27B, 9B, and 2B on a healthcare benchmark, and found that on-device LLMs can match frontier models on medical diagnosis. That future may arrive sooner than most people think. See more: x.com/BoWang87/statu…

@euromaximal Italy doing worse than Poland and Czech Republic, literally communist countries less than 40 years ago, while being #8 global economy.

There are no words to describe the economic failure that is Portugal. It is now almost as poor as Ukraine, a country literally at war. 50 years after a brutal dictatorship, you’d think an economic miracle would’ve followed like Poland and the Baltics. Total incompetence.

@garrytan Works with Claude Code reviewing its own code, too

ADHD CEO really needs their Autistic CTO

@bcherny Doesn’t this waste context on the current instance? I use a side terminal window for side quests

9/ Use /btw for side queries I use this all the time to answer quick questions while the agent works

I wanted to share a bunch of my favorite hidden and under-utilized features in Claude Code. I'll focus on the ones I use the most. Here goes.

@amritwt @ClementDelangue Tested it and I don't know how they measure but it's so behind Claude. Still really good but def not close

> open source model > is 7x cheaper than opus > is barely behind something you should defn use for all your general tasks

@levie The curve should show pre-COVID numbers; 2021-22 numbers were inflated by cheap cash policies

Jevons paradox is happening in real time. Companies, especially outside of tech, are realizing that they can now afford to take on software projects that they wouldn’t have been able to tackle before because now AI lets them do so. We’re going to start to use software for all new things in the economy because it’s incrementally cheaper to produce. Marketing teams at big companies will have engineers helping to automate workflows. Engineers in life sciences and healthcare will automate research. Small businesses will hire engineers for the first to build better digital experiences. And as long as AI agents still require a human who understands what to prompt, how to review when an agent goes off the rails, how it guide back, how to maintain the system that was built, how to fix the ongoing bugs, and more, we will still have humans managing these agents. This is why all the advice you get of not going into engineering is wrong. The world is going to increasingly be made up of software, and the people that understand it best will be in a strong economic position. This will happen in other roles as well where output goes up and demand increases.

@garrytan Relatable

Yeah we have an AI addiction and it’s awesome

When you realize it is easier to cut cancer than to get it IRB-approved

@daniel__russo22 It’s hard being an average Italian these days

Italy getting good at literally every single sport you could possibly imagine besides the one sport they actually care for is probably the most Italian thing to ever happen in history

The EU regulations are surely doing a great deal leveling AI competition…with disadvantaged countries. EU countries don’t get access to the latest tools but literally any other country on Earth does