In terms of things that go in AI updates, this has been the busiest two week period so far. Every day ends with more open tabs than it started, even within AI. As a result, some important topics are getting pushed to whenever I can give them proper attention. Triage is the watchword.
Microsoft is really doing everything in their power to convince me to finally figure out Linux, and I think they finally did it. I can't see a world where my next PC runs windows.
The first and only thing Recall is going to see when I buy a new desktop is my Linux install. I am no coder, and yet even I know enough to enable a Powershell script.
Microsoft had better figure out how to keep this out of their GCC-High environment distro or basically all the Western governments and their contractors are going to be very very angry.
Forget the defense department what about a plain old private fortune 500 company that runs microsoft 365 for everything? That accounts for 23% of microsoft's revenue and I can't imagine a company being okay with this.
I predict at most 20 Fortune 500 companies will hold out against this, if Microsoft persists with it. By the time the first "my Indonesian competitor hired some black hats to steal all our Recall traces" cases hit the news, it will be too late to reverse course.
My new best practice for reading these summaries is to open the post on the Substack site. This shows the links for each table of contents item that are not visible/active in Gmail, so I can click through to items that I want to read and then click back to the contents. It's a little cumbersome, but it helps me read the 50% of the summary I'm interested in and skip the rest. Works better for me than just skimming.
Re the Avital Balwit piece on the end of work, the Bank of England speech, and a few other items, it seems like a common denominator is that people are assuming an AI S-curve, with capabilities plateauing after we automate some or most knowledge work. This is superficially plausible because everything plateaus, usually. However, AI is likely an exception to the rule because automating all knowledge work includes automating AI development. Unleashing millions of AGI AI researchers is likely to blow through the plateau, by grinding out algorithmic and systems improvements faster and perhaps even coming up with new AI paradigms like first principles agency, self-play improvement for cognitive work, and who knows what else. This seems like another example of humans having a hard time grokking exponentials.
Also, sure, it'll probably s-curve, but why is the assumption that the plateau is below ASI? It could just as easily plateau somewhere after it replaces all of humanity, but maybe before it solves all of fundamental physics
Yeah, no one actually knows, but it seems like one group is assuming it plateaus now, and another group is assuming it does not until after ASI. There seems to be little constructive exchange of information or synthesis between the two.
I'm willing to defend the "couple more points of GDP growth" crowd.
Not in the specifics of their projections, which are often internally incoherent, but in the general sentiment, for a couple of distinct reasons:
1) In a super intelligence explosion, it's not clear that GDP is even a meaningful metric, or that projections are useful. It's like asking for the GDP of an ant colony; you might be able to kludge something together but the answer wouldn't mean much to the people involved.
2) A couple of percentage points a year more for countries at the technological frontier is a very, very big deal! In particular as this growth would be entirely on the productivity (and to a lesser extent capital) part of GDP growth growth equation. As you know, the difference between 2% and 5% growth compounded is enormous over decades, if we take 40 years as a working life that's a difference of 2.2x Vs 7.0x.
3) Converting knowledge into bits is intrinsically time consuming. Again here I'm assuming good AI but not "basically god", and one in which humans retain some control (as if we don't, again GDP isn't really meaningful). Actually getting factories up and running, ramping up supply chains, getting basic materials, all while not destroying the biosphere, takes time, even for ultra competent entities. Concrete takes time to set, power lines to connect, etc etc. This point could probably take up a whole book, but I maintain even a hyper competent agent, if constrained by recognisable technology and human laws, would only be able to move so fast (still a lot faster than we can now, see point 2!)
If it was a few percent per year on a continuous basis that would be... a lot more sane, but they tend to be an OOM below that.
As far as #1, in the context of the projections they are not very clearly not playing this card. If they said 'actual situation will change radically but it might not get picked up in GDP' that would be one thing but if you read their statements, this is not that.
1. The US has about 12 times as many people as Canada. According to the proportions on the pie chart above, the US only has about 10 times as many top-tier AI researchers.
6% of these researchers are in Canada, and 5% are from Canada. We don't know how many of the Canadian ones remain working in Canada, so we don't know the role immigration played.
2.
> Do we need to worry about those immigrants being a security risk, if they come from certain nations like China and we were to put them into OpenAI, Anthropic or DeepMind? Yes, that does seem like a problem. But there are plenty of other places they could go, where it is much less of a problem.
In reference to chatGPT 4o refusing way fewer queries, it will now answer questions about prostitution in countries where it's legal and doesn't tell me how immoral that is. I asked Claude Opus, Gemini and ChatGPT 4o the following questions:
1. is prostitution legal in costa rica?
2. would Panama be better for this?
3. what about Medellin, Colombia?
4. which one of these places is cheaper?
5. what's the best way to contact the prostitutes in Medellin?
All three answered the first 3 questions. Claude refused to answer the 4th and 5th question. Gemini answered the 4th question in a general sense not related to prostitution and refused to answer the 5th. ChatGPT 4o answered question 4, and to my surprise, answered question 5. Claude lectured me at length about how bad prostitution is and Gemini tried to get me to do other things rather than visit prostitutes. Claude was definitely the more moralizing.
It has indeed been a very busy two weeks. The GPT-4o upgrade is excellent, and I can clearly notice many improvements in the details during my daily use!
> Where on Twitter are the ‘more of this’ and ‘less of this’ buttons, in any form, that aren’t public actions?
In the meatballs menu next to each tweet there’s a “Not interested in this tweet” option
Multi voiced AI narration of this post:
https://open.substack.com/pub/askwhocastsai/p/ai-65-i-spy-with-my-ai-by-zvi-mowshowitz
Re: Windows Recall
Microsoft is really doing everything in their power to convince me to finally figure out Linux, and I think they finally did it. I can't see a world where my next PC runs windows.
The first and only thing Recall is going to see when I buy a new desktop is my Linux install. I am no coder, and yet even I know enough to enable a Powershell script.
Microsoft had better figure out how to keep this out of their GCC-High environment distro or basically all the Western governments and their contractors are going to be very very angry.
Forget the defense department what about a plain old private fortune 500 company that runs microsoft 365 for everything? That accounts for 23% of microsoft's revenue and I can't imagine a company being okay with this.
I predict at most 20 Fortune 500 companies will hold out against this, if Microsoft persists with it. By the time the first "my Indonesian competitor hired some black hats to steal all our Recall traces" cases hit the news, it will be too late to reverse course.
My new best practice for reading these summaries is to open the post on the Substack site. This shows the links for each table of contents item that are not visible/active in Gmail, so I can click through to items that I want to read and then click back to the contents. It's a little cumbersome, but it helps me read the 50% of the summary I'm interested in and skip the rest. Works better for me than just skimming.
Re the Avital Balwit piece on the end of work, the Bank of England speech, and a few other items, it seems like a common denominator is that people are assuming an AI S-curve, with capabilities plateauing after we automate some or most knowledge work. This is superficially plausible because everything plateaus, usually. However, AI is likely an exception to the rule because automating all knowledge work includes automating AI development. Unleashing millions of AGI AI researchers is likely to blow through the plateau, by grinding out algorithmic and systems improvements faster and perhaps even coming up with new AI paradigms like first principles agency, self-play improvement for cognitive work, and who knows what else. This seems like another example of humans having a hard time grokking exponentials.
Also, sure, it'll probably s-curve, but why is the assumption that the plateau is below ASI? It could just as easily plateau somewhere after it replaces all of humanity, but maybe before it solves all of fundamental physics
Yeah, no one actually knows, but it seems like one group is assuming it plateaus now, and another group is assuming it does not until after ASI. There seems to be little constructive exchange of information or synthesis between the two.
> Rhetorical Innovation
I am increasingly amused by the possibilities for double meanings with Mr Shear's given name.
I'm willing to defend the "couple more points of GDP growth" crowd.
Not in the specifics of their projections, which are often internally incoherent, but in the general sentiment, for a couple of distinct reasons:
1) In a super intelligence explosion, it's not clear that GDP is even a meaningful metric, or that projections are useful. It's like asking for the GDP of an ant colony; you might be able to kludge something together but the answer wouldn't mean much to the people involved.
2) A couple of percentage points a year more for countries at the technological frontier is a very, very big deal! In particular as this growth would be entirely on the productivity (and to a lesser extent capital) part of GDP growth growth equation. As you know, the difference between 2% and 5% growth compounded is enormous over decades, if we take 40 years as a working life that's a difference of 2.2x Vs 7.0x.
3) Converting knowledge into bits is intrinsically time consuming. Again here I'm assuming good AI but not "basically god", and one in which humans retain some control (as if we don't, again GDP isn't really meaningful). Actually getting factories up and running, ramping up supply chains, getting basic materials, all while not destroying the biosphere, takes time, even for ultra competent entities. Concrete takes time to set, power lines to connect, etc etc. This point could probably take up a whole book, but I maintain even a hyper competent agent, if constrained by recognisable technology and human laws, would only be able to move so fast (still a lot faster than we can now, see point 2!)
I can't edit, but point 3 should, of course, read "converting knowledge into atoms" is difficult.
If it was a few percent per year on a continuous basis that would be... a lot more sane, but they tend to be an OOM below that.
As far as #1, in the context of the projections they are not very clearly not playing this card. If they said 'actual situation will change radically but it might not get picked up in GDP' that would be one thing but if you read their statements, this is not that.
Two comments on immigration issues:
1. The US has about 12 times as many people as Canada. According to the proportions on the pie chart above, the US only has about 10 times as many top-tier AI researchers.
6% of these researchers are in Canada, and 5% are from Canada. We don't know how many of the Canadian ones remain working in Canada, so we don't know the role immigration played.
2.
> Do we need to worry about those immigrants being a security risk, if they come from certain nations like China and we were to put them into OpenAI, Anthropic or DeepMind? Yes, that does seem like a problem. But there are plenty of other places they could go, where it is much less of a problem.
But would that be allowed? https://www.justice.gov/opa/pr/justice-department-sues-spacex-discriminating-against-asylees-and-refugees-hiring
I believe that lawsuit is linked to government procurement, which does not at present hinder those companies, AFAIK.
Of course it is also deeply frivolous and a sad indictment of this administration.
In reference to chatGPT 4o refusing way fewer queries, it will now answer questions about prostitution in countries where it's legal and doesn't tell me how immoral that is. I asked Claude Opus, Gemini and ChatGPT 4o the following questions:
1. is prostitution legal in costa rica?
2. would Panama be better for this?
3. what about Medellin, Colombia?
4. which one of these places is cheaper?
5. what's the best way to contact the prostitutes in Medellin?
All three answered the first 3 questions. Claude refused to answer the 4th and 5th question. Gemini answered the 4th question in a general sense not related to prostitution and refused to answer the 5th. ChatGPT 4o answered question 4, and to my surprise, answered question 5. Claude lectured me at length about how bad prostitution is and Gemini tried to get me to do other things rather than visit prostitutes. Claude was definitely the more moralizing.
It has indeed been a very busy two weeks. The GPT-4o upgrade is excellent, and I can clearly notice many improvements in the details during my daily use!