Michael Italia

Working on helping healthcare and life science ease into the 21st century. Currently building something new.

[Michael Italia @mitalia]

I know everyone is complaining about Siri. I tried to have an open mind but this is just comically bad. Adding songs to a playlist is a task that should be a slam dunk for @Apple to get right.

Several HOURS went by while Siri claimed to be “working”.

[Michael Italia @mitalia]

Two real ones from today:
1) I’m building a massive curated dataset that used to require teams and people working for years. I need research agents working in parallel around the clock to refine and add to the data or I’ll never be able to launch it.

2) Sweep the entire codebase and look for any comments that indicate unhandled edge cases or to-do items. For each one write ticket, then a failing test, and finally implement it. Agent swarm gets this done overnight vs days of grinding away one by one.

[Michael Italia @mitalia]

The thing that people who are embedded in the software industry forget is that ordinary people do not wake up in the morning eager to use and try out new software. They have a job that needs to be done. Sometimes, software is the way that gets done. I learned (painfully!) working in healthcare that no doctor is ever eager to hear about your new piece of software that purports to solve their problem. They just want to help their patients. So in that arena, I think two things are true at the same time: 1) Doctors will not want to vibe code an Electronic Health Record. and 2) AI will completely eliminate this overly complex software category entirely. The result won't look anything like what we have today. To pick a silly illustration: when Picard asks the Star Trek computer to analyze something, he doesn't ask it to build him a python script to do statistical analysis and data integration off thousands of sensor outputs ("make no mistakes!"). He just gives it a job to do and it does it. This future is hard to imagine for people who are anchored to the world as it exists today.

[Michael Italia @mitalia]

One of the things I keep going back and forth on is whether abstractions are still helpful. In human maintained code, you are constantly looking for ways to create shared components and useful abstractions that reduce the size of your codebase and make it easier to understand. I’m not sure agents need any of that in the same way humans do. Which is a deeply uncomfortable feeling.

[Michael Italia @mitalia]

@dzhng I’ve often assumed there are benefits to things like react because 1) the LLMs have been trained on them and 2) they implement a lot of nuanced things you may not realize you need unless you really know front end development super well. So I guess the LLMs are just that good now?

[Michael Italia @mitalia]

@andrewmccalip I’ve been thinking about this a lot too. Banks have know your customer rules. Why can’t model providers implement something similar?

[Michael Italia @mitalia]

So much of computing is centered around the archetype of a “user” who runs some form of prepackaged software (open source or commercial). The fact that a majority of people will be not just users of software but builders (via ai) means a big re-think of what an operating system needs to include. This is the biggest transformation since the GUI.

[Michael Italia @mitalia]

@dhh @spacexai @grok Omarchy phone project seems like an obvious next step.

[Michael Italia @mitalia]

@Steve_Yegge I have tried to use Claude code without beads on client projects and it’s shocking how much worse the overall experience is when I don’t have beads. The quality of LLM decisions, especially on micro-architecture and systems design, is objectively worse.

[Michael Italia @mitalia]

Sorry to disappoint you but there’s actually quite a number of us that check those boxes in the industry. The structural problems combined with inertia and entrenched interests make getting anything innovative done agonizingly slow and painful. This is why many healthcare startups die trying to get to PMF.

[Michael Italia @mitalia]

Lots of unhelpful comments here. I had this problem on an iPad mini that I only use for books and light tasks. It’s not all the OS. The l cache wasn’t getting cleaned up. I backed up, wiped, and restored from backup and now I have plenty of free space. It’s an outrage that Apple allows this to occur with no option other than “nuclear.”

[Michael Italia @mitalia]

@BishopBarron @TheoVon I think this is a photo of what Pope Francis meant by the Church being a “field hospital”.

[Michael Italia @mitalia]

Research area I'd love to see this team go after: There's been a hypothesis for years that synonymous mutations --> rare codons --> transcriptional stalling (from tRNA exhaustion) --> protein mis-folding. I think there's probably some surprising undiscovered SNPs in this category.

[Michael Italia @mitalia]

@clairevo If it’s cheap to fix, spend zero time reviewing (lots of front end code is like this). If it’s expensive to fix later, be more careful (eg bad data models lead to painful migrations and refactoring). The faster LLMs get, the less the latter category matters.

[Michael Italia @mitalia]

"A man who touches no medicine and sits between the two will decide the price in secret"

"What's this man called?"

"A Pharmacy Benefit Manager"
🔥🤣

[CrowdHealth @JoinCrowdHealth]

Washington's dream for healthcare (inspired by Nate Bargatze).

[Michael Italia @mitalia]

@RoupenMD Modernizing clinical trial operations is the single largest thing that would give more patients access to new therapies and get those drugs to market faster. There is io much wasted time and inefficiency.

[Michael Italia @mitalia]

I worked on a commercial system to do this. Even with full access to EHR data it’s extremely hard to eliminate both false positives and false negatives. The problem is missing data and how missingness is handled in the context of inclusion and exclusion criteria. Temporality is also an issue: if an inclusion criterion says a patient’s ALT must be < X, what time range do you use to make sure they meet the criterion? LLMs are a huge advance for lots of medical records tasks, but I’m skeptical this is one.

[Michael Italia @mitalia]

Opus has been doing this all week to me. It’s aggravating how many tokens it wastes on self-flagellation because of its own bad assumptions and sloppy work. It’s also overly dramatic, “My assumption is incorrect, and I want to be straight with you: things are far worse than you know”. It’s might be the most annoying coworker ever.

[Michael Italia @mitalia]

Having spent lots of time in and around biology research labs, I’ve always thought it almost comical that some of the most intelligent people on earth spend so much time on what amounts to manual labor. The actual scientific discovery part is a tiny, tiny slice of the wall clock time. For people who don’t know this world: it would be like programmers having to build their computer from component parts every day just to get some work done. Except sometimes you make a tiny mistake and the while day is ruined.

[Michael Italia @mitalia]

For many of us who enjoy computing, the idea that one has to pick (and stick with) a "tribe" was instilled at an early age. As an 11 year old I would have fights on PC BBS systems over the superiority of the Amiga's graphics over PC's VGA. That attitude is so unhealthy! There's a lot of cool stuff out there if you allow yourself the chance to experience it!