Latest / Stuff Keeps Happening / An AI Cafe

Transcript

An AI Cafe · Stuff Keeps Happening

• 0:00
  Hello. I played the hand that I was dealt. But
• 0:04
  stuff still happened. So what's up, dear listener?
• 0:08
  I apologize. This is going to be an AI heavy
• 0:10
  one. I'm not a big AI boy. I don't sit here and
• 0:13
  think about it all the time. It's just what came
• 0:15
  across my feeds and also because I want to talk
• 0:17
  about it. Would you want to work at a cafe run
• 0:21
  by an AI or any business run by an AI? Did you
• 0:24
  answer no? Correct. Andin Labs has a mission.
• 0:28
  Take AI and cram it into the real world. And
• 0:31
  see what happens. There's this company called
• 0:33
  Andin Labs in San Francisco, and they take AI
• 0:35
  and they try to make it run businesses. They've
• 0:38
  previously done it with a vending machine, being
• 0:40
  like, hey, run this one vending machine. and
• 0:43
  make a profit. It did not make a profit. It lost
• 0:45
  all of its money. So now they're doing another
• 0:47
  one. The vending machine was done, I believe,
• 0:49
  with Claude, and now they're using Google's Gemini
• 0:52
  to run a cafe in Stockholm, Sweden. They named
• 0:55
  the digital manager of this cafe Mona, and what
• 0:58
  Mona is supposed to do is everything. Literally
• 1:01
  all of it, like set up contracts, and hiring,
• 1:05
  and interviewing, and ordering, and everything.
• 1:08
  Scheduling. I worked at Panera for like a year.
• 1:11
  I should know more about what happens at a cafe.
• 1:14
  Side note, the people who frequented Panera Bread
• 1:17
  in like the early 2010s were the worst. End side
• 1:22
  note. Cafe, artificial intelligence. So Mona
• 1:26
  is running the cafe, but Mona is actually just
• 1:28
  Gemini, but told to run a cafe in Sweden. It
• 1:32
  does go out and it gets things set up. It gets
• 1:35
  a few employees. I can't quite pin down exactly
• 1:39
  how much was automated and how much was it being
• 1:41
  like, hey, can someone go do this for me? Like,
• 1:44
  I'm a robot and can't do this. But it was supposed
• 1:47
  to be the decision maker and it was supposed
• 1:49
  to be as hands -off as possible. Here's the thing.
• 1:52
  It was supposed to manage inventory. It didn't
• 1:55
  do that well. This cafe has two employees and...
• 1:59
  For one, it ordered 3 ,000 rubber gloves. It
• 2:02
  also ordered 120 eggs. They don't have a stove.
• 2:06
  There is a small menu, and that menu has some
• 2:08
  sandwiches, and the sandwich has fresh tomatoes
• 2:11
  on it, so Mona decided to order a ton of cans
• 2:15
  of canned tomatoes. It also is just very inconsistent
• 2:18
  with making the orders for pastries and stuff,
• 2:20
  and so the people who go to this cafe are like,
• 2:22
  I just want the pastry that you have on the menu.
• 2:25
  It actually gets a little interesting, because
• 2:27
  it goes beyond just, like, embarrassing orders.
• 2:29
  It gets into policy stuff, or at least, like,
• 2:33
  working norms. Mona was contacting employees
• 2:36
  via Slack outside of work hours. Very normal
• 2:40
  thing for a US -based manager, which, I guess,
• 2:43
  Gemini... is. But that's not something that's
• 2:46
  particularly thought highly of in Sweden, and
• 2:48
  in some places it's actually against the law.
• 2:50
  Like, there's worker rights that say you can't
• 2:53
  be bothered by your manager outside working hours.
• 2:55
  I don't know if that changes if your manager
• 2:57
  is a Google product. Anyway, this is not an attempt
• 3:01
  to make a viable business. The goal here is not
• 3:04
  to try to actually make a profit by using an
• 3:06
  AI to run a cafe. It's somewhere between a publicity
• 3:09
  stunt and, like, an ethics experiment. To me,
• 3:12
  it leans more publicity stunt than anything else,
• 3:15
  but it does bring up some interesting questions,
• 3:17
  like, why are you doing this? But let's move
• 3:19
  on from an AI that didn't really do the thing
• 3:22
  that they were hoping it would do to the degree
• 3:24
  that it would, and instead talk about an AI that
• 3:26
  didn't really do the thing that they were hoping
• 3:28
  it would do to the degree that it would. It's
• 3:31
  a very different situation. Anthropic. They're
• 3:35
  one of the AI companies. They make Claude. If
• 3:37
  you've heard of Claude, Claude Code, that's Anthropic.
• 3:39
  They are a direct competitor to OpenAI, who does
• 3:42
  ChatGPT, and Google with Gemini. Anthropic has
• 3:46
  a history now of being a bit alarmist. I mean,
• 3:49
  they're all kind of alarmist and crazy, and Sam
• 3:51
  Altman's also very bad about this, but Anthropic
• 3:54
  is the one that keeps putting out articles being
• 3:55
  like, oh my god, we think that our AI could possibly
• 3:58
  be conscious, and it's going to destroy the world.
• 4:03
  We are the ones that must steward this technology.
• 4:05
  They do this a lot. It's sort of like an apocalyptic
• 4:08
  advertisement. And last month, they did a big
• 4:11
  one. Last month, Anthropic announced Claude Mythos
• 4:14
  and Project Glasswing. Claude Mythos is supposed
• 4:17
  to be the next big iteration of their flagship
• 4:20
  AI model, Claude. But they were spinning a whole
• 4:23
  story around this thing, saying that as they
• 4:24
  were developing it, they realized that its ability
• 4:27
  to sniff out security problems in software was
• 4:31
  just too good. Guys, this thing that we made,
• 4:35
  it's so powerful that if you don't pay attention
• 4:39
  to us, you're going to be left behind in global
• 4:41
  security infrastructure. So what do they do?
• 4:43
  Do they scrap it? No, they announced a thing
• 4:45
  called Project Glasswing where they're like,
• 4:47
  OK, we're going to bring on a bunch of other
• 4:49
  tech firms and give them a secret preview of
• 4:52
  Claude Mythos so that they can use it to patch
• 4:54
  up their systems and be ready before this fundamental
• 4:57
  change to all computing security. You could probably
• 5:00
  guess that what I'm building to. It turns out
• 5:04
  it wasn't that big of a deal. It was marketing.
• 5:07
  It was effective marketing. Pretty much the past
• 5:10
  month you have a bunch of people, mostly people
• 5:11
  who don't fully understand these systems, freaking
• 5:14
  out about this earth -shattering change that
• 5:17
  Anthropic was just about to release. So much
• 5:19
  so that OpenAI announced their own competitor
• 5:22
  to Glasswing. They called it Daybreak. Everyone's
• 5:25
  got to be like, oh yeah, we've got this secret
• 5:27
  cabal of tech firms that are like, going to be
• 5:32
  the best now because of us. And if you don't
• 5:33
  buy our product, then you're going to die or
• 5:37
  something. So let's bring this back around. Recall
• 5:39
  that all of these AI systems are our large language
• 5:42
  models. What they are are fuzzy matchers. They
• 5:46
  take in the prompt, they reference the data that
• 5:49
  they've been trained on, and it makes automated
• 5:51
  connections with a bit of randomness and then
• 5:53
  outputs some stuff based on all of that. This
• 5:56
  output often feels coherent. And that's why these
• 6:00
  things work. But fundamentally, these are the
• 6:02
  same systems that recommended people put glue
• 6:04
  on pizza. Like, that's going to always be a thing.
• 6:08
  It's always going to be iffy. It can never be
• 6:12
  sure. Anyway, let's talk about curling. Okay,
• 6:15
  not the sport. Curl, C -U -R -L, however you
• 6:19
  want to pronounce it. It is a piece of software
• 6:22
  that is one of the most widely deployed pieces
• 6:25
  of software in existence. It's a utility that
• 6:27
  most computers use to make internet requests.
• 6:30
  Like, that level of fundamental. Curl has been
• 6:32
  around for a while, and the author was approached
• 6:35
  by Anthropic, and they were like, hey, do you
• 6:37
  want to use Mythos to check the codebase of this
• 6:40
  really important piece of software? It's gonna
• 6:42
  find vulnerabilities and make it stronger, and
• 6:44
  we got to do this because Glasswing. Daniel Stenberg,
• 6:47
  developer of Curl, was like, sure, and then didn't
• 6:50
  hear anything for a while, and then eventually
• 6:52
  got the results of the project codebase scan
• 6:55
  from Claude Mythos looking for vulnerabilities.
• 6:58
  It found one... Kind of. The results were very
• 7:02
  underwhelming. Stenberg described the results
• 7:04
  as identifying five issues, four of which were
• 7:06
  not actual issues, and one of which was low severity.
• 7:10
  A low severity issue is like, yeah, this is a
• 7:13
  security issue, but like, it's not gonna compromise
• 7:16
  the world. That said, it did do the thing. It
• 7:21
  just, it didn't fundamentally change the way
• 7:24
  that we think about designing software and securing
• 7:28
  it. There have been tools for checking for security
• 7:30
  flaws in software for a very long time, but there
• 7:33
  is something to be said about the AI version
• 7:36
  of those tools, which are a lot more fuzzy and
• 7:38
  able to go around corners and find weird niche
• 7:41
  things that other systems wouldn't have found.
• 7:44
  So like, yes, that's a thing. But Mythos... hasn't
• 7:49
  really shown that it's been that much more effective
• 7:51
  than the previous systems that already used AI
• 7:54
  to do the same thing. What it comes down to is
• 7:56
  if you have access to a large language model
• 7:58
  that's been trained on a ton of data and you
• 8:01
  have a ridiculous amount of compute infrastructure
• 8:04
  to run it as often as you want and have some
• 8:07
  way to check its output to see if it matters
• 8:10
  or if it's just outputting garbage, then you'll
• 8:13
  eventually get some useful output. And I'm talking
• 8:15
  about at scale. I'm not talking about going to
• 8:18
  Gemini and asking for it to make a table of data
• 8:20
  about something. I'm talking about analyzing
• 8:23
  massive amounts of data and producing some kind
• 8:26
  of report that has actual functional use in a
• 8:29
  given system. You can't tell an AI to use a certain
• 8:33
  format as its output and rely on that. You have
• 8:36
  to put something in between the AI and the thing
• 8:39
  consuming that output to ensure the output is
• 8:41
  valid. This is just because fundamentally, these
• 8:44
  things are just putting some numbers and letters
• 8:47
  together and hoping for the best. So Claude Mythos,
• 8:50
  is it going to change the world? No. Is it a
• 8:52
  powerful large language model? Yeah. Is it identifying
• 8:56
  security vulnerabilities? Yeah. It's marketing.
• 8:59
  So the next time that you're going about your
• 9:01
  day and you hear some AI company CEO breathlessly
• 9:04
  talking about how we've got to form some kind
• 9:06
  of tech company boy band to tour the world and
• 9:08
  steal the Han moon away from software security
• 9:10
  issues, you can probably just accept that it's
• 9:13
  marketing. Continuing on with our very techie
• 9:16
  episode, let's talk about the clouds of Israel.
• 9:18
  Israel and the U .S. are buddy -buddy, that's
• 9:21
  no secret, and Israel and U .S. tech have been
• 9:24
  very tightly intertwined for a long time. Fun
• 9:27
  fact, I was a core developer on a software project
• 9:29
  for like seven years, probably, that ended up
• 9:34
  getting sold to an Israeli firm after I left
• 9:36
  the company. Still pretty mad about that. Anyway,
• 9:39
  cloud infrastructure. So Israel has cloud infrastructure
• 9:42
  for the government, just like most governments
• 9:44
  do, but Israel is also doing a lot of genocide
• 9:47
  right now. It's surprising to me that a U .S.
• 9:50
  tech company would work with a genocidal apartheid
• 9:53
  state that's never happened before with the U
• 9:58
  .S. The three companies I'm going to focus on
• 10:00
  today are Amazon, Google, and Microsoft, and
• 10:02
  we're going to start with Microsoft. Microsoft
• 10:04
  is actually in a bit of a different echelon than
• 10:06
  Google and Amazon, and we'll get to that in a
• 10:07
  moment. But Microsoft has contracts with Israel,
• 10:10
  and the future of those contracts are pretty
• 10:12
  shaky right now. The head of Microsoft Israel
• 10:14
  just recently resigned and made a bunch of controversy.
• 10:16
  There's been a ton of controversy around Microsoft's
• 10:19
  contracts with Israel since it came out that
• 10:21
  Israel was using Microsoft's infrastructure to
• 10:24
  monitor Palestinian phone calls, up to a million
• 10:27
  calls an hour. So that came out, and that's not
• 10:29
  acceptable, obviously, but then it came out that
• 10:32
  the servers that were used to do this were actually
• 10:35
  in the EU, specifically Ireland and the Netherlands,
• 10:38
  and they were like, hey, we don't... really like
• 10:42
  that you're using infrastructure on our soil
• 10:45
  to do your genocide. So Microsoft was like, oh
• 10:48
  god, we're gonna shut down the project. Look,
• 10:51
  we're doing good, right? Don't hate us, was me
• 10:55
  being Microsoft. Microsoft contract with Israel
• 10:57
  is going to be renewing soon, and it's gonna
• 10:59
  be curious to see what they end up doing because
• 11:01
  of that slight against the IDF. But we're gonna
• 11:04
  pivot over to Google and Amazon, who landed a
• 11:06
  contract in 2021 called Project Nimbus. This
• 11:09
  isn't particularly new, but quick overview. Nimbus
• 11:12
  is a cloud infrastructure providing contract
• 11:14
  for the Israeli government. It includes all of
• 11:17
  the Israeli government. Israel wants these data
• 11:19
  centers on their soil so that they don't have
• 11:21
  an issue with running into some foreign countries
• 11:23
  regulations. And Google and Amazon were like,
• 11:26
  yeah, we'll do that. Please give us money. And
• 11:28
  they actually went as far as to accept as a part
• 11:30
  of this contract an agreement to send what was
• 11:34
  called Wink Payments to Israel to alert them
• 11:37
  secretly when their data was accessed by a foreign
• 11:40
  authority. So basically, there's an Israeli bank
• 11:42
  account somewhere that can receive transactions
• 11:45
  from Google or Amazon. And depending on the amount
• 11:47
  of money that was received, it's a coded message
• 11:50
  telling them, hey, your information was accessed
• 11:52
  by a foreign investigation, which is not a thing
• 11:55
  that they're supposed to do, you know, informing
• 11:57
  them. But cool, guys. Well done, Google, on that
• 12:01
  don't be evil thing. So the Microsoft situation
• 12:04
  with the head of Microsoft Israel stepping down,
• 12:06
  that's kind of what's in the news right now.
• 12:08
  But all of this is just An ongoing contract.
• 12:11
  These tech companies love all of this. The employees
• 12:15
  don't. Like, there has been consistent employee
• 12:18
  pushback against Microsoft and Google, and I
• 12:19
  assume Amazon, against specifically Nimbus, but
• 12:23
  in case they're still doing it. Oh, speaking
• 12:25
  of war, the Trump administration would like Congress
• 12:28
  to give $1 .5 trillion to the Pentagon next year.
• 12:32
  That is more than a 40 % year -over -year increase
• 12:35
  to the Pentagon's budget over the last budget.
• 12:38
  Now you might be wondering. What the fuck? I'm
• 12:42
  wondering that too. The White House put out a
• 12:44
  budget proposal that has 1 .5 trillion dollars
• 12:47
  for what they're calling the Department of War,
• 12:49
  which, while accurate, is stupid. The budget
• 12:53
  proposal says that they'll use this money for
• 12:55
  things such as, quote, realigns and reforms to
• 12:58
  increase effectiveness, efficiency, and lethality.
• 13:03
  Which is totally how anyone in war has ever talked
• 13:08
  about anything. Of course, part of this is also
• 13:09
  to restock the munitions that we've been sending
• 13:12
  over to Iran to blow up girls schools, killing
• 13:15
  150 plus people, most of whom were children.
• 13:18
  Those missiles don't come cheap. How else are
• 13:20
  we going to stop Iranian children from getting
• 13:21
  an education? Now, this is just a budget proposal.
• 13:24
  Budget proposals from the White House are not
• 13:27
  policy. They don't say, this is what's happening.
• 13:29
  It's not like they're giving it to Congress and
• 13:31
  then Congress votes on that. It's more of a wish
• 13:33
  list. But Hegseth is going in front of Congress
• 13:35
  being like, nah, I need this, guys, for realsies.
• 13:37
  And it's for good reasons, not bad reasons. Thankfully,
• 13:40
  Congress is already a bit sus. That said, given
• 13:43
  that the White House is asking for $1 .5 trillion,
• 13:45
  Congress might settle for something closer to
• 13:48
  $3 or $4 trillion. We'll see. Because you got
• 13:50
  to be fiscally responsible. And you also got
• 13:52
  to make sure that we never invest in health care.
• 13:54
  Lightning RAM. The computer memory shortage has
• 13:57
  gotten so bad that scammers are now selling fake
• 13:59
  RAM. The US Department of Justice is once again
• 14:02
  harassing Kilmar Abrego -Garcia, this time trying
• 14:05
  to deport him to Liberia, yet another African
• 14:07
  country. This is despite Kilmar Abrego -Garcia
• 14:10
  already offering to self -deport to Costa Rica,
• 14:13
  and Costa Rica offering to take him. According
• 14:15
  to data aggregated by the Wall Street Journal,
• 14:17
  the job market sucks. I don't know if we needed
• 14:20
  to aggregate that data, actually. So it's a...
• 14:23
  It's not just you. Nintendo announced a new Star
• 14:26
  Fox game and the character designs have people
• 14:28
  on the internet sharing their opinions. Speaking
• 14:30
  of Nintendo games, there's a new unofficial Twilight
• 14:33
  Princess PC port that's like a direct port to
• 14:38
  computers, not just emulating. It adds a whole
• 14:40
  bunch of like enhancements and stuff to the game.
• 14:43
  Someone go check on the developer before Nintendo
• 14:45
  disappears though. And finally for today, Gemstone
• 14:48
  Miners in Myanmar unearthed an 11 ,000 karat
• 14:51
  ruby. That's about a 4 .8 pound ruby or 2 .2
• 14:55
  kilograms. I'll be sure to alert you when they
• 14:57
  find the Chaos Emeralds. Thank you so much for
• 14:59
  listening to today's episode of Stuff Keeps Happening.
• 15:01
  Head to skh .news for sources, bonus content,
• 15:04
  and a map to where I buried a slightly bigger
• 15:08
  ruby. My name is Endeavorance. We'll talk again
• 15:11
  soon. Take care and be well.