I'm bringing a copy of the PT Cube and the RoboRosewater Masters Cube this weekend, lmk if you'd like to draft either (though you'll bring the PT Top 8s average up quite a bit 😅). Either way, I'd appreciate you signing one of the Solution cards that are in the former.
Re language learning I've been having a lot of success using the khan academy gpt on chatgpt to quiz me on mandarin. Being able to get corrections from it and explanations for grammatical points, identifying where I've made errors, etc is very useful. Where most language leering programs just give a binary correct/incorrect output
Caveat that this is getting information on one of the most popular global languages in the other, and generally dealing with pretty basic grammar and vocab. So may not extend to more obscure languages and more advanced levels
Possibly of interest, from Google and academic researchers: "LLMs achieve adult human performance on higher-order theory of mind tasks"
"This paper examines the extent to which large language models (LLMs) have developed higher-order theory of mind (ToM); the human ability to reason about multiple mental and emotional states in a recursive manner (e.g. I think that you believe that she knows). This paper builds on prior work by introducing a handwritten test suite – Multi-Order Theory of Mind Q&A – and using it to compare the performance of five LLMs to a newly gathered adult human benchmark. We find that GPT-4 and Flan-PaLM reach adult-level and near adult-level performance on ToM tasks overall, and that GPT-4 exceeds adult performance on 6th order inferences. Our results suggest that there is an interplay between model size and finetuning for the realisation of ToM abilities, and that the best-performing LLMs have developed a generalised capacity for ToM. Given the role that higher-order ToM plays in a wide range of cooperative and competitive human behaviours, these findings have significant implications for user-facing LLM applications."
Enjoyed your breakdown but I have to disagree with your point on Lina Khan. If you read the article, she doesn't actually say that scraping someone else's data violates antitrust laws, only that it's an "unfair method of competition", which is essentially a direct quote from the FTC Act. This seems hardly controversial--if OpenAI is scraping news articles to train models without consent from news sites which then directly contribute to less traffic to those news sites, then that's pretty unfair.
It seems like you might be focused too much on the use of "antitrust laws" in the article which seems more like a poor editorial choice / misunderstanding by the reporter (she should have just said FTC Act for more precision).
I mean Khan totally does keep doing that, so it's understandable to get confused that instead Khan is simply saying the FTC can use OTHER 'authority' to clamp down on anything they don't like.
Either the scraping should be legal or illegal and that really should not depend on the FTC deciding on 'fairness.'
I think it's fair to say that the FTC Act might grant the FTC overly broad regulatory discretion, but as the law written right now it's technically within her jurisdiction. It's only becoming an issue now because previous administrations didn't really bother interpreting the law that way.
I think that might be a very innovative and unprecedented use of that phrase.
Lina Khan is one of the reasons I am pretty sanguine about the upcoming election, we might get Trump who might for various reasons actually do all the bad things he somehow didn't get around to last time, but that risk is to be weighed against definitely getting rid of Lina Khan and hopefully getting rid of the CFTC commissioners and the NLRB.
Are you really expecting that someone with a username that hails from a call to genocide largely over money and power by Cato that resulted in not simply the obliteration of the city of Carthage but all remaining citizens sold into slavery to be concerned about the rights of marginalized people when there's money to be made by further stripping workers' rights and limiting the power of big business? After all the end of the Punic Wars only further empowered those in Rome who wanted conquest.
At the very least, the phrase delenda est tells me that someone is either historically illiterate, or they are in favor of the most powerful being stronger. Neither of which I consider to be good things.
"Once again, it feels like if someone runs over rule of law and imposes tons of arbitrary rules, the internet stops to ask if it might plausibly stop us from dying. If not, then they get a free pass."
Regarding AGI/ASI timelines, a few data points have made me think it's not very close: LeCun's recent comments about needing a different model architecture, studies on LLM planning (https://x.com/rao2z/status/1795595801177260311), and work on counterfactual tasks (https://aiguide.substack.com/p/evaluating-large-language-models). I still think the current models are very useful, and there's still a fuzzy line between retrieval and reasoning in practice, but I think we're going to need some new paradigms in addition to scaling to get real agents. It'll be interesting to see to what extent you can throw massive amounts of money and human capital at the problem to force faster timelines on model architectures. I suspect big idea innovation is hard to accelerate.
I'm bringing a copy of the PT Cube and the RoboRosewater Masters Cube this weekend, lmk if you'd like to draft either (though you'll bring the PT Top 8s average up quite a bit 😅). Either way, I'd appreciate you signing one of the Solution cards that are in the former.
Happy to sign a card. Whether or not I play depends how things are going otherwise.
Re language learning I've been having a lot of success using the khan academy gpt on chatgpt to quiz me on mandarin. Being able to get corrections from it and explanations for grammatical points, identifying where I've made errors, etc is very useful. Where most language leering programs just give a binary correct/incorrect output
Caveat that this is getting information on one of the most popular global languages in the other, and generally dealing with pretty basic grammar and vocab. So may not extend to more obscure languages and more advanced levels
Yann provides humor.
The "run off a cliff without looking down" image is there twice.
The MLP stuff is also probably false https://x.com/TobbyMLP/status/1795951341875155301
Make sure to tag your 10^26 FLOPS model as emotional support if you want to bring it to less online
Whew this one was a bit of a doozy, multi voiced AI narration where everyone gets their own voice:
https://open.substack.com/pub/askwhocastsai/p/ai-66-oh-to-be-less-online-by-zvi
A grain of salt is a small rock
Fyi, your post says Sundar is CEO of Microsoft. Should say Alphabet/Google.
Yep, my mistake, good catch, fixed.
The Hasbro thing seems to be untrue, according to follow-ups to the original tweets and messages. May want to add an edit to that section.
Huh. At the time I checked the threads and didn't see anyone contradicting it.
Confirmed that the story was fake (details here are kind of wild), put in an edit but left the segment up for posterity.
Hi Zvi--
Possibly of interest, from Google and academic researchers: "LLMs achieve adult human performance on higher-order theory of mind tasks"
"This paper examines the extent to which large language models (LLMs) have developed higher-order theory of mind (ToM); the human ability to reason about multiple mental and emotional states in a recursive manner (e.g. I think that you believe that she knows). This paper builds on prior work by introducing a handwritten test suite – Multi-Order Theory of Mind Q&A – and using it to compare the performance of five LLMs to a newly gathered adult human benchmark. We find that GPT-4 and Flan-PaLM reach adult-level and near adult-level performance on ToM tasks overall, and that GPT-4 exceeds adult performance on 6th order inferences. Our results suggest that there is an interplay between model size and finetuning for the realisation of ToM abilities, and that the best-performing LLMs have developed a generalised capacity for ToM. Given the role that higher-order ToM plays in a wide range of cooperative and competitive human behaviours, these findings have significant implications for user-facing LLM applications."
https://arxiv.org/pdf/2405.18870
Enjoyed your breakdown but I have to disagree with your point on Lina Khan. If you read the article, she doesn't actually say that scraping someone else's data violates antitrust laws, only that it's an "unfair method of competition", which is essentially a direct quote from the FTC Act. This seems hardly controversial--if OpenAI is scraping news articles to train models without consent from news sites which then directly contribute to less traffic to those news sites, then that's pretty unfair.
It seems like you might be focused too much on the use of "antitrust laws" in the article which seems more like a poor editorial choice / misunderstanding by the reporter (she should have just said FTC Act for more precision).
I mean Khan totally does keep doing that, so it's understandable to get confused that instead Khan is simply saying the FTC can use OTHER 'authority' to clamp down on anything they don't like.
Either the scraping should be legal or illegal and that really should not depend on the FTC deciding on 'fairness.'
I think it's fair to say that the FTC Act might grant the FTC overly broad regulatory discretion, but as the law written right now it's technically within her jurisdiction. It's only becoming an issue now because previous administrations didn't really bother interpreting the law that way.
I think that might be a very innovative and unprecedented use of that phrase.
Lina Khan is one of the reasons I am pretty sanguine about the upcoming election, we might get Trump who might for various reasons actually do all the bad things he somehow didn't get around to last time, but that risk is to be weighed against definitely getting rid of Lina Khan and hopefully getting rid of the CFTC commissioners and the NLRB.
I think my female, black and gay relatives would disagree with your assertion that those are equivalent.
Are you really expecting that someone with a username that hails from a call to genocide largely over money and power by Cato that resulted in not simply the obliteration of the city of Carthage but all remaining citizens sold into slavery to be concerned about the rights of marginalized people when there's money to be made by further stripping workers' rights and limiting the power of big business? After all the end of the Punic Wars only further empowered those in Rome who wanted conquest.
At the very least, the phrase delenda est tells me that someone is either historically illiterate, or they are in favor of the most powerful being stronger. Neither of which I consider to be good things.
Fair points all.
"Once again, it feels like if someone runs over rule of law and imposes tons of arbitrary rules, the internet stops to ask if it might plausibly stop us from dying. If not, then they get a free pass."
Oh no
Regarding AGI/ASI timelines, a few data points have made me think it's not very close: LeCun's recent comments about needing a different model architecture, studies on LLM planning (https://x.com/rao2z/status/1795595801177260311), and work on counterfactual tasks (https://aiguide.substack.com/p/evaluating-large-language-models). I still think the current models are very useful, and there's still a fuzzy line between retrieval and reasoning in practice, but I think we're going to need some new paradigms in addition to scaling to get real agents. It'll be interesting to see to what extent you can throw massive amounts of money and human capital at the problem to force faster timelines on model architectures. I suspect big idea innovation is hard to accelerate.
Re: opting out to Facebook training AI on your data:
Fill in the form like a [Dangerous Professional](https://www.kalzumeus.com/2017/09/09/identity-theft-credit-reports/), as Patrick McKenzie would put it.
I've written examples for UK / EU, and California, on [LessWrong](https://www.lesswrong.com/posts/vSPdRg8siXCh6mLvt/ai-66-oh-to-be-less-online?commentId=hv5A4Jjxb2hX2PMau).