I genuinely don’t understand why Biden thinks it’s a good idea to propose something to Trump, who tends to reflexively take the opposite stance of whatever his rival’s positions are.
i always read this post to keep in the loop and always leave with a sense that AI isn't coming soon or doing anything meaningful soon. which is funny because i feel as if the vibe of the post is saying the opposite. i just don't see it in the numbers yet anywhere, and AI is just not useful for my own life. It doesn't move the needle for me personally, and I don't see it move the needle in the economy broadly either
There have been lots of companies trying to automate scheduling, including in the 2010s as you mention, including the past owners of x.ai. Not an easy problem to hit the requires accuracy rate.
Just wanted re-state to those wondering why this exists: The value add for this podcast is that I go through and give all the uniquely quoted people their own consistent voices throughout the episode, and describe images for the audio medium. This is more than just a simple TTS pass. Please give it a listen if you are the kind of person who finds audio more accessible.
> Never? No, never. What, never? Well, actually all the time.
> I also find it hard to believe that students are this slow, especially given this is a very low bar - it’s whether you even once asked for ‘help’ at all, in any form.
I get what you are saying. I can also understand why the survey results look the way it does without using the phrase "the kids are lying."
It's happened a few times now. I am helping my kids with their homework and I pull out my phone to either look up something on Gemini/Claude/ChatGPT or help me answer a question. They'd look at me, aghast, and say, "Dad! We're not suppose to use AI!"
I tried to explain there is a difference between using an AI to "Do your work for you and not think" and using an AI to "help you learn." But teenagers have this amazing superpower of shutting you out before you can even complete a sentence. It's 2025. We literally have thinking machines now! Yet, teens still don't listen to their parents.
Would my children's claim that they have not used AI, "not even once", hold up to scrutiny? I don't know. But I can understand why the survey came back the way it did.
FINALLY someone has published a study on the impacts of AI tutoring. This either suggests that A) AI tutors are not yet as good as human tutors or B) the "2-sigam" results that Sal Khan cited in his Khanmigo announcement don't fully replicate (I've read some critiques of the effect before, although they were only marginally convincing)
The effect size is good, but I would be interested in seeing it on a group of students that _aren't_ as (presumably) behind as African school kids.
Does this mean that the persuader goes in cold against a random target and has to get them to do something they don't want to do purely via verbal arguments?
Or does this mean a full stack influence effort, where the persuader picks a target, gathers information, and builds a plan that includes operating in the real world?
Being "superhuman" at #1 is not so interesting in that a) human level performance at this is very, very low and b) a malign actor trying to do this isn't going to limit themself to verbal arguments.
As for #2, top level human performers are already capable of operating somewhere close to the maximum against undefended targets given sufficient resources. Imagine the CIA bringing its full force to bear on getting your banking passwords. There's already a 100% chance they would succeed.
So, I think the concern here is not so much an AI that performs at a super human level in terms of that kind of manipulation but rather the possibility of scaling performance that matches the best existing manipulation organizations. That seems different than super persuasion, though? And is a vector you guard against in a different way.
There are disagreements about how high a threshold it is to count. To some S-P means better than the best human (e.g. better than certain historical figures or their reputations) when at their full powers, to others it means even more than that.
I see several categories of super persuasion but coercion is a separate topic.
- Systematically taking advantage of our cognitive biases and failure modes, especially individualized and at scale.
- Ability to create unprecedented mass movements - revolutions, cults, cultural shifts, marketing campaigns.
- Ability to bypass the normal cognitive processes - hypnosis, subliminal messaging, and inducing altered states similar to hallucinogens and meditation.
- Direct brain access - signal in the noise (see Snowcrash), implants.
We have historical examples of most of those in humans so its just a matter of degree. I see LLMs with reasoning as ideally suited for the first two categories, multi-modal models can probably induce hypnosis and altered states already, and ASI could find other routes if we're not dead already.
One problem for LLM recommendations is the knowledge cutoff. I've been writing ratings and reviews of all the books I've read for the last three years (~150 books), and Claude gave me great recommendations after I uploaded those as a CSV. However, most if not all of those books were written before the cutoff. For movies, I think there is far more recency bias in what people watch than with books, and my guess is you'd have to scrape and put a lot of data in to context in order to get decent recommendations for anything after the knowledge cutoff?
DougDoug (who uses a ton of LLM stuff in his content) made a pretty good video on AI copyright; not strictly new info but old info in an entertaining format: https://www.youtube.com/watch?v=pt7GtDMTd3k
I genuinely don’t understand why Biden thinks it’s a good idea to propose something to Trump, who tends to reflexively take the opposite stance of whatever his rival’s positions are.
i always read this post to keep in the loop and always leave with a sense that AI isn't coming soon or doing anything meaningful soon. which is funny because i feel as if the vibe of the post is saying the opposite. i just don't see it in the numbers yet anywhere, and AI is just not useful for my own life. It doesn't move the needle for me personally, and I don't see it move the needle in the economy broadly either
There have been lots of companies trying to automate scheduling, including in the 2010s as you mention, including the past owners of x.ai. Not an easy problem to hit the requires accuracy rate.
Clocking in at a chunky 2 hours, here is the podcast episode for this post:
https://open.substack.com/pub/dwatvpodcast/p/ai-99-farewell-to-biden
Just wanted re-state to those wondering why this exists: The value add for this podcast is that I go through and give all the uniquely quoted people their own consistent voices throughout the episode, and describe images for the audio medium. This is more than just a simple TTS pass. Please give it a listen if you are the kind of person who finds audio more accessible.
> Never? No, never. What, never? Well, actually all the time.
> I also find it hard to believe that students are this slow, especially given this is a very low bar - it’s whether you even once asked for ‘help’ at all, in any form.
I get what you are saying. I can also understand why the survey results look the way it does without using the phrase "the kids are lying."
It's happened a few times now. I am helping my kids with their homework and I pull out my phone to either look up something on Gemini/Claude/ChatGPT or help me answer a question. They'd look at me, aghast, and say, "Dad! We're not suppose to use AI!"
I tried to explain there is a difference between using an AI to "Do your work for you and not think" and using an AI to "help you learn." But teenagers have this amazing superpower of shutting you out before you can even complete a sentence. It's 2025. We literally have thinking machines now! Yet, teens still don't listen to their parents.
Would my children's claim that they have not used AI, "not even once", hold up to scrutiny? I don't know. But I can understand why the survey came back the way it did.
FINALLY someone has published a study on the impacts of AI tutoring. This either suggests that A) AI tutors are not yet as good as human tutors or B) the "2-sigam" results that Sal Khan cited in his Khanmigo announcement don't fully replicate (I've read some critiques of the effect before, although they were only marginally convincing)
The effect size is good, but I would be interested in seeing it on a group of students that _aren't_ as (presumably) behind as African school kids.
What is super-persuasion?
Does this mean that the persuader goes in cold against a random target and has to get them to do something they don't want to do purely via verbal arguments?
Or does this mean a full stack influence effort, where the persuader picks a target, gathers information, and builds a plan that includes operating in the real world?
Being "superhuman" at #1 is not so interesting in that a) human level performance at this is very, very low and b) a malign actor trying to do this isn't going to limit themself to verbal arguments.
As for #2, top level human performers are already capable of operating somewhere close to the maximum against undefended targets given sufficient resources. Imagine the CIA bringing its full force to bear on getting your banking passwords. There's already a 100% chance they would succeed.
So, I think the concern here is not so much an AI that performs at a super human level in terms of that kind of manipulation but rather the possibility of scaling performance that matches the best existing manipulation organizations. That seems different than super persuasion, though? And is a vector you guard against in a different way.
There are disagreements about how high a threshold it is to count. To some S-P means better than the best human (e.g. better than certain historical figures or their reputations) when at their full powers, to others it means even more than that.
I see several categories of super persuasion but coercion is a separate topic.
- Systematically taking advantage of our cognitive biases and failure modes, especially individualized and at scale.
- Ability to create unprecedented mass movements - revolutions, cults, cultural shifts, marketing campaigns.
- Ability to bypass the normal cognitive processes - hypnosis, subliminal messaging, and inducing altered states similar to hallucinogens and meditation.
- Direct brain access - signal in the noise (see Snowcrash), implants.
We have historical examples of most of those in humans so its just a matter of degree. I see LLMs with reasoning as ideally suited for the first two categories, multi-modal models can probably induce hypnosis and altered states already, and ASI could find other routes if we're not dead already.
I was considering stepping up for reproducing the medical diagnosis paper, but discovered that the authors actually make the code available on Github (https://github.com/rajpurkarlab/craft-md?tab=readme-ov-file) and they have an updated benchmark here (https://rajpurkarlab.github.io/craft-md-pages/) which has o1-preview at the top, then 4o, and o1-mini in third.
I had Claude whip up some diagrams here:
https://claude.site/artifacts/25449836-0ed4-429d-9773-fee9305833bc
and here:
https://claude.site/artifacts/169dffb5-87d2-4578-a7ac-abad75065477
One problem for LLM recommendations is the knowledge cutoff. I've been writing ratings and reviews of all the books I've read for the last three years (~150 books), and Claude gave me great recommendations after I uploaded those as a CSV. However, most if not all of those books were written before the cutoff. For movies, I think there is far more recency bias in what people watch than with books, and my guess is you'd have to scrape and put a lot of data in to context in order to get decent recommendations for anything after the knowledge cutoff?
DougDoug (who uses a ton of LLM stuff in his content) made a pretty good video on AI copyright; not strictly new info but old info in an entertaining format: https://www.youtube.com/watch?v=pt7GtDMTd3k
Please do the MATH on MATH, and tell me your thoughts me