Ian Sefferman
@iseff.com
👨💻 ai-first software engineer.
✨ applying ai to make healthcare cheaper at goodbill.com.
✋ detroiter.
🇮🇱 zionist.
🌲 yellowstone.
🎣 fly fisherman.
🐶 english setter lover and wannabe bird dog trainer.
🌐 more at iseff.com.
✨ applying ai to make healthcare cheaper at goodbill.com.
✋ detroiter.
🇮🇱 zionist.
🌲 yellowstone.
🎣 fly fisherman.
🐶 english setter lover and wannabe bird dog trainer.
🌐 more at iseff.com.
Every once in a while I still get hit over the head and realize programming will never be the same.
Instead of writing deterministic code with strong backgrounds in algorithms, it’s going to be who makes the AI feel better about itself and willing to answer queries better.
Instead of writing deterministic code with strong backgrounds in algorithms, it’s going to be who makes the AI feel better about itself and willing to answer queries better.
This paper is wild - a Stanford team shows the simplest way to make an open LLM into a reasoning model
They used just 1,000 carefully curated reasoning examples & a trick where if the model tries to stop thinking, they append "Wait" to force it to continue. Near o1 at math. arxiv.org/pdf/2501.19393
They used just 1,000 carefully curated reasoning examples & a trick where if the model tries to stop thinking, they append "Wait" to force it to continue. Near o1 at math. arxiv.org/pdf/2501.19393
February 7, 2025 at 3:10 AM
Every once in a while I still get hit over the head and realize programming will never be the same.
Instead of writing deterministic code with strong backgrounds in algorithms, it’s going to be who makes the AI feel better about itself and willing to answer queries better.
Instead of writing deterministic code with strong backgrounds in algorithms, it’s going to be who makes the AI feel better about itself and willing to answer queries better.
If we really do take control of Gaza, we should build a technology corridor there and work with Israel to have a totally open border for trade and work, so that we can have Israelis working “in the US” on technology from Israel with no restrictions or red tape whatsoever.
February 5, 2025 at 12:39 AM
If we really do take control of Gaza, we should build a technology corridor there and work with Israel to have a totally open border for trade and work, so that we can have Israelis working “in the US” on technology from Israel with no restrictions or red tape whatsoever.
People can’t seem to just be normal. We can hire based on merit and teach history at the same time.
January 25, 2025 at 3:25 PM
People can’t seem to just be normal. We can hire based on merit and teach history at the same time.
New Lock Screen. Name that movie.
January 25, 2025 at 1:43 AM
New Lock Screen. Name that movie.
Remember what happened after Trump disbanded the Global Health Security and Biodefense team?
I have not seen people talking about this so far w/ 47's cuts to roles.
"Members of the Aviation Security Advisory Committee received a memo Tuesday" eliminating their role. This is a mandated committee by law dating back to PanAm 103. @jonostrower.com @wandrme.paxex.aero
apnews.com/article/coas...
"Members of the Aviation Security Advisory Committee received a memo Tuesday" eliminating their role. This is a mandated committee by law dating back to PanAm 103. @jonostrower.com @wandrme.paxex.aero
apnews.com/article/coas...
Trump fires heads of TSA, Coast Guard and guts key aviation safety advisory committee
President Donald Trump has moved quickly to remake the Department of Homeland Security by firing the heads of the Transportation Security Administration and Coast Guard before their terms are up while...
apnews.com
January 24, 2025 at 2:47 AM
Remember what happened after Trump disbanded the Global Health Security and Biodefense team?
Attention is all you need.
(also… wtf is wrong with these people)
(also… wtf is wrong with these people)
January 24, 2025 at 2:07 AM
Attention is all you need.
(also… wtf is wrong with these people)
(also… wtf is wrong with these people)
I’m worried the dog has a problem.
January 11, 2025 at 4:04 PM
I’m worried the dog has a problem.
What can I say, Sarahs love me on this app.
January 7, 2025 at 4:33 PM
What can I say, Sarahs love me on this app.
s/13yo/college/
s/debian/macos/
s/debian/macos/
January 4, 2025 at 4:55 PM
s/13yo/college/
s/debian/macos/
s/debian/macos/
i wonder if zuck listens to the social network soundtrack when he codes?
January 2, 2025 at 2:43 PM
i wonder if zuck listens to the social network soundtrack when he codes?
First ride in a Waymo was impressive in how unimpressive and normal the ride was.
December 26, 2024 at 8:37 PM
First ride in a Waymo was impressive in how unimpressive and normal the ride was.
👻🎄
Why do developers confuse Halloween for Christmas?
Because OCT31 = DEC25.
Happy Halloween 🎃👻
Because OCT31 = DEC25.
Happy Halloween 🎃👻
December 25, 2024 at 11:30 PM
👻🎄
Interesting to think about the difficulty in creating *good* evals.
A look at the more challenging AI evaluations emerging in response to the rapid progress of models, including FrontierMath, Humanity's Last Exam, and RE-Bench (Tharin Pillay/Time)
Main Link | Techmeme Permalink
Main Link | Techmeme Permalink
December 25, 2024 at 11:27 PM
Interesting to think about the difficulty in creating *good* evals.
Very interesting. Makes me think FHIR should be extended to clinical guidelines for interoperability.
December 24, 2024 at 1:48 AM
Very interesting. Makes me think FHIR should be extended to clinical guidelines for interoperability.
While MoCA is the standard, it’s still not very good. But this is objectively hilarious.
Susceptibility of large language models to cognitive impairment. #bmj
"With the exception of ChatGPT 4o, almost all large language models subjected to the MoCA test showed signs of mild cognitive impairment."
www.bmj.com/content/387/...
"With the exception of ChatGPT 4o, almost all large language models subjected to the MoCA test showed signs of mild cognitive impairment."
www.bmj.com/content/387/...
December 24, 2024 at 1:43 AM
While MoCA is the standard, it’s still not very good. But this is objectively hilarious.
Oregon v Ohio State on New Year's Day at the Rose Bowl is oddly familiar.
Oregon v Ohio State, both Big Ten teams, playing in the CFP quarterfinals is oddly different.
Oregon v Ohio State, both Big Ten teams, playing in the CFP quarterfinals is oddly different.
December 22, 2024 at 1:14 PM
Oregon v Ohio State on New Year's Day at the Rose Bowl is oddly familiar.
Oregon v Ohio State, both Big Ten teams, playing in the CFP quarterfinals is oddly different.
Oregon v Ohio State, both Big Ten teams, playing in the CFP quarterfinals is oddly different.
To deploy at 5:10p on a Friday, or not to deploy at 5:10p on a Friday?
*That* is the question.
*That* is the question.
December 20, 2024 at 10:10 PM
To deploy at 5:10p on a Friday, or not to deploy at 5:10p on a Friday?
*That* is the question.
*That* is the question.
hooboy, here we go!
let's see how fast the costs can come down now.
let's see how fast the costs can come down now.
87.5% ARC-AGI besting human 85% human performance
December 20, 2024 at 7:23 PM
hooboy, here we go!
let's see how fast the costs can come down now.
let's see how fast the costs can come down now.
the rate of ai progress is incredible.
“oh, that doesn't work with this model? just wait 3 months for the next model.”
“oh, that doesn't work with this model? just wait 3 months for the next model.”
December 19, 2024 at 5:18 PM
the rate of ai progress is incredible.
“oh, that doesn't work with this model? just wait 3 months for the next model.”
“oh, that doesn't work with this model? just wait 3 months for the next model.”
As LLMs/AI continue to improve at math and reasoning, I'm excited for them to be used to real-time fact-check scientific studies.
I suspect there are far more errors in papers than we expect.
arstechnica.com/health/2024/...
I suspect there are far more errors in papers than we expect.
arstechnica.com/health/2024/...
Huge math error corrected in black plastic study; authors say it doesn’t matter
Correction issued for black plastic study that had people tossing spatulas.
arstechnica.com
December 17, 2024 at 5:47 PM
As LLMs/AI continue to improve at math and reasoning, I'm excited for them to be used to real-time fact-check scientific studies.
I suspect there are far more errors in papers than we expect.
arstechnica.com/health/2024/...
I suspect there are far more errors in papers than we expect.
arstechnica.com/health/2024/...
Startup hilarity:
Us: how many members do you have?
Them: about half a million.
(the next week)
Us: remind us, how many members do you have?
Them: about a million.
(the next day)
Us: so those million... is that currently enrolled?
Them: oh no. We have 10 currently enrolled.
Us: how many members do you have?
Them: about half a million.
(the next week)
Us: remind us, how many members do you have?
Them: about a million.
(the next day)
Us: so those million... is that currently enrolled?
Them: oh no. We have 10 currently enrolled.
December 16, 2024 at 6:17 PM
Startup hilarity:
Us: how many members do you have?
Them: about half a million.
(the next week)
Us: remind us, how many members do you have?
Them: about a million.
(the next day)
Us: so those million... is that currently enrolled?
Them: oh no. We have 10 currently enrolled.
Us: how many members do you have?
Them: about half a million.
(the next week)
Us: remind us, how many members do you have?
Them: about a million.
(the next day)
Us: so those million... is that currently enrolled?
Them: oh no. We have 10 currently enrolled.
Case in point: It will never not amaze me that Stripe built a $100b business bc they simply had a better API and developer experience than anyone else.
This, a thousand times this!
Be as descriptive as possible with error messages in your customer facing APIs.
Everyone makes mistakes when integrating with a new service, and by providing all the info they need to correct the error quickly, not only are you saving a support email, ur gaining a fan
Be as descriptive as possible with error messages in your customer facing APIs.
Everyone makes mistakes when integrating with a new service, and by providing all the info they need to correct the error quickly, not only are you saving a support email, ur gaining a fan
December 14, 2024 at 2:05 AM
Case in point: It will never not amaze me that Stripe built a $100b business bc they simply had a better API and developer experience than anyone else.
this is so good. so so good.
A Waymo robotaxi got stuck in a roundabout loop
A Waymo robotaxi got stuck in a roundabout loop
Roundabouts. They trapped Clark Griswold in National Lampoon’s European Vacation and now they’ve tripped up a Waymo robotaxi. A video is circulating on social media showing a Waymo robotaxi going round and round on a roundabout — as if it is stuck in a…
tcrn.ch
December 13, 2024 at 3:46 PM
this is so good. so so good.
Nice to see Goodbill listed in BI's 10 Startups Using AI to Disrupt Healthcare Payments. Lots of work to be done. We're steadfast in our mission to reduce hospital bill costs in this country and are chipping away at the problem a little bit at a time.
www.businessinsider.com/startups-usi...
www.businessinsider.com/startups-usi...
These 10 startups are using AI to disrupt healthcare payments as public outrage toward insurers mounts
Health insurers are increasingly denying paying for patient care. A growing crop of startups think AI can help.
www.businessinsider.com
December 11, 2024 at 7:54 PM
Nice to see Goodbill listed in BI's 10 Startups Using AI to Disrupt Healthcare Payments. Lots of work to be done. We're steadfast in our mission to reduce hospital bill costs in this country and are chipping away at the problem a little bit at a time.
www.businessinsider.com/startups-usi...
www.businessinsider.com/startups-usi...