Juan Sebastian Lozano
@jslozano.bsky.social
Wandering awed about on a splintered wreck. CTO @ cofactorai.com . Z-fellow. Prev: Quantum algorithms @ D-Wave, Empirical Studies of Conflict @ Princeton, Math @ UT Austin.
Not every bubble is the same in terms of the impact of the underlying technology. Did railroads have a massive impact on America in many ways beyond finance before and after 1873? Yes. Was there a bubble that popped in 1873? Also yes.
October 27, 2025 at 10:00 PM
Not every bubble is the same in terms of the impact of the underlying technology. Did railroads have a massive impact on America in many ways beyond finance before and after 1873? Yes. Was there a bubble that popped in 1873? Also yes.
Everyone in tech is scarred by the dot com bubble, and the bubble description is explicitly about the insane amounts of capital being used in AI right now. The terminology is not arbitrary, it's a financialized sector talking in the terms of finance.
October 27, 2025 at 10:00 PM
Everyone in tech is scarred by the dot com bubble, and the bubble description is explicitly about the insane amounts of capital being used in AI right now. The terminology is not arbitrary, it's a financialized sector talking in the terms of finance.
I don't think ambient scribes are fully LLM based, Kaiser has been using them for a while before ChatGPT even existed, based off of the same models that powered Dragon voice-to-text. I wonder if Nuance (the company that makes the scribe) has added LLMs.
July 1, 2025 at 9:02 AM
I don't think ambient scribes are fully LLM based, Kaiser has been using them for a while before ChatGPT even existed, based off of the same models that powered Dragon voice-to-text. I wonder if Nuance (the company that makes the scribe) has added LLMs.
Very interesting insights, what software are you using? I've heard that some are way better than others at accuracy, but not all of them organize the note well.
July 1, 2025 at 8:58 AM
Very interesting insights, what software are you using? I've heard that some are way better than others at accuracy, but not all of them organize the note well.
Anyway if you can't tell I've been thinking about how to make a recommender feed ever since AT Proto came out lol but I think only the official bsky app has enough usership to be able to get this right.
June 30, 2025 at 5:49 AM
Anyway if you can't tell I've been thinking about how to make a recommender feed ever since AT Proto came out lol but I think only the official bsky app has enough usership to be able to get this right.
Another really important thing to consider is that a like is a socially visible form of expression, as is following, and posting even more so. This can be good (it's unlikely to simply recommend rage bait) but also limiting bc for people using their they get an employer friendly version of a feed.
June 30, 2025 at 5:44 AM
Another really important thing to consider is that a like is a socially visible form of expression, as is following, and posting even more so. This can be good (it's unlikely to simply recommend rage bait) but also limiting bc for people using their they get an employer friendly version of a feed.
I.e. The post where a like/non-like have the highest correlation accross the cluster. Then you rank these discriminative posts by the normal like prediction algo and show them in some % of the feed. The % should be higher for new users and lower for more established users.
June 30, 2025 at 5:25 AM
I.e. The post where a like/non-like have the highest correlation accross the cluster. Then you rank these discriminative posts by the normal like prediction algo and show them in some % of the feed. The % should be higher for new users and lower for more established users.
Another approach which can make discover feel magical is creating an algorithmic feed which mixes high probability liked posts with less certain but highly predictive posts.
This would mean that you do semantic clustering of posts and then within each cluster find the highest discrimitive post.
This would mean that you do semantic clustering of posts and then within each cluster find the highest discrimitive post.
June 30, 2025 at 5:25 AM
Another approach which can make discover feel magical is creating an algorithmic feed which mixes high probability liked posts with less certain but highly predictive posts.
This would mean that you do semantic clustering of posts and then within each cluster find the highest discrimitive post.
This would mean that you do semantic clustering of posts and then within each cluster find the highest discrimitive post.
Another simple approach is to generate their user embedding from the set of posts they liked, inversely weighted by when they liked them.
If you collected interaction time data (👀 TikTok) you could generate this embedding by the seen posts weighted by interaction time.
Many different approaches.
If you collected interaction time data (👀 TikTok) you could generate this embedding by the seen posts weighted by interaction time.
Many different approaches.
June 30, 2025 at 5:25 AM
Another simple approach is to generate their user embedding from the set of posts they liked, inversely weighted by when they liked them.
If you collected interaction time data (👀 TikTok) you could generate this embedding by the seen posts weighted by interaction time.
Many different approaches.
If you collected interaction time data (👀 TikTok) you could generate this embedding by the seen posts weighted by interaction time.
Many different approaches.
Now you use the non ML based candidate group sampling algos for this group, giving you something like a hybrid topic + social set of posts.
Of course most users are not posting enough to actually generate a useful user embedding this way, but you can do this all sorts of ways.
Of course most users are not posting enough to actually generate a useful user embedding this way, but you can do this all sorts of ways.
June 30, 2025 at 5:25 AM
Now you use the non ML based candidate group sampling algos for this group, giving you something like a hybrid topic + social set of posts.
Of course most users are not posting enough to actually generate a useful user embedding this way, but you can do this all sorts of ways.
Of course most users are not posting enough to actually generate a useful user embedding this way, but you can do this all sorts of ways.
A naive but illustrative way of doing this is for every user average the embeddings of all of their posts, this gives you a rough "user embedding". Now you create a cohort of the user of those people who's user embeddings are in a neighborhood of size \delta.
June 30, 2025 at 5:25 AM
A naive but illustrative way of doing this is for every user average the embeddings of all of their posts, this gives you a rough "user embedding". Now you create a cohort of the user of those people who's user embeddings are in a neighborhood of size \delta.
I haven't interacted with bit which I really want to see. Usually this is done by combining the two approaches for candidate groups you talked about: it's not follow based OR topic, it's _both_.
To do this what you want to do is create a semantic clustering algorithm for _users_.
To do this what you want to do is create a semantic clustering algorithm for _users_.
June 30, 2025 at 5:25 AM
I haven't interacted with bit which I really want to see. Usually this is done by combining the two approaches for candidate groups you talked about: it's not follow based OR topic, it's _both_.
To do this what you want to do is create a semantic clustering algorithm for _users_.
To do this what you want to do is create a semantic clustering algorithm for _users_.
You need to build a dynamic topic model, which you can do by dynamically identifying the clusters only, or by identifying clusters + updating embeddings, which I think is more powerful.
Moreover, what can feel "magical" about a truly good recommender system is when it recommends something that...
Moreover, what can feel "magical" about a truly good recommender system is when it recommends something that...
June 30, 2025 at 5:25 AM
You need to build a dynamic topic model, which you can do by dynamically identifying the clusters only, or by identifying clusters + updating embeddings, which I think is more powerful.
Moreover, what can feel "magical" about a truly good recommender system is when it recommends something that...
Moreover, what can feel "magical" about a truly good recommender system is when it recommends something that...
Super interesting thread!
As far as generating candidate groups: topic similarity using an off the shelf embedding model with a fixed training cutoff means that you won't be able to topic cluster along newer slang, memes, etc, which is super important for social media.
As far as generating candidate groups: topic similarity using an off the shelf embedding model with a fixed training cutoff means that you won't be able to topic cluster along newer slang, memes, etc, which is super important for social media.
June 30, 2025 at 5:25 AM
Super interesting thread!
As far as generating candidate groups: topic similarity using an off the shelf embedding model with a fixed training cutoff means that you won't be able to topic cluster along newer slang, memes, etc, which is super important for social media.
As far as generating candidate groups: topic similarity using an off the shelf embedding model with a fixed training cutoff means that you won't be able to topic cluster along newer slang, memes, etc, which is super important for social media.
The bill is going to kick off a massive number of people from insurance, and make those with insurance (both commercial and governmental) substantially worse off, including via the AI regulation ban. bsky.app/profile/jslo...
Don't want AI to auto-deny your health insurance claim?
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
Enter your address to get more accurate results
www.deny-ai.com
June 4, 2025 at 2:57 PM
The bill is going to kick off a massive number of people from insurance, and make those with insurance (both commercial and governmental) substantially worse off, including via the AI regulation ban. bsky.app/profile/jslo...
Not only is it going to kick off a massive number of people from insurance, and make those with insurance (both commercial and governmental) substantially worse off, including via the AI regulation ban. bsky.app/profile/jslo...
Don't want AI to auto-deny your health insurance claim?
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
Enter your address to get more accurate results
www.deny-ai.com
June 4, 2025 at 2:55 PM
Not only is it going to kick off a massive number of people from insurance, and make those with insurance (both commercial and governmental) substantially worse off, including via the AI regulation ban. bsky.app/profile/jslo...
It's going to kick off a massive number of people from insurance, and make those with insurance (both commercial and governmental) substantially worse off, including via the AI regulation ban. bsky.app/profile/jslo...
Don't want AI to auto-deny your health insurance claim?
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
Enter your address to get more accurate results
www.deny-ai.com
June 4, 2025 at 2:54 PM
It's going to kick off a massive number of people from insurance, and make those with insurance (both commercial and governmental) substantially worse off, including via the AI regulation ban. bsky.app/profile/jslo...
We're going to kick off a massive number of people from insurance, and make those with insurance (both commercial and governmental) substantially worse off, including via the AI regulation ban. bsky.app/profile/jslo...
Don't want AI to auto-deny your health insurance claim?
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
Enter your address to get more accurate results
www.deny-ai.com
June 4, 2025 at 2:53 PM
We're going to kick off a massive number of people from insurance, and make those with insurance (both commercial and governmental) substantially worse off, including via the AI regulation ban. bsky.app/profile/jslo...
Don't want AI to auto-deny your health insurance claim?
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
June 4, 2025 at 2:44 PM
Don't want AI to auto-deny your health insurance claim?
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
The Big Beautiful Bill will give insurance companies the ability to deny your claims without a human ever reading them.
We need to make sure this doesn't happen. CALL TODAY: www.deny-ai.com/call-your-re...
It's crazy how far reaching this is, it would also effect health insurance.
May 15, 2025 at 1:50 AM
It's crazy how far reaching this is, it would also effect health insurance.
May 15, 2025 at 1:48 AM
This thread has more resources for the non-insurance implications of this bill
Update: The TERRIBLE 10-Year Ban on State AI Regulations advanced House Committee. What does this mean?
It means we BLAST the HELL out of their phones! Call your House Members AND Senators today! Called a few days ago? Call again!
Tell them NO on 10-year State AI Regulation Ban!
(Info in thread)
It means we BLAST the HELL out of their phones! Call your House Members AND Senators today! Called a few days ago? Call again!
Tell them NO on 10-year State AI Regulation Ban!
(Info in thread)
May 15, 2025 at 1:44 AM
This thread has more resources for the non-insurance implications of this bill
This also includes state legislation to prevent insurance companies from automatically denying your claims using AI. Here is a website to find and call your reps to talk to them about the insurance aspect of this issue:
www.deny-ai.com/call-your-re...
www.deny-ai.com/call-your-re...
Enter your address to get more accurate results
www.deny-ai.com
May 13, 2025 at 6:31 PM
This also includes state legislation to prevent insurance companies from automatically denying your claims using AI. Here is a website to find and call your reps to talk to them about the insurance aspect of this issue:
www.deny-ai.com/call-your-re...
www.deny-ai.com/call-your-re...
This also includes state legislation to prevent insurance companies from automatically denying your claims using AI. Here is a website to find and call your reps to talk to them about this issue:
www.deny-ai.com/call-your-re...
www.deny-ai.com/call-your-re...
Enter your address to get more accurate results
www.deny-ai.com
May 13, 2025 at 6:02 PM
This also includes state legislation to prevent insurance companies from automatically denying your claims using AI. Here is a website to find and call your reps to talk to them about this issue:
www.deny-ai.com/call-your-re...
www.deny-ai.com/call-your-re...