I am hosting Bits in Bio Presents, A Mixer in Miami. Join me! Bits in Bio is a global community of more than 13,000 people working where software, AI, and biology meet. On Wednesday, November 18, we're bringing it to Miami for the first time, 6:30 to 9:00 PM at Cerveceria La Tropical in Wynwood. No panels. No pitches. No slides. Just everyone in South Florida building at the intersection of computation and biology, in one room. Researchers, clinicians, founders, engineers, students, investors, and anyone trying to break in. UM, FIU, the Miller School, Nova Southeastern, and everyone in between. Why La Tropical? It started in Havana in 1888, became Cuba's most beloved brewery, and was reborn here in Wynwood. As the son of Cuban immigrants starting something new in this city, that one's personal. Coming alone? Good. Most people will be. Introducing people is the job that night, so come find me. Huge thanks to @AaronBlotnick and @bitsinbio for making this happen. luma.com/uk92lg9g via @LumaHQ

If you can't show what failed, you can't prove what worked.

Prospective: you predict, then the answer comes in. Retrospective: the answer was already in, then you predicted. Only one of those is a test.

If you only report the wins, your hit rate is a highlight reel.

Three hits out of ten is a breakthrough 🍾 Three out of a hundred thousand is a lottery ticket. Same press release 😵‍💫

Every hit rate is a fraction. Most announcements only show you the top half.

A diligence tip for anyone buying on an AI model's performance: ask for the predictions, not the score. A score is a summary someone else computed. The predictions, matched to the measured results and the exact split, let you recompute it yourself. If the number changes when you do, you've learned more than any slide could tell you. And if nobody can give you the predictions, you've learned something too🧐

Every "AI discovered" headline is missing the same number: how many things it tried first.

Did you know Florida has the second-largest life sciences industry in the U.S. by company count? Most people in it have never met each other. On Thursday, November 19, @bitsinbio is bringing that community under one roof in Miami to fix that. Bits in Bio Miami Kickoff Thursday, November 19 | 6:30 to 9:00 PM Cerveceria La Tropical | 42 NE 25th St, Wynwood RSVP: luma.com/uk92lg9g What it is: the first-ever Bits in Bio Miami event. No panels. No pitches. No slide decks. Nobody will make you stand up and introduce yourself. Where: a brewery that started in Havana in 1888 and got a second life in Wynwood. A new chapter for an old brewery felt like the right place to start a new chapter for Miami biotech. Who it's for: anyone working on or interested in biotech, and the people building technology for it. Wet lab, computational, clinical, founders, students, investors, and the curious. Coming alone? Good. Most people are. Find an organizer and we'll introduce you to someone. That is the job. Dress code: it's Miami in November. Shorts are correct. Why now: I keep meeting brilliant people in South Florida biotech. Usually it turns out they live 20 minutes from each other and have never met. That seemed fixable. Some quick numbers on the Sunshine State: - More medical device manufacturers than any state but one - Over $136B in life sciences GRP in 2025 - The Miami Health District is the second-largest medical district in the country And a partial list of who's already here: - @umiamimedicine, @FIU, @NSUFlorida, and the @Diabetes_DRI - Converge Miami and @cicofficial, home to South Florida's first shared wet lab - @OpenEvidence, the $12B medical AI company headquartered in Miami - OPKO Health, @BioTissueInc, @organabio, @Click_Tx, Noven, Cordis, and @BeckmanCoulter - Miami Biotech Collective, Life Sciences South Florida, @BioFlorida, and the @beaconcouncil - Investors like @Ocean_Azul Huge thanks to @RippaSatss for co-hosting and pushing to get this chapter off the ground.

Stanford just published a virtual biotech in Science: up to 37,000 AI agents organized like a drug company, with a chief-scientist agent directing specialist teams. They cataloged about 56,000 clinical trials in under a week. The detail worth noticing is how they described their best result. When a drugmaker independently arrived at the same lung cancer strategy, the lead researcher called it an independent, third-party validation. That's the one thing agents can't do for themselves. As they generate more of the science, independent checking stops being nice to have and becomes what makes any of it usable.

AI is making scientific analysis cheaper and faster. It isn't making it more trustworthy 💀 That part was never going to come free with the speed.

An AI agent that loses a connection mid-request will just try again. If the system on the other end isn't built for that, the same experiment gets run and logged twice. At scale, the record stops telling you how many attempts it really took. The denominator problem doesn't go away with agents. It gets automated.

Some evaluations shouldn't be run. LP-GATE checks first whether a submission can be evaluated at all, and turns it away with a reason if it can't. A confident answer to the wrong question is worse than no answer.

114 AI Drug Discovery partnerships were signed in 2025, up from 84 the year before Most of the headline value is milestones, Not the Cash. Pharma pays a little up front for access and a lot later for proof. What happens between those two payments is the part nobody can measure?

If a faster method misses the right answer some of the time, it isn't an optimization. It's a different answer, delivered sooner💀

Held-out test sets are becoming table stakes. The next question is harder: when was success defined, and who was holding it?

Novo Nordisk is deploying Claude across drug discovery workflows. That moves the bottleneck. Generating a computational claim is getting cheap. Checking one isn't 👀 When claims get produced faster than any team can review them, verification stops being a service and becomes infrastructure.

A benchmark number means something only if you know three dates. When the test was frozen? When the model was trained? When success was defined? Ask for all three🧠

A clean benchmark result has an expiration date. Nobody prints it. When a model is evaluated, the test set is new to it. That's the whole point. But public data doesn't stand still. The test set gets posted, cited, scraped, discussed, and folded into the next round of training corpora. A year later, a newer model runs the same benchmark and scores higher. Some of that is real progress. Some of it is the test set leaking into training through the ordinary movement of the internet, with nobody doing anything wrong. From the outside the two look identical. Same benchmark, higher number. This is why an evaluation has to be bound to a dated snapshot of what the model could have seen. Without the date, a result isn't a measurement. It's a number whose meaning changes every time the corpus does.

It's time.