How does one actually overcome Claude's limits on the Pro plan? It downgrade for around 5 hours to 3.5 haiku which is worse than 4o and in no way a substitute for big conversations. Use API instead of subscription?
Zvi often mentions it as his base model, but you can reach a 5-hour limit pretty quickly with an extensive use.
I don't use enough queries to run into the limit often (it's happened but it's rare). But yeah, if you are running out on the regular start using the API.
There are probably *some* limits on the API, but I've never encountered them; unless you're doing something at scale (eg running parallel queries in a loop) they're effectively unlimited. If you want it to still feel like conversation (plus bonuses like being able to fork off multiple threads) you can use something like https://msty.app.
I've been using the API through OpenRouter, but I'm considering a subscription since I have ongoing conversations throughout the day (for addiction tracking) that I need to access from both my computer and phone on the go. The issue with interfaces like OpenRouter Chatroom or Msty is that conversations are stored locally. Is there a service that would let me use my OpenRouter API key while storing conversations on their servers (or maybe I can rent a server and connect it to some interface accessible both from Mac and mobile either via a browser or app)? This way I could access my chat history from any device at any time (whether through a browser or app). Any recommendations?
Thanks for the fantastic AI posts as always. I want to hope that the tariff policy was not in fact spat out by a language model, but rather that the AI finds the same behavior because it's the most direct naive apparent solution to the problem that it was prompted with, which is tariffing in a way that could seemingly equalize the trade deficit (though this naive solution completely ignores elasticity). Truly a logical goal, since what American wants to be a blue collar worker for $20/hr when they could work in a sweatshop for 50 cents an hour instead, am I right? The thing that makes the most frustrated is that several of the key idiots, Trump included, have economics degrees. It really makes me wonder what they were doing instead of studying.
It's not about the economy, stupid. What people don't understand is that trump is a nationalist. The whole point is deglobalization. Even if these tarrifs hurt the US economy, as long as they help to deglobalize the world I think his administration is all for it. They actually really don't care. Nationalism is more important for them than the economy. This is also why the border is shut down.
Remember, we are going to 'build a wall' and mexico is going to pay for it?
Well, they are building an economic wall with the rest of the planet right now.
Nationalism is something that is foreign to our generation as we grew up in the liberal unipolar world order. But that world ended with the rise of China.
Here is an obvious prediction. Nationalism will continue to rise everywhere on the planet. Also in the USA, a future leftwing presidency will keep most of what trump is doing to trade and everything else in place. Biden kept all the tarrifs from trump 1.0 in place and only increased them (esp vis a vis China) and it will be true in the future as well.
Hey, only my wife calls me stupid. I guess maybe that was an attempt at referencing that Bill Clinton advisor's political strategy.
Anyway, these tariffs are more like passing economic sanctions on ourselves. Is that actually nationalism? Perhaps attempted nationalism, in the sense that self injury could be attempted self-medication.
While I disagree with much of what you're implying, your comment about tariffs being historically very sticky is actually true but suggests that it was a very bad idea. See, for example: Smoot-Hawley, the British Corn Laws, the Dairy tariffs in Canada. We still have a stupid chicken tax between the United States and Germany from the 1960s because of a trade dispute that was never resolved. The problem is that when you do this sort of thing, both parties don't really benefit.
I'm actually in Robin Hanson's camp in all of this -in-kind de-escalatory reciprocal (actually reciprocal) tariffs could be reasonable in certain situations, and President Trump could have done that and I'd be commending him for it if he did or if he course corrected.
However, you also have to be a little targeted. We are a rich country because we manufacture high on the value chain - airplanes, medical devices, software. The massive wealth we take for granted as Americans was built on moving up that value chain. Intentionally moving down it would be moving backwards in income and standard of living. It really doesn't make any logical sense to tariff Cambodia because we don't want to live in mud huts and work in sweatshops for pennies an hour. We benefit from letting them make us cheap Nike shoes and focusing on other things. If the aim of nationalism is doing what is good for America, free trade is a win for almost everyone in the country.
I think the goals are far more to have a complete industrial chain for the very-long term the same way China does it to deal with the reality of a multipolar world and not be dependent on China for /everything/ (which is currently the case). You do need high tarrif barriers (the same way emerging economies do btw. china included) to pull this off + targeted industrial policy. As when you build industries from zero you can not compete in the globalized economy, so you put up high tarrif barriers while you build that industry. This is standard emerging market asian tiger type plays.
Think of the USA as an emerging economy and the policies make sense. Which it is by the way, like the energy grid is going to collapse over the next 50 years or so because everyone that can keep it afloat is retiring. This is true for the USA's entire industrial chain, including Boeing, GM, etc. Only Elon Musk can still produce things.
It is literally, all collapsing right now. Just go chat with people in industrial manufacturing in the USA. everyone is retiring, nobody has the skills because nobody is trained for it, and China can do everything 10x cheaper anyway.
There are deep structural problems the USA has due to having offshored all of its manufacturing.
But I agree that the US its political system will not allow it as the next person in charge will just undo it all. So you know, if you have children there is no future either way for america and the west in a non ASI world. In an ASI world there is probably also no future.
I think there is an intelligible argument that China's manufacturing is a problem and so China should be tariffed, but the intelligent way to do tthat would be to coordinate with our allies to all contain their manufacturing together with joint tariffs, not to single handedly tariff China and everyone else. The way the tariffs are being done to everyone chases all strategic partners that could contain China right into their arms.
And this isn't hypothetical; South Korea, Japan and China have strengthened their trade deals and announced that they will become closer trade partners and distance themselves from the USA. And these countries all hate each other! The picture of their economic ministers all holding hands on the 3rd is quite a sight.
yes i'm not even sure if tarrifs are the right way to go about this. But the blame should squarely be on all the neocons and unipolar era liberals like clinton, obama, bush and biden. rather than trump for trying to do something about it.
The USA helped China become rich, and now China wants to dominate the USA. greatest strategic failure in the history of mankind.
My general comment that nationalism will continue to rise everywhere on the planet stands however. I suspect a US-led block and a China-led block. And these two will compete for security, as happened in the first cold war.
The formation of the blocks is happening right now as we speak as the USA starts to renegotiate trade relations around the globe.
I actually don't know if tarrifs are the right way to go about this. But I understand where they are coming from.
University-wide Claude availability is going to be a fantastic initiative to help combat any sort of two tier education fallout caused by differences in LLM access. If everyone is working in the same environment then a lot more focus can be directed at learning essential rather than incidental skills.
(Although Learning mode feels like its "just" a tweaked system prompt in projects.)
re: anton leicht's position: so best one can hope for is some kind of AI Chernobyl (bad example, I know, nuclear energy good etc) that is disastrous but limited and makes people wake up?
"We are still putting a huge tax on GPUs. Are we trying to lose the future?"
I'm confused about your general position on AI. If you think AI doom is possible, shouldn't anything that slows down AI be good? Is it that the US should stay ahead of China? If so, why? (Obviously from the perspective of the American government AI development is good so they should not do tarrifs, but you seem to be arguing that you also personally think tarrifs that slow down American AI are bad, which is what I don't understand).
Do you think the US is more likely to make aligned AGI? How certain of this are you? Does it not seem more likely that any international competition that speeds up AI development will outweigh this and thus increase pdoom?
You have thought about this a lot more than I have but I haven't been able to piece together your opinion on this. I have read a lot of your writing (you're a great writer) but not all (you're a prolific writer) so it's possible I've missed an article where you more clearly lay out your position.
I don't think most people, especially most people in government, make a habit of checking their results with multiple AIs. Most people have one they use regularly, if they use AI at all. Because of that, I don't think it's drastically stronger evidence of AI use that multiple AI give the equation actually used, relative to if just one AI gave that result. I think the more likely explanation is that all the LLMs and government policy folks are drawing from a common source, like some 1910 economics book or something that said this is how to do tariffs.
There has been lot of commentary about LLM performance on the USAMO, saying it proves something that the LLMs all got a zero, but I think people are (understandably) getting the AIME level competitions and USAMO level competitions confused.
The AIME is hard enough that a smart high schooler can score a zero. The USAMO is much harder, hard enough that someone who does very well on the AIME can score a zero. The IMO is roughly USAMO level, and Alphaproof did well on it, but no LLM has ever done well on this level.
So this is all a sign that Gemini 2.5 is impressive, rather than that the LLMs were overrated before because they did badly on the USAMO while crushing the AIME.
A paper book from the 1980s called "Feeling Good" is already a pretty good therapist. As a paper book, "Feeling Good" doesn't have any theory of mind at all, and can't update its text based on the reader's particular situation. But it can still successfully treat depression and anxiety symptoms for many people.
So to me, "Decent AI therapist" has for many years seemed like an obvious easy AI product (probably doable adequately even with a 3.5-level model, except maybe requiring big context though). Getting a well-designed and reliable version of this product out there seems like a good thing.
On "beating a Pokémon game zero shot" we *really* need a definition of what the AI here is. TAS' are arguably a form of "rules based AI" that have been able to beat Pokémon for at least 20 years. Hell, in the next 18 months I'd expect a basic agent model (not even a good one!) to hack a solution to this by downloading the solution from this video: https://www.youtube.com/watch?v=jNMWkD5VsZ8 and just playing it (great video for those who enjoy game analysis YouTube, which I acknowledge isn't everyone's thing).
If it is limited to "feed a series of screenshots into a ChatGPT equivalent and ask for 'what to do next' repeatedly, with no ability to search the internet" I decrease my probability from 100% to more like 20%. The critical factor here will be navigating context window length, which I still think is pretty doable.
How does one actually overcome Claude's limits on the Pro plan? It downgrade for around 5 hours to 3.5 haiku which is worse than 4o and in no way a substitute for big conversations. Use API instead of subscription?
Zvi often mentions it as his base model, but you can reach a 5-hour limit pretty quickly with an extensive use.
I don't use enough queries to run into the limit often (it's happened but it's rare). But yeah, if you are running out on the regular start using the API.
There are probably *some* limits on the API, but I've never encountered them; unless you're doing something at scale (eg running parallel queries in a loop) they're effectively unlimited. If you want it to still feel like conversation (plus bonuses like being able to fork off multiple threads) you can use something like https://msty.app.
I've been using the API through OpenRouter, but I'm considering a subscription since I have ongoing conversations throughout the day (for addiction tracking) that I need to access from both my computer and phone on the go. The issue with interfaces like OpenRouter Chatroom or Msty is that conversations are stored locally. Is there a service that would let me use my OpenRouter API key while storing conversations on their servers (or maybe I can rent a server and connect it to some interface accessible both from Mac and mobile either via a browser or app)? This way I could access my chat history from any device at any time (whether through a browser or app). Any recommendations?
Thanks for the fantastic AI posts as always. I want to hope that the tariff policy was not in fact spat out by a language model, but rather that the AI finds the same behavior because it's the most direct naive apparent solution to the problem that it was prompted with, which is tariffing in a way that could seemingly equalize the trade deficit (though this naive solution completely ignores elasticity). Truly a logical goal, since what American wants to be a blue collar worker for $20/hr when they could work in a sweatshop for 50 cents an hour instead, am I right? The thing that makes the most frustrated is that several of the key idiots, Trump included, have economics degrees. It really makes me wonder what they were doing instead of studying.
It's not about the economy, stupid. What people don't understand is that trump is a nationalist. The whole point is deglobalization. Even if these tarrifs hurt the US economy, as long as they help to deglobalize the world I think his administration is all for it. They actually really don't care. Nationalism is more important for them than the economy. This is also why the border is shut down.
Remember, we are going to 'build a wall' and mexico is going to pay for it?
Well, they are building an economic wall with the rest of the planet right now.
Nationalism is something that is foreign to our generation as we grew up in the liberal unipolar world order. But that world ended with the rise of China.
Here is an obvious prediction. Nationalism will continue to rise everywhere on the planet. Also in the USA, a future leftwing presidency will keep most of what trump is doing to trade and everything else in place. Biden kept all the tarrifs from trump 1.0 in place and only increased them (esp vis a vis China) and it will be true in the future as well.
Hey, only my wife calls me stupid. I guess maybe that was an attempt at referencing that Bill Clinton advisor's political strategy.
Anyway, these tariffs are more like passing economic sanctions on ourselves. Is that actually nationalism? Perhaps attempted nationalism, in the sense that self injury could be attempted self-medication.
While I disagree with much of what you're implying, your comment about tariffs being historically very sticky is actually true but suggests that it was a very bad idea. See, for example: Smoot-Hawley, the British Corn Laws, the Dairy tariffs in Canada. We still have a stupid chicken tax between the United States and Germany from the 1960s because of a trade dispute that was never resolved. The problem is that when you do this sort of thing, both parties don't really benefit.
I'm actually in Robin Hanson's camp in all of this -in-kind de-escalatory reciprocal (actually reciprocal) tariffs could be reasonable in certain situations, and President Trump could have done that and I'd be commending him for it if he did or if he course corrected.
However, you also have to be a little targeted. We are a rich country because we manufacture high on the value chain - airplanes, medical devices, software. The massive wealth we take for granted as Americans was built on moving up that value chain. Intentionally moving down it would be moving backwards in income and standard of living. It really doesn't make any logical sense to tariff Cambodia because we don't want to live in mud huts and work in sweatshops for pennies an hour. We benefit from letting them make us cheap Nike shoes and focusing on other things. If the aim of nationalism is doing what is good for America, free trade is a win for almost everyone in the country.
I think the goals are far more to have a complete industrial chain for the very-long term the same way China does it to deal with the reality of a multipolar world and not be dependent on China for /everything/ (which is currently the case). You do need high tarrif barriers (the same way emerging economies do btw. china included) to pull this off + targeted industrial policy. As when you build industries from zero you can not compete in the globalized economy, so you put up high tarrif barriers while you build that industry. This is standard emerging market asian tiger type plays.
Think of the USA as an emerging economy and the policies make sense. Which it is by the way, like the energy grid is going to collapse over the next 50 years or so because everyone that can keep it afloat is retiring. This is true for the USA's entire industrial chain, including Boeing, GM, etc. Only Elon Musk can still produce things.
It is literally, all collapsing right now. Just go chat with people in industrial manufacturing in the USA. everyone is retiring, nobody has the skills because nobody is trained for it, and China can do everything 10x cheaper anyway.
There are deep structural problems the USA has due to having offshored all of its manufacturing.
But I agree that the US its political system will not allow it as the next person in charge will just undo it all. So you know, if you have children there is no future either way for america and the west in a non ASI world. In an ASI world there is probably also no future.
I think there is an intelligible argument that China's manufacturing is a problem and so China should be tariffed, but the intelligent way to do tthat would be to coordinate with our allies to all contain their manufacturing together with joint tariffs, not to single handedly tariff China and everyone else. The way the tariffs are being done to everyone chases all strategic partners that could contain China right into their arms.
And this isn't hypothetical; South Korea, Japan and China have strengthened their trade deals and announced that they will become closer trade partners and distance themselves from the USA. And these countries all hate each other! The picture of their economic ministers all holding hands on the 3rd is quite a sight.
yes i'm not even sure if tarrifs are the right way to go about this. But the blame should squarely be on all the neocons and unipolar era liberals like clinton, obama, bush and biden. rather than trump for trying to do something about it.
The USA helped China become rich, and now China wants to dominate the USA. greatest strategic failure in the history of mankind.
My general comment that nationalism will continue to rise everywhere on the planet stands however. I suspect a US-led block and a China-led block. And these two will compete for security, as happened in the first cold war.
The formation of the blocks is happening right now as we speak as the USA starts to renegotiate trade relations around the globe.
I actually don't know if tarrifs are the right way to go about this. But I understand where they are coming from.
I think it’s worth eventually speaking more on the LSE / Claude for Education announcements.
https://www.anthropic.com/news/introducing-claude-for-education
University-wide Claude availability is going to be a fantastic initiative to help combat any sort of two tier education fallout caused by differences in LLM access. If everyone is working in the same environment then a lot more focus can be directed at learning essential rather than incidental skills.
(Although Learning mode feels like its "just" a tweaked system prompt in projects.)
re: anton leicht's position: so best one can hope for is some kind of AI Chernobyl (bad example, I know, nuclear energy good etc) that is disastrous but limited and makes people wake up?
yeah, afaict: https://x.com/QiaochuYuan/status/1884668625942175983
odd that it worked out like this with all of these EA's at the top of these organisations.
"We are still putting a huge tax on GPUs. Are we trying to lose the future?"
I'm confused about your general position on AI. If you think AI doom is possible, shouldn't anything that slows down AI be good? Is it that the US should stay ahead of China? If so, why? (Obviously from the perspective of the American government AI development is good so they should not do tarrifs, but you seem to be arguing that you also personally think tarrifs that slow down American AI are bad, which is what I don't understand).
Do you think the US is more likely to make aligned AGI? How certain of this are you? Does it not seem more likely that any international competition that speeds up AI development will outweigh this and thus increase pdoom?
You have thought about this a lot more than I have but I haven't been able to piece together your opinion on this. I have read a lot of your writing (you're a great writer) but not all (you're a prolific writer) so it's possible I've missed an article where you more clearly lay out your position.
I thought at first the title was about AI and wondered if someone had taken over the world while I was sleeping. Well, yes and no
Re: Multiple AI agreeing on tariffs
I don't think most people, especially most people in government, make a habit of checking their results with multiple AIs. Most people have one they use regularly, if they use AI at all. Because of that, I don't think it's drastically stronger evidence of AI use that multiple AI give the equation actually used, relative to if just one AI gave that result. I think the more likely explanation is that all the LLMs and government policy folks are drawing from a common source, like some 1910 economics book or something that said this is how to do tariffs.
There has been lot of commentary about LLM performance on the USAMO, saying it proves something that the LLMs all got a zero, but I think people are (understandably) getting the AIME level competitions and USAMO level competitions confused.
The AIME is hard enough that a smart high schooler can score a zero. The USAMO is much harder, hard enough that someone who does very well on the AIME can score a zero. The IMO is roughly USAMO level, and Alphaproof did well on it, but no LLM has ever done well on this level.
So this is all a sign that Gemini 2.5 is impressive, rather than that the LLMs were overrated before because they did badly on the USAMO while crushing the AIME.
Podcast episode for this post:
https://open.substack.com/pub/dwatvpodcast/p/ai-110-of-course-you-know?r=67y1h&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
> The new ‘Liberation Day’ tariffs are suicidal insanity.
I guess the silver lining is that replacing human leaders with AGI seems increasingly less like a horrifying prospect.
I kid of course, but FFS
"they must see robots kill ppl in their hometown or AI isn't real"
Build a MMO hometown. Let the AIs kill people in it. Profit?
If I may throw in a request - Would love to know what you think about the AGI timelines of the epoch guys.
A paper book from the 1980s called "Feeling Good" is already a pretty good therapist. As a paper book, "Feeling Good" doesn't have any theory of mind at all, and can't update its text based on the reader's particular situation. But it can still successfully treat depression and anxiety symptoms for many people.
So to me, "Decent AI therapist" has for many years seemed like an obvious easy AI product (probably doable adequately even with a 3.5-level model, except maybe requiring big context though). Getting a well-designed and reliable version of this product out there seems like a good thing.
Typo: TxGamma -> TxGemma
Couldn't find anything about TxGamma by googling at first XD (twitter is blocked on my end).
On "beating a Pokémon game zero shot" we *really* need a definition of what the AI here is. TAS' are arguably a form of "rules based AI" that have been able to beat Pokémon for at least 20 years. Hell, in the next 18 months I'd expect a basic agent model (not even a good one!) to hack a solution to this by downloading the solution from this video: https://www.youtube.com/watch?v=jNMWkD5VsZ8 and just playing it (great video for those who enjoy game analysis YouTube, which I acknowledge isn't everyone's thing).
If it is limited to "feed a series of screenshots into a ChatGPT equivalent and ask for 'what to do next' repeatedly, with no ability to search the internet" I decrease my probability from 100% to more like 20%. The critical factor here will be navigating context window length, which I still think is pretty doable.
> Anthropic releases more information with information from the Anthropic Economic Index
Side gripe: Is anyone else annoyed by the use of the noun “index” in the sense of “report” or “dataset”?