Ben has second thoughts about this whole Iran War thing; congressional candidate Alex Bores joins us to discuss the DoD-Anthropic fight; McDonald's CEO promotes the Big Arch, an exciting new product.
On burgers, I think Josh was conflating the McDLT, from the late 80s with cool/hot separate in styrofoam, and the Arch Deluxe.
Megan for your next LA conference I very much recommend a lunch at the Apple Pan. I think of it as the platonic ideal that In-n-Out is trying to replicate, surprisingly delicious.
Overall I'm in camp Josh, it's decent but wildly overrated.
Love the burger talk, design flaws and all. I request one frivolous topic per week with hilariously analytical discussion.
On the particulars - the Agreed with Megan that in-n-out is great for the price and @Josh Barro I’d like your substantiated case why burgers are unhealthy. I think if you no cheese and no spread and skip the fries/coke they aren’t inherently too bad.
The "Arch Deluxe" failed because it smacks of elitism, whereas the "Big Arch" will be a success because it is the massive burger of the common man.
I worked at McD in the late 1980s. After every shift, my uniform shirt felt steeped in grease from the fryers, which I guess it was. So I get where Josh was coming from.
I live about 3 miles from one of the In-N-Out Burgers in Boise that Ben mentioned. After the podcast was over, I looked at the time and saw it was 9PM, thought, why not, and drove down to get a cheeseburger. This being Wednesday at 9PM, there was a line around the building as always.
Will listen to the episode on the weekend but very glad to hear Ben is realising his support for the war was wrong, and that it only took three week's - took some people year's, even decades to admit they were wrong about Iraq!
I hope Bores and Lasher don’t spilt the vote (they have the same profile) accidentally giving the nomination to the no nothing Schlossberg.
It’s pretty unfortunate that there’s no exciting left wing energy to either Bores or Lasher even though they both support abolishing ICE and Medicare for All. Schlossberg is the only one riding on social media “influencer” energy. I guess the reason the left isn’t excited about any of them is because they are all pro-Israel (makes sense given the district).
I do think it’s risky but imo Lasher should accept Mamdani’s endorsement. Lasher, while having v hawkish views on Israel, endorsed Mamdani right after the primary and said nice reassuring things about him. So unlike left’s blinding hate towards Dan Goldman, they don’t hate Lasher the same way and getting Mamdani to Mike Bloomberg behind him might allow Lasher to consolidate a victory. Borres looks like a long shot since most of the votes are in the Upper West Side and Bores is an Upper East Side guy.
2001 is one of those movies that I'd only ever consider watching on the big screen. Much like slow cinema, it's a sensory experience that I can best appreciate when I'm surrounded by it without distraction.
I am with Josh on this one; a film should not be boring. It can be slow-moving, sure, but that's not the same thing. "The Stepford Wives" moves slowly, but it's interesting throughout. Hell, "Alien" is a hell of a slow burn, but it's great!
I guess I'm just at a loss as to what you mean by "boring" then. Do you mean lack of plot development, or something like that? Because I've watched films where not a lot happens (like 2001) and been entranced by the shot design, the music, the spectacle, the sense that I've never seen anything like this before. That, to me, is not boring.
Conversely, I've seen action movies where I've been bored out of my mind. Just CGI robots mashing against each other endlessly, with no sense of weight or purpose.
Obviously this is mostly a subjective thing. But just saying "no movie should be boring" isn't enough, because people feel that in different ways.
Admittedly, what people find boring is somewhat subjective. However, I was responding to Ben's assertion that a movie can be boring but still good, which I find odd. That's like me saying that I hate the taste of bananas (true), but that I find them delicious anyway.
Note: I realize that I come across as though I am always disagreeing with Ben, but I really don't spend the podcast muttering objections. It just seems that way lately. :-)
The most common argument that I hear from people are not anti-mass surveillance is, “Why should I care if the government is surveilling my online habits. I’m not doing anything wrong”. They don’t seem to understand that the definition of “wrong” is not fixed and can change without warning. Also, many people have things online that that are not wrong in a legal sense but are still things they wouldn’t want shared with the world. That is exactly the kind of stuff that bad actors will seek out to use against people and bad actors can include members of a government or even a whole government.
Long time listener, first time poster, and probably late to this conversation. At the risk of typing into a void, I must take issue with a couple of Alex Bores takes. First, as Josh points out, stringing together data to find unique or near-unique suspects for a crime is a much better way to approach investigation than the more haphazard low-data methods currently in use. Bores says Congress should set limits, but politicians do not have the right incentives because the loudest constituents are at the extremes (ban use, or government free-for-all.) The issue is arbitrary or capricious use of new powerful investigation tools, and it is a problem better handled by constitutional interpretation in the courts. (FWIW the Supreme Court might have something to say about this in its upcoming Chatrie v. United States case. I have written about how the courts should handle these new tools here- https://southerncalifornialawreview.com/2024/07/17/filtered-dragnets-and-the-anti-authoritarian-fourth-amendment/) Also, Megan McArdle is right to press Bores on how people feel about privacy. Bores said that defaults explain why people give away so much data, but that's not quite right because usually the opt-out versus opt-in choices involve situations where the data subject gets the same price or service either way. As soon as they can get a slight discount for opting into tracking, they do. Privacy is very useful to avoid all sorts of abuse in a noisy, low efficiency world, but once tech and expectations adjust to high data, high efficiency, high sorting, we have to think more about the purposes of privacy and preserve those into law.
I don't know what to believe in this Anthropic vs DoD saga. I don't trust Hegseth, but I also don't trust Amodei to be candid about what his company's technology actually does. "AI" can mean so many different things - sometimes it refers to current cutting-edge tech, sometimes it refers to imaginary tech that will perhaps one day exist, and sometimes it refers to tech that's been around for a long time and has just been rebranded as "AI". Some examples from this episode:
- Deanonymization methods have been around forever. It's been almost 30 years since Latanya Sweeney famously sent Massachussetts governor Bill Weld a copy of his own medical record that she had recovered from supposedly anonymized health data.
- If "mass surveillance" means combining lots of data sources and identifying entries across those sources that represent the same person... yeah, that's been around forever too. It's a classic machine learning application, except now everything that we used to call machine learning has to be called AI, so as to insinuate that it's brand new.
- I don't know what Amodei is referring to when he writes about "autonomous weapons", but obviously automation plays a huge role in existing military technology, and has for a very long time.
Amodei seems to want us to imagine that Anthropic is making magical "everything" technology in the form of a pure intelligence that can then be hooked up to whatever application the user desires, and he's stepping in to bravely stop Pete Hegseth from misusing it. (And he's doing it for humanity, of course. Everything these people do is for humanity). After all, there is a dirt-shit-obvious way for Amodei to ensure that the goverment won't use Anthropic's technology for "mass surveillence" or "autonomous weapons": DON'T DEVELOP MASS SURVEILLENCE AND AUTONOMOUS WEAPONS TECHNOLOGY!!!
Apologies for the all-caps, but this shit drives me crazy. How is DoD supposed to use Anthropic's AI to do mass surveillence and autonomous warfare unless Anthropic makes and then hands over AI that does mass surveillence and autonomous warfare?
Because I refuse to believe that Anthropic, while working on some different AI application, just happened come up with technology that does a better job of "mass surveillence" than actual already existing mass surveillence technology, the sort developed by people who specialize in it and have worked in the field for decades. Likewise with "autonomous weapons" - did they whoopsie-daisy stumble into creating something that outperforms the current state-of-the-art from Lockheed Martin and Raytheon? Either they're working on this stuff or they aren't. If it's the former, then they're bullshitting us with this whole "b-b-b-but it's t-t-too p-p-p-p-powerful" sci-fi LARPer morality play they're been putting on for the cameras. And if it's the latter, then they're bullshitting us about the nature of the technology that they've been leasing out to DoD. (I strongly suspect it is the latter)
I know I'm being cynical, but ever since OpenAI released ChatGPT, the big tech CEOs have been playing this game of insinuating that they're making all-purpose electronic brains. And they get away with it because chatbots really do seem like people. But they aren't making brains, or "general intelligence". Chatbot technology is designed specifically for generating text. AI image technology is designed specifically for generating images. Record-linking technology is designed specifically for linking records. Etc. Yes, they share some commonalities under the hood, but they're still made for purpose - there are no magical everything machines. The notion that Hegseth is gonna get take Anthropic's newest version of Claude and then have his lackeys do some hacking and turn it into The Terminator is just silly.
Silly, and apparently exactly what Amodei wants us all to imagine. Because it's s-s-s-s-soooo p-p-p-p-p-p-powerful....
> Because I refuse to believe that Anthropic, while working on some different AI application, just happened come up with technology that does a better job of "mass surveillence" than actual already existing mass surveillence technology
Maybe reconsider that refusal, because that's exactly what happened. LLMs are a general-purpose technology, such that inventing better LLMs can make mass surveillance easier, just like inventing better cameras or better GPUs. Mass surveillance requires collecting giant amounts of data that humans cannot sift through manually, and LLMs are a tool that can be used to help sift through the data automatically.
Can you be more specific? I don't know what you mean when you say LLMs are "general purpose technology". LLMs are text generation technology. Specifically, they are next-token prediction technology. The user inputs a bunch of text, the LLs chops that text up into component tokens, vectorizes them, feeds them through a long series of matrix multiplications that culminate in a probability distribution across possible next tokens, and then the LLM selects a token. Then the LLM takes the user input plus the token it just generated and does the same process again. And then again, and again, until it selects the "stop generating text" token, and then it's done.
It turns out that this method can produce really impressive results. But I don't see how it's "general purpose", except insofar as it will always generate some kind of answer to any question you feed it. Doesn't mean it's *doing* record-linkage, or tracking individuals, or analyzing large datasets, or anything like that. It's always just doing the one thing it does: generating and outputting a series of tokens.
On burgers, I think Josh was conflating the McDLT, from the late 80s with cool/hot separate in styrofoam, and the Arch Deluxe.
Megan for your next LA conference I very much recommend a lunch at the Apple Pan. I think of it as the platonic ideal that In-n-Out is trying to replicate, surprisingly delicious.
Overall I'm in camp Josh, it's decent but wildly overrated.
Right yes, oops
And the McDLT was also the subject of an ad campaign with pre-Seinfeld Jason Alexander: https://youtu.be/nieB0VoRbVU?si=WXOoKqdr6OOjEOBJ
I demand a retraction on the next podcast.
WHAT the cheese was on the cool side?! It's this the Mandella effect of burgers.
Ben, it takes a big man to kinda-sorta, conditionally, tentatively admit that he *might* have made a mistake. ;-)
Love the burger talk, design flaws and all. I request one frivolous topic per week with hilariously analytical discussion.
On the particulars - the Agreed with Megan that in-n-out is great for the price and @Josh Barro I’d like your substantiated case why burgers are unhealthy. I think if you no cheese and no spread and skip the fries/coke they aren’t inherently too bad.
The "Arch Deluxe" failed because it smacks of elitism, whereas the "Big Arch" will be a success because it is the massive burger of the common man.
I worked at McD in the late 1980s. After every shift, my uniform shirt felt steeped in grease from the fryers, which I guess it was. So I get where Josh was coming from.
I live about 3 miles from one of the In-N-Out Burgers in Boise that Ben mentioned. After the podcast was over, I looked at the time and saw it was 9PM, thought, why not, and drove down to get a cheeseburger. This being Wednesday at 9PM, there was a line around the building as always.
I haven’t been there since last summer but I 100 waited like 45 minutes or something on a random weekday in the drive thru haha
Will listen to the episode on the weekend but very glad to hear Ben is realising his support for the war was wrong, and that it only took three week's - took some people year's, even decades to admit they were wrong about Iraq!
I hope Bores and Lasher don’t spilt the vote (they have the same profile) accidentally giving the nomination to the no nothing Schlossberg.
It’s pretty unfortunate that there’s no exciting left wing energy to either Bores or Lasher even though they both support abolishing ICE and Medicare for All. Schlossberg is the only one riding on social media “influencer” energy. I guess the reason the left isn’t excited about any of them is because they are all pro-Israel (makes sense given the district).
I do think it’s risky but imo Lasher should accept Mamdani’s endorsement. Lasher, while having v hawkish views on Israel, endorsed Mamdani right after the primary and said nice reassuring things about him. So unlike left’s blinding hate towards Dan Goldman, they don’t hate Lasher the same way and getting Mamdani to Mike Bloomberg behind him might allow Lasher to consolidate a victory. Borres looks like a long shot since most of the votes are in the Upper West Side and Bores is an Upper East Side guy.
2001 is one of those movies that I'd only ever consider watching on the big screen. Much like slow cinema, it's a sensory experience that I can best appreciate when I'm surrounded by it without distraction.
I am with Josh on this one; a film should not be boring. It can be slow-moving, sure, but that's not the same thing. "The Stepford Wives" moves slowly, but it's interesting throughout. Hell, "Alien" is a hell of a slow burn, but it's great!
I guess I'm just at a loss as to what you mean by "boring" then. Do you mean lack of plot development, or something like that? Because I've watched films where not a lot happens (like 2001) and been entranced by the shot design, the music, the spectacle, the sense that I've never seen anything like this before. That, to me, is not boring.
Conversely, I've seen action movies where I've been bored out of my mind. Just CGI robots mashing against each other endlessly, with no sense of weight or purpose.
Obviously this is mostly a subjective thing. But just saying "no movie should be boring" isn't enough, because people feel that in different ways.
Admittedly, what people find boring is somewhat subjective. However, I was responding to Ben's assertion that a movie can be boring but still good, which I find odd. That's like me saying that I hate the taste of bananas (true), but that I find them delicious anyway.
Note: I realize that I come across as though I am always disagreeing with Ben, but I really don't spend the podcast muttering objections. It just seems that way lately. :-)
The most common argument that I hear from people are not anti-mass surveillance is, “Why should I care if the government is surveilling my online habits. I’m not doing anything wrong”. They don’t seem to understand that the definition of “wrong” is not fixed and can change without warning. Also, many people have things online that that are not wrong in a legal sense but are still things they wouldn’t want shared with the world. That is exactly the kind of stuff that bad actors will seek out to use against people and bad actors can include members of a government or even a whole government.
Long time listener, first time poster, and probably late to this conversation. At the risk of typing into a void, I must take issue with a couple of Alex Bores takes. First, as Josh points out, stringing together data to find unique or near-unique suspects for a crime is a much better way to approach investigation than the more haphazard low-data methods currently in use. Bores says Congress should set limits, but politicians do not have the right incentives because the loudest constituents are at the extremes (ban use, or government free-for-all.) The issue is arbitrary or capricious use of new powerful investigation tools, and it is a problem better handled by constitutional interpretation in the courts. (FWIW the Supreme Court might have something to say about this in its upcoming Chatrie v. United States case. I have written about how the courts should handle these new tools here- https://southerncalifornialawreview.com/2024/07/17/filtered-dragnets-and-the-anti-authoritarian-fourth-amendment/) Also, Megan McArdle is right to press Bores on how people feel about privacy. Bores said that defaults explain why people give away so much data, but that's not quite right because usually the opt-out versus opt-in choices involve situations where the data subject gets the same price or service either way. As soon as they can get a slight discount for opting into tracking, they do. Privacy is very useful to avoid all sorts of abuse in a noisy, low efficiency world, but once tech and expectations adjust to high data, high efficiency, high sorting, we have to think more about the purposes of privacy and preserve those into law.
the lack of love for full metal jacket on this episode was disheartening... :(
I don't know what to believe in this Anthropic vs DoD saga. I don't trust Hegseth, but I also don't trust Amodei to be candid about what his company's technology actually does. "AI" can mean so many different things - sometimes it refers to current cutting-edge tech, sometimes it refers to imaginary tech that will perhaps one day exist, and sometimes it refers to tech that's been around for a long time and has just been rebranded as "AI". Some examples from this episode:
- Deanonymization methods have been around forever. It's been almost 30 years since Latanya Sweeney famously sent Massachussetts governor Bill Weld a copy of his own medical record that she had recovered from supposedly anonymized health data.
- If "mass surveillance" means combining lots of data sources and identifying entries across those sources that represent the same person... yeah, that's been around forever too. It's a classic machine learning application, except now everything that we used to call machine learning has to be called AI, so as to insinuate that it's brand new.
- I don't know what Amodei is referring to when he writes about "autonomous weapons", but obviously automation plays a huge role in existing military technology, and has for a very long time.
Amodei seems to want us to imagine that Anthropic is making magical "everything" technology in the form of a pure intelligence that can then be hooked up to whatever application the user desires, and he's stepping in to bravely stop Pete Hegseth from misusing it. (And he's doing it for humanity, of course. Everything these people do is for humanity). After all, there is a dirt-shit-obvious way for Amodei to ensure that the goverment won't use Anthropic's technology for "mass surveillence" or "autonomous weapons": DON'T DEVELOP MASS SURVEILLENCE AND AUTONOMOUS WEAPONS TECHNOLOGY!!!
Apologies for the all-caps, but this shit drives me crazy. How is DoD supposed to use Anthropic's AI to do mass surveillence and autonomous warfare unless Anthropic makes and then hands over AI that does mass surveillence and autonomous warfare?
Because I refuse to believe that Anthropic, while working on some different AI application, just happened come up with technology that does a better job of "mass surveillence" than actual already existing mass surveillence technology, the sort developed by people who specialize in it and have worked in the field for decades. Likewise with "autonomous weapons" - did they whoopsie-daisy stumble into creating something that outperforms the current state-of-the-art from Lockheed Martin and Raytheon? Either they're working on this stuff or they aren't. If it's the former, then they're bullshitting us with this whole "b-b-b-but it's t-t-too p-p-p-p-powerful" sci-fi LARPer morality play they're been putting on for the cameras. And if it's the latter, then they're bullshitting us about the nature of the technology that they've been leasing out to DoD. (I strongly suspect it is the latter)
I know I'm being cynical, but ever since OpenAI released ChatGPT, the big tech CEOs have been playing this game of insinuating that they're making all-purpose electronic brains. And they get away with it because chatbots really do seem like people. But they aren't making brains, or "general intelligence". Chatbot technology is designed specifically for generating text. AI image technology is designed specifically for generating images. Record-linking technology is designed specifically for linking records. Etc. Yes, they share some commonalities under the hood, but they're still made for purpose - there are no magical everything machines. The notion that Hegseth is gonna get take Anthropic's newest version of Claude and then have his lackeys do some hacking and turn it into The Terminator is just silly.
Silly, and apparently exactly what Amodei wants us all to imagine. Because it's s-s-s-s-soooo p-p-p-p-p-p-powerful....
> Because I refuse to believe that Anthropic, while working on some different AI application, just happened come up with technology that does a better job of "mass surveillence" than actual already existing mass surveillence technology
Maybe reconsider that refusal, because that's exactly what happened. LLMs are a general-purpose technology, such that inventing better LLMs can make mass surveillance easier, just like inventing better cameras or better GPUs. Mass surveillance requires collecting giant amounts of data that humans cannot sift through manually, and LLMs are a tool that can be used to help sift through the data automatically.
Can you be more specific? I don't know what you mean when you say LLMs are "general purpose technology". LLMs are text generation technology. Specifically, they are next-token prediction technology. The user inputs a bunch of text, the LLs chops that text up into component tokens, vectorizes them, feeds them through a long series of matrix multiplications that culminate in a probability distribution across possible next tokens, and then the LLM selects a token. Then the LLM takes the user input plus the token it just generated and does the same process again. And then again, and again, until it selects the "stop generating text" token, and then it's done.
It turns out that this method can produce really impressive results. But I don't see how it's "general purpose", except insofar as it will always generate some kind of answer to any question you feed it. Doesn't mean it's *doing* record-linkage, or tracking individuals, or analyzing large datasets, or anything like that. It's always just doing the one thing it does: generating and outputting a series of tokens.