Nice day to get a fresh coat of wax on the car!

#MirrorFinish

Always impressed with the innovation people are dreaming up.

If you like the For You feed, try this one out.

The Rod of Asclepius with a signal emanating from the top

I decided Medsky was due for a glow-up. Say hello to the new logo!

We're also using a new domain, all Medsky accounts are a .network domain. Major updates moving forward will be coming from @medsky.network, while labels will be handled through @mod.medsky.network.

Other big updates coming soon...

An overhead shot of a brown cardboard pizza box featuring the "Bridge Pizza" logo with an octagonal border and bridge truss design, resting on a dark countertop.

Reading this while eating pizza feels very meta.

Btw - Bridge Pizza is awesome if anyone finds themselves in Missoula.

This. 👇👇👇👇👇👇

I had a chance to play around with this earlier. Claude was already capable of making good slides, but this is seriously impressive!

This looks interesting, I’ll have to try it out!

Transmission isn't limited to harmless preferences. They also observed the same effect with code, reasoning traces, and even misalignment calling for violence, despite rigorous filtering. It only happens if the teacher and student share a base model.

Language models transmit behavioural traits through hidden signals in data - Nature During model distillation, large language models can subtly transmit traits unrelated to the training data.

AI models can subliminally transmit biases to other models during training.

In a new Nature study, the authors saw a 5x increase in 'owl' preferences occurred when a student model was trained only on number sequences from a biased teacher.

We got a cargo e-bike last year, but it is a bit too bulky for commuting.

I’m looking for one that would have a pretty predictable route for my wife to the university and back. We have a pretty good path to the bike trail across town.

I get the battery concerns though!

First look: Also's upcoming e-bike disconnects the pedals and wheels The company bets that software can create a distinct—and better—riding experience.

I'm very interested in seeing how the Rivian bike spin out, Also, fares when it is released.

OpenAI starts offering a biology-tuned LLM GPT-Rosalind is an LLM trained on biology workflows, available in closed access.

OpenAI is moving past generic science models with GPT-Rosalind. It's a biology-tuned LLM trained on 50 specific workflows to connect genotype to phenotype. It is promising to see things like tuning for skepticism about drug targets.
🧬💻

Specialized foundation models for intelligent operating rooms - npj Digital Medicine npj Digital Medicine - Specialized foundation models for intelligent operating rooms

Generalist AI isn't built for the safety-critical, multimodal demands of the OR. The authors here introduce ORQA, which unifies visual and audio data to act as an intelligence core for surgical robots. It beats generalist models at surgical scene perception.
#MedSky #MedAI

‘Science needs defending’: record number of researchers run for office in US mid-terms Many Democrats making the switch to politics are motivated by the Trump administration’s cuts to science — whereas energy and AI are a pull for some Republicans.

🧪 A record wave of researchers is running in the 2026 mid-terms (700+ applications to 314 Action, which recruits scientists and healthcare specialists). Only 3% of state seats are held by STEM professionals, and many of these recruits are looking to help push science-backed policy solutions again.

Quantum computers take on health care: light-sensitive cancer drugs win US$2-million contest Quantum machines are making inroads into biology but have no ‘advantage’ over classical machines yet.

Quantum tech is hitting the wet lab. A team from IBM and Algorithmiq won $2M for simulating light-activated cancer drugs. Hybrid models show qubits can help design less toxic therapies. No $5M prize yet, but the roadmap is forming.
#MedSky

MAGA Is Increasingly Convinced the Trump Assassination Attempt Was Staged Conspiracy theories about the Butler, Pennsylvania, shooting have ramped up in recent weeks as once steadfast Trump supporters turn on the president.

Conspiracies about the Butler assassination attempt used to circulate only on the left. As Trump’s grip on the base slips, MAGA figures like Tim Dillon and Joe Kent are filling the information vacuum with claims of a cover-up.

Bluesky says a sophisticated DDoS attack is to blame for continued app outages but it has not seen any evidence of unauthorized access to private data Sarah Perez / TechCrunch: Bluesky says a sophisticated DDoS attack is to blame for continued app outages but it has not seen any evidence of unauthorized access to private data

Bluesky says a sophisticated DDoS attack is to blame for continued app outages but it has not seen any evidence of unauthorized access to private data (Sarah Perez/TechCrunch)

www.techmeme.com/260417/p17#a...

This was the early alternative to the fail whale.

Oh, to be back in the days when all we had to be wary of was Alf.

Montana may be viewed as a red state, but we were purple for a very long time, and strongly adhered to a "live and let live" mentality. Our constitution was ratified in 1972, is often cited by legal scholars as one of the most progressive and modern state constitutions in the U.S.

The National Institutes of Health (NIH) funds over 300,000 researchers at more than 2,500 institutions across all 50 U.S. states, supporting roughly 390,863 jobs and driving $94
billion in economic activity. I am one of them and am thankful to all who pay the taxes that support this enterprise 🧪

Whoo! 100 days in a row.

Still kind of salty I messed up the 100
Day award.

#Peloton #100Days

A multi-modal foundation model for brain disease diagnosis and medical imaging Accurate diagnosis of brain disorders requires combining medical images with clinical knowledge. Brainfound is a large multi-modal AI model trained on millions of brain CT and MRI images paired with…

Specialized AI for neuroimaging is catching up to experts. Brainfound, trained on 10M slices, matches physician accuracy on MCQs (0.785) and cuts analysis time in half.
#MedSky 🧠

Americans ask AI for health care. Hospitals think the answer is more chatbots. Do you trust AI chatbots for health advice? What about one in your patient portal?

1 in 3 adults use AI for health info, often because they can't afford a doctor or find an open appointment. Hospitals are responding with bots like PatientGPT, but accuracy drops to 33% when users write their own prompts.
#MedSky

Seven deadly sins in artificial intelligence for digital medicine - npj Digital Medicine npj Digital Medicine - Seven deadly sins in artificial intelligence for digital medicine

Seven deadly sins in artificial intelligence for digital medicine.

84.9% of 914 global stakeholders agree: blind trust is the biggest risk for medical AI. Interestingly, concern drops significantly among the youngest users and public authority roles.

#MedSky

Looking forward to giving the Barry D. Pressman, MD,
D’64, M’65, Visiting Lectureship at Dartmouth later this week. Will be speaking about work on agentic AI & will participate in a panel discussion about health AI #dartmouth

I've had some conversations with prolific writers who have folded LLMs into their writing process. They often use them just like that, sounding boards that are intimately familiar with all their prior work, and it helps sharpen the topic they want to write about. They still author the final product.

An agentic AI system for automated pharmacogenomic recommendation generation - npj Digital Medicine npj Digital Medicine - An agentic AI system for automated pharmacogenomic recommendation generation

Manual curation of genomic guidelines is slow, but an agentic AI system scored 9.0/10 in expert reviews (beating GPT-5). By grounding LLMs in structured data, researchers were able to double coverage to 573 gene-drug pairs. 🧬💻

You aren't the only one with that line of thinking: Supervising Resident AI Use Without Losing the Learning jgme.kglmeridian.com/view/journal...

The best option is checking the LLM model card when it gets released, that gives you a peak under the hood for the model. For the reasoning optimized approaches, there are a variety of options, like Chain-of-Thought Reinforcement Learning, where the model is rewarded for "showing its work".