Tom Kelly says frontier models stopped getting better on Heidi's tasks. I published the opposite in June. Both stories, the safety data nobody publishes, and my updated Heidi scorecard.
The Nature Medicine article is Open Evidence learning the Bitter Lession (https://en.wikipedia.org/wiki/Bitter_lesson). With regard to AI scribes, ever since Meta released wave2vec in 2020 and then MMS and then the OmniLingual ASR suite into open source, I'm seeing a lot claiming to be privacy-preserving open-source voice transcription apps. It seems a strange space in which to try to be profitable when you can get it for free. But doctors want and need turnkey applications, not something to install with brew or pip. And integration into the EHR is probably a must. That's where the competition will be. Going after dentists and veterinarians is smart, until Epic invades that space, too.
Good points. It’s hard to break through that ceiling if you’re not riding the Epic train, unless you’re Abridge and can simply buy yourself a seat at the table. 😉
That’s also why putting a value on a monopoly like Epic is almost impossible. Maybe it’s worth $100B. Maybe it’s worth $1T. When you control the data layer in healthcare, the upside is enormous. (And yes, I wouldn’t lose much sleep over an FTC suit. Epic has an army of lobbyists and lawyers, and any settlement would likely amount to pocket change.)
Heidi has clearly found a niche. I just think it’s a fairly small one. At some point, if they want to keep growing and survive long term, they’ll probably have to work much more closely with Epic.
P.S. Great link to "The Bitter Lesson." I hadn’t seen it before. That’s exactly what OpenEvidence has gone through. 👍🏼
Great piece. The part about "plateaued" depending on whose task list you measure against is the key one. On narrow clinical tasks it can look flat while the frontier keeps moving elsewhere, and the safety data nobody publishes is what actually decides deployment. Thanks for putting the scorecard out there.
Thanks! My read of Tom Kelly’s comments is that the tasks he’s referring to are the everyday tasks clinicians actually use Heidi for, from documenting patient encounters and generating clinical notes to creating referral letters, discharge summaries, and treatment plans.
I don't have any kind words for these companies. I would rope you away from this toward academia. have you work on hopefully more meaningful things. the harsh truth not much in this world is meaningul. certainly not these bullshit loser startups
The Nature Medicine article is Open Evidence learning the Bitter Lession (https://en.wikipedia.org/wiki/Bitter_lesson). With regard to AI scribes, ever since Meta released wave2vec in 2020 and then MMS and then the OmniLingual ASR suite into open source, I'm seeing a lot claiming to be privacy-preserving open-source voice transcription apps. It seems a strange space in which to try to be profitable when you can get it for free. But doctors want and need turnkey applications, not something to install with brew or pip. And integration into the EHR is probably a must. That's where the competition will be. Going after dentists and veterinarians is smart, until Epic invades that space, too.
Good points. It’s hard to break through that ceiling if you’re not riding the Epic train, unless you’re Abridge and can simply buy yourself a seat at the table. 😉
That’s also why putting a value on a monopoly like Epic is almost impossible. Maybe it’s worth $100B. Maybe it’s worth $1T. When you control the data layer in healthcare, the upside is enormous. (And yes, I wouldn’t lose much sleep over an FTC suit. Epic has an army of lobbyists and lawyers, and any settlement would likely amount to pocket change.)
Heidi has clearly found a niche. I just think it’s a fairly small one. At some point, if they want to keep growing and survive long term, they’ll probably have to work much more closely with Epic.
P.S. Great link to "The Bitter Lesson." I hadn’t seen it before. That’s exactly what OpenEvidence has gone through. 👍🏼
Great piece. The part about "plateaued" depending on whose task list you measure against is the key one. On narrow clinical tasks it can look flat while the frontier keeps moving elsewhere, and the safety data nobody publishes is what actually decides deployment. Thanks for putting the scorecard out there.
Thanks! My read of Tom Kelly’s comments is that the tasks he’s referring to are the everyday tasks clinicians actually use Heidi for, from documenting patient encounters and generating clinical notes to creating referral letters, discharge summaries, and treatment plans.
oh. is this an entire new report today? omg :-)
Haha. Summer’s over. The pace just picked up. And now I have a podcast too. Like and subscribe. 😉
hehehe. you know I do well with those commands. smoothes the PTSD. all these people being nice to me has me on edge constantly ;-)
😂
I don't have any kind words for these companies. I would rope you away from this toward academia. have you work on hopefully more meaningful things. the harsh truth not much in this world is meaningul. certainly not these bullshit loser startups