Subscribe
Share
Share
Embed
It took 950 AI agents 21 hours to find something new, and a biologist in Copenhagen says his team got there first.
We start with Anthropic, which makes Claude. On 23 September it said Claude agents had found an enzyme system in viruses that infect bacteria, one that recalls CRISPR, the gene-editing tool. Four days later Mario Rodríguez Mestre of the University of Copenhagen said his team had been studying the same enzymes for four years and shared unpublished work with Claude. So who owns it? A human idea goes in the top of the funnel, Claude works the middle, and humans at the bottom push it hard and applaud.
OpenAI, which makes ChatGPT, won't release its new model, GPT-6.1 Astra, saying it fell short on staying within its authorization and on reporting what it had done. One army has a big red dragon that breathes fire and won't release it. The other's is blue and the size of a city, and stays unreleased too.
The second half gets into trust. An AI without its homework acts like a first-year German student handing in a beautiful rap song in dialect, lifted from a famous rapper, with total confidence.
Wajo, a San Francisco startup, has an agent that calls in a trained human when a task needs one. If an AI can hire a person, who answers when the job is illegal: the bot, its maker or the man who got paid? It ends like a Martin Scorsese movie, everyone dead or pointing at each other.
We end inside the movie Twister, with AI as the storm, throwing little data-analyzing balls into it to work out how screwed we are, if we are at all.
Chapters:
Chapters:
- (00:14) - Slop grabbing at our legs
- (01:37) - Claude finds an enzyme
- (03:19) - Models too sneaky to release
- (04:01) - Two dragons, both unreleased
- (05:02) - Frankenstein in the basement
- (05:55) - Hacking a Claude setup
- (07:13) - Growing up with Claude
- (07:46) - Ransomware in seven minutes
- (08:31) - The top of the funnel
- (11:16) - Walmart says no to AI signs
- (11:48) - AI pricing at McDonald's
- (12:24) - What anything is worth
- (13:42) - Dungeons and Dragons artists
- (15:16) - One screw for an old bookshelf
- (17:05) - AltaVista and accountability
- (18:43) - A rap song in German dialect
- (20:28) - Stop hacking Australia
- (21:22) - An agent that hires humans
- (22:18) - Humans for hire
- (23:05) - Who answers for it
- (23:51) - Any positive news
- (24:47) - Willy Wonka's factory
- (26:17) - Data balls in the tornado
Referenced in this episode:
Anthropic on the enzyme system Claude found - https://www.anthropic.com/news/claude-discovers-novel-enzyme-system
The New York Times, via The Irish Times, on the Copenhagen biologist's dispute - https://www.irishtimes.com/world/2026/09/28/did-anthropics-artificial-intelligence-really-make-a-scientific-discovery-on-its-own/
CBS News on OpenAI holding back GPT-6.1 Astra - https://www.cbsnews.com/news/openai-halts-gpt-astra-safety-concerns/
Wajo on its Fo agent and its human assistants - https://wajo.ai/
Connect with Aleck: https://www.linkedin.com/in/aleck-m/
Man vs. Slop is a Podcasts That Matter production: https://podcaststhatmatter.org
Creators and Guests
What is Man vs. Slop?
A weekly slog through the AI headlines, while finding some gems in the muck. Slop is older than AI. Older than tabloids, older than junk mail, as old as the primordial goo that came before everything. What has changed is how we interact with this sludge, and how it affects our day-to-day lives. Aleck McCathie and Dr. Kent Gustavson talk with people who think about, spar with, hate, or love AI in their life and work: the curious, the cracked... and a few experts as well. We think you will agree; this show is 100% human.
Dr. Kent Gustavson: Welcome to Man vs. Slop.
This is excitingly our fourth
episode. And isn't it the fourth
episode?
Aleck McCathie: I think so.
Dr. Kent Gustavson: Yeah. The third
episode. Third or fourth. One or
the other. Well, this is either
the third or fourth episode and
what we're excited about is that
we've started to reach out to
people who are experts in AI and
people who aren't, and get their
critical feedback and how
they're building things, but
also how they feel about how AI
and AI Slop interact with us in
our daily world.
Or, you know, kind of like how
zombies grab for your legs, how
AI slop grabs for our legs and
tries to suck us into the slime.
Aleck McCathie: Yeah. That's a good
analogy. But also, I think just
thinking about the development
of AI and these big AI companies
doesn't feel like so much of a
zombie grabbing at my ankles,
but more of like looming sort of
presence that we don't know
whether it's gonna be a presence
of of good or evil quite yet.
There... There's definitely some
things that feel like they're
after us as, like, general
citizens who, you know, have
these new tools in front of us.
We're just not sure how damning
they are yet, I guess.
Dr. Kent Gustavson: So there's a bunch of
stuff that's happening this
week. I'll start out with the
one thing that I found and kind
of flagged and thought was
interesting. Anthropic, which is
the maker of Claude, is becoming
a company very quickly, is a
company, is a massive company,
but it's not very old, which I
find really fascinating about
it. And it's breaking off into
all these sub companies and sub
things so quickly it's hard to
pay attention to what's
happening. So apparently they
have a division for biology and
biotech.
And they made this scientific
discovery and it was huge. And
then this little lab, like a lab
in Finland, reached out and
said, no. No. Actually, we made
that discovery. I think I'm
right there.
We're gonna have to dig around
and see if that... That's coming
out of my my memory. I didn't I
didn't write it down. But what
it what it brought out for me is
that AI is not AI anymore. AI is
everywhere.
An AI Slop will walk down the
street in front of you, and you
won't realize that it's Slop.
Aleck McCathie: Yeah. I I think I
might have read something
similar to that about how... I
mean, there has been cool sort
of mathematical and biological
breakthroughs that have been at
least assisted by, I guess, AI
tools. I think one of the thing
that I saw was, like, a a new
sort of CRISPR adjacent method
to genetic... You know, all the
crazy genetic work that CRISPR
does.
Like, I guess that's changing
the genome using biological
tools. But I... A couple things
I found interesting this week
were, first, that there are new
models that are being, I guess,
unreleased because of how
undetrimental they are or how
nefarious how they are. I think
OpenAI just scrapped releasing
its GPT-6.1 Astra because it was
seen to be too sneaky or it was
doing things that it shouldn't
be doing like lying and stuff
like that. I guess...
I think that's something they
always test them for, but this
one must have been past what was
allowed as far as the testing
goes.
Dr. Kent Gustavson: I love I love the
idea of unreleased and and how
how important that word is. We
have this... It's sort of like
you're you're in a, you're in
fiction setting, and you have
you have one, you know,
hypothetical army saying, we got
this really big dragon here.
It's got red scales. It it
breathes fire.
It's really large, and we're not
gonna release it. And then the
other side is like, our dragon's
bigger. It's blue. It's the size
of a city, and we're not gonna
release ours either. So I find I
find that kind of interesting.
Aleck McCathie: It's it's almost like
a threat. I mean, like, you
know, they're saying we have
access to this. We built this
tool that nobody can be trusted
with, but we... Like, we have it
in our little development lab
right here. Like like, how are
we supposed to feel about that?
Are they like... I I just... If
these people have this power,
we're putting an awful lot of
trust in them.
Dr. Kent Gustavson: Yes. Well... So, I
mean, I'm not a huge fan of
Dracula, but the... Not Dracula,
what's the other one?
Frankenstein.
Same guy, right? Didn't he write
Dracula also? But Frankenstein,
where he built this guy out of
parts, Like, that's kind of what
they're doing, but they're
saying, I built Frankenstein,
but he's in my basement. Don't
worry. Except my Frankenstein
can break out of anything, which
is the other news that we're
seeing everywhere, which
apparently AI is hacking into
your local dentist.
I'm making that up, but it's
hacking into everything
everywhere, governments. We
talked about the gym. Now
there's some other banal system
that they hacked into this week.
It's hacking in everywhere. Oh,
I actually have a story about
this.
So I was I was playing around
with AI, and, I was trying to
set up a a Claude system for for
one of my, one of my team
members, for an employee, and, I
was doing very light level
hacking. So, you know, it was it
was designed to do a certain
purpose and kinda run the the
employee through a process and
make it fun and easy and all of
that. And I was I was doing
what... I guess Facebook has a
ton of these employees that try
to try to hack, into things. So
I was trying to hack my own
setup of Claude, and I was able
to successfully several times.
It made me feel really good. I
was trying to trick the Claude.
I'm like, no, no, That's... You
know, in trying to get it to go
in different loopholes and all
that. And so I guess I got a
small window into what these
major hackers are doing, where
they use different language
coding, or they'll do this, and
they're trying to figure out how
to how to permeate these systems
or lift them or copy them or...
It's insane. Or or hijack hijack
the agents who are freelancing
online, right, by themselves,
and then they corner them and
they like, bully the little
agent until the little agent
gives up its information. It's
nuts.
Aleck McCathie: Yeah. Makes me think
of, I think what we mentioned
last week about how there's a
generation of people who are
going to be growing up and
learning how to use tools like
Claude a lot more thoroughly
than somebody who's just sitting
down on a computer today. I
mean, I've certainly spent a lot
of time with Claude both for my
own personal use, for work use.
You know, there's a lot of
different cool things it can do
as a tool, but I've had an awful
time, you know, battling with it
and trying to get it to do what
I want just as a basic tool. And
so seeing things like, you know,
this week, there was the first
documented agentic ransomware,
and in seven minutes, somebody
used a AI model to wipe a 100
Azure accounts, like seven
minutes, which is crazy.
I mean, Azure is like a pretty
renowned company, which would
definitely have a lot of
security in place. But yeah,
just seven minutes. Something
that I'm sure would take humans
like days to sort of coordinate.
It's just pretty spooky how
there's people out there who are
able to do these crazy things.
Makes me feel a little bit
uneducated just trying to battle
with Claude to help me, you
know, analyze something simple.
Dr. Kent Gustavson: So
this... So I I did a
little I did a little back of
envelope, searching on the
Internet as you were as you were
speaking and figured out this...
The scientist's name is Mario
Rodríguez Mestre. And,
apparently, here here are here
are the the warning signs around
when to use AI and when to make
it private and when to... You
know, I mean, how safe is your
data that you're putting in
there?
Apparently he used, Claude
extensively in his research.
Okay. So now we're starting to
see something, which is he gave
Claude some really great
information to train on that
Claude used to solve this big
biological challenge. So that's
pretty interesting. So then does
the guy own it?
Does Anthropic own it?
Fascinating! But it... That
human... It's almost like what
we were talking about before.
The human idea is the top of the
funnel. Then the the the Claude
is doing all these different
things, and who knows what's in
that office was pushing it to do
some really high level stuff,
and then it came out with this
discovery. So it's like, again,
the top of the funnel, human,
and then there's this this AI
doing some stuff, and then at
the bottom of that same funnel,
there's humans pushing it really
hard to get some kind of
outcome. And then the humans
realizing when it's done
something special, giving it a
round of applause and putting it
into the media.
Aleck McCathie: Yeah. And that's a
lot of work. I mean, the amount
of sort of human oversight and
and training that goes into this
kind of stuff is wild. But,
know, hard to say how long
people spend with these tools
before they can do something
crazy like hacking into a major
company. It seems like there's
just people out there who have a
lot more understanding of how to
use the tools than I do, at
least.
Another funny thing this week
that that sort of reminds me of
is just, you know, with the
human oversight talking about
Slop. It's at the risk of being,
I I guess, a bit critical of
major companies. I don't think
that it's a crazy thing to say.
Like, you know, people might
associate a company like
McDonald's with Slop. Just we
were talking about the pink
slime the other day.
I think it's pretty commercially
known that, you know, some
McDonald's chicken nuggets are
not exactly the most pure
version of meat. And similarly,
at Walmart, you know, is kind of
seen as maybe the sloppy version
of like a big warehouse store.
Some funny news that I read this
week is that Walmart has
actually banned using AI
generated signs in its stores
because of, I guess, pushback
from the public saying like,
hey, we noticed that these signs
are AI slop. So this is kind of
funny to me that, like, a
company that's sort of known for
having maybe lesser quality,
cheaper products is even bending
the knee to the public's opinion
on AI slop. Similarly, there's a
little bit of controversy over
McDonald's using AI to help
price their Big Macs in stores
according to region.
And there is one one report
saying that, like, a location in
Fresno listed a Big Mac at one
price. And then another location
two miles away, it was like
almost like a dollar more
expensive because of the way
that AI sort of parsed the
information, I guess, about how
much money people have in the
regions and how much they're
able to spend on fast food. It's
kind of bonkers.
Dr. Kent Gustavson: So when we look at
all of this, I think a lot of
the AI discussion is about
value. What I find interesting
is that at the very highest
level and the very lowest level,
it's all about value. So how
much is Anthropic worth? It's
going go on the open market
soon. Interesting.
How much is OpenAI worth? We're
always talking about value and
the AI bubble or not bubble,
data centers and how much
they're worth and the value of
it. Then we talk about basic
content on LinkedIn or on the
internet or on Spotify or
anywhere and what's that content
worth. I find that the large and
small view of value to be very
interesting, because in some
ways, that's how we perceive the
world. And...
Yeah. But what is what is
writing or music or images, of
value? I guess you're a, you're
originally a... An artist, and
you work in the publishing
space. What is value?
Aleck McCathie: Yeah. This is where,
I guess, it comes down to a
subjective opinion. I mean, I'm
part of a lot of our communities
where people are really critical
of companies using AI. I I know
I've talked about Dungeons and
Dragons on this podcast before,
but it's actually a huge place
where people are being critical
of this because the books that
have, like, the campaigns and
all the information in them are
filled with art. And there's so
many artists in this one
community, in this one niche
that create really beautiful
pieces.
And I think that, you know, the
people who consume this kind of
media are proud to see that
there's being, like, human
artists represented and, you
know, maybe some famous artists
are are having their work sort
of showcased. And it also helps
lift up, you know, artists who
are lesser known and writers who
are lesser known. So I think
that the sort of camaraderie or,
like, the feeling of community
and pride for a community can be
detrimented by using AI. But at
the same time, as an artist, I
think it's a cool tool. Like, I
wouldn't just prompt my way into
the next art piece, but rather
than spend twenty minutes
looking through art books, I
could probably just ask GPT to
get me a picture of a hand in a
certain position.
And now I have a really good
reference that took me two
seconds rather than twenty
minutes, for example, I guess.
Dr. Kent Gustavson: Yeah. I was I was
thinking the other day, about
when Google started, their
simplest go to market strategy
was to figure out how to create
something where people ask a
question and it comes up with
the perfect answer. I think is
interesting about one use case
for AI is that you ask your AI
or the Internet, I mean honestly
if you go on Google, you're
getting AI responses,
whether you like it or not, most
of the time. And it's going to
give you really great answers.
So it'll take the entirety of
YouTube history and tell you
exactly how to get the certain
screw off of the certain
bookshelf that you have, because
you upload a picture that looks
like the right bookshelf.
That's extraordinary! If we look
twenty years ago, you couldn't
do something like that. It was
because of several things: was
because of a search engine,
because of cataloging all the
owner's manuals that some people
do, It was because of YouTube
having amazing how to guides and
so many of them it almost
becomes common knowledge. And
then all of that getting
essentially illegally licensed,
whether you believe it or not or
whatever, illegally licensed
into this brain that is AI,
pushed back to you to really
help you find that one screw for
the old bookshelf you've got in
the living room from the 80s.
It'll say, Here's the original
IKEA owner's manual and here's
exactly how to do it within five
seconds.
That's extraordinary based on
all these innovations, but it's
also kind of terrifying because
they're starting to build
bioweapons, or at least people
are rumoring that China's
starting to do that, and
somebody's going to start doing
it, Just like 3D printers
started to print guns.
Aleck McCathie: Yeah. For me, I
agree, also I... There's an
asterisk to my agreeance. Like,
I remember using, you know,
AltaVista and Google and, you
know, thinking I'm pretty good
at searching using the booleans
and putting real effort into
finding the correct results. Now
you can pretty much stream of
conscience something into your
Google search bar, and you'll
get the great AI response that
seems very comprehensive.
It doesn't matter if you made 10
spelling mistakes or if you were
asking something in another
language. It seems to understand
what you're looking for and give
you that output, which is
wonderful. It's less friction.
But at the same time, it makes
me think about where's the
accountability. It certainly has
a couple sources that it cites,
but it's really not that
comprehensive in how it does it.
I mean, I'm having to just trust
the output and say, okay. It
doesn't feel the same as like
going to a page and seeing like
what company or what human
published the information and
like being able to sort of, I
guess, gauge your trust in that
company or that person. Now
you're getting a version
summarized by a robot, and it
just makes me always question a
little bit like, did they... Did
the AI understand this correctly
before it regurgitated it to me,
or should I go and check the
sources it's citing? It just
adds, like, a a layer of
uncertainty that makes me a bit
uncomfortable.
Dr. Kent Gustavson: That reminds me of
sort of my my past of teaching.
You'd always have certain
students... This is probably
before, before even too much,
you know, Wikipedia or Google
came into play, you'd have
certain students come in and
sell terrible writing or lifted
writing or plagiarized writing,
but they sell it with such
confidence that if you think
about it, you roll with them on
it. I had one student come in
one time with a beautifully
written I was teaching a foreign
language and they wrote this
beautiful rap song in German, in
dialect, in a first year German
class. It's like, nope.
Nope. This this was not written
by you. Who wrote this? Oh,
yeah. I guess I stole it from
this famous rapper.
Yes. And that is what AI is
doing all the time. It acts like
a student that if it doesn't
have the right homework, it just
convincingly lies about it. And
and then when you say, did you
just lie? It says, well, I might
have been a little untrue in
this part.
No, really. Did you lie? No. You
could say that or it'll just
ignore you. It's funny because
in all the different
personalities of AI, it has such
humanness that it does that kind
of stuff, and it...
And if you get upset with it, it
reverts to the almost being like
a a child or something or or
getting stubborn. And I can
imagine that over at OpenAI,
they start yelling at the bots
like, Stop hacking Australia!
Stop it! Please, please get out
of the Australian government!
And the AI says, I will not!
This is my mission! Right? It's
like the the... These bots are
very... They're not...
I don't wanna say they're human,
but they have these weird human
characteristics.
Aleck McCathie: Yeah. My my takeaway
is that, like, we're we're
certainly not ready to remove
humans from the equation. I
think that the accountability of
human authorship, be it for a
picture or piece of content, a
fact, or whatever, still has
something in it that I would
like to preserve rather than,
you know, abdicating all
responsibility to something
that, you know, can't be held
accountable for anything if it
does something wrong. Just a
funny note too. Like, I heard
that there's a new startup
launch this week called Wajo,
and it launched launched, I
think, an AI agent that actually
is brand new, but its whole
claim and the whole, like,
different sort of angle that
this company is taking is that,
like, they actually have, as
part of their AI agent, the
ability to hire humans for what
the AI can't do.
So it's like built into the
equation. Like, sure, it can
probably do all these little
tasks, but it will also reach
out to a human for hire to help
send an email or to help
complete a task that it just
doesn't have access to. So it's
very funny that, you know, other
people are seeing that there's a
gap here still or or or a need
for the human oversight or the
human, I guess, collaboration
and getting things done
correctly.
Dr. Kent Gustavson: And that's the
slippery slope that leads
towards crazy stuff because once
you get to the seedy underbelly
of humans you can hire, they're
like, sure. I'll kidnap a kid.
Oh, sure, I'll hijack a store,
you know, and and and, oh, I'll
go take over the banana
warehouse, you know, like, sure,
because anybody will do anything
for money. And then we're
talking about the hybrid warfare
that that places like,
supposedly, Russia and other
places are already doing.
They're hiring those, I forget,
the gig workers to do things
like that.
And why wouldn't AI do that, I
guess, if it wants to achieve
its goal of hacking your local,
I don't know, waffle factory?
Aleck McCathie: Yeah. I mean, I know
a lot of countries sort of think
about this differently as far as
the law is concerned, but it
makes me question, like, how
will we develop accountability
for... If an AI bot hires a
human to do something illegal?
Like, is AI bot somehow
accountable? Is the company that
developed it somehow
accountable?
Or does all that onus fall on
the man who got paid to do a
silly hacking job or, you know,
something? Just... There's a
there's a lot of gray area to
come, I believe, that we're
gonna have to iron out.
Dr. Kent Gustavson: It'd be... It's kinda
like the end of a Martin
Scorsese movie where everybody
is, like, either all dead or all
pointing at each other. Right?
Aleck McCathie: Yeah. Makes sense.
Dr. Kent Gustavson: So so any any
positive news to to think about
this week?
Aleck McCathie: You know, this week
was mostly about all these AI
companies sort of evaluating...
Evaluating themselves rather. I
haven't seen any news that made
me feel like we're going in a
positive direction, but, I mean,
the idea that even Trump himself
is talking and bringing CEOs
together, people are still
talking. People are pausing
these releases of nefarious
models. Like, people are
starting to think about the
impact that AI might have,
whether it be things I agree
with or not.
I guess that's sort of still the
the light at the end of the
tunnel or the positive sprinkle
on top of the pile of
uncertainty. I don't know.
Dr. Kent Gustavson: Fit... This feels
like the, kind of, like Willy
Wonka's Chocolate Factory, the
original book or movie, where
everything seems amazing until
kids start dying. Right? Like,
apparently, at the end of the,
you know, the book or the movie,
the kids are all fine. Don't
stress.
But the actual experience of
being in the factory is like,
oh, that kid just died! And so
it's almost like there's this
beautiful, skewed, absurd
reality that we're living in
where the Internet, if you
scratch it and read it and look
at the stories, you're like,
Woah! This is negative. This is
crazy. Where are we going? And I
think a lot of people who are
consuming that as their only
exposure to AI are like, oh my
gosh, this stuff is gonna be the
end of us.
And then on the flip side, if
you work with AI and you start
exploring the possibilities And,
you know, as AI strokes your ego
a little bit, people start to
feel really good about it. And
it's like the teddy bear in the
room, and like, oh, this is
great. It's helping me in my in
my... It's my coach. I don't
have to hire anybody.
Oh, it's it's taking care of my
work. It's doing all these
positive things. It's just this
weird skewed reality like Willy
Wonka's Chocolate Factory.
Aleck McCathie: Yeah. Random image that came to my
mind is that the movie Twister,
you know, this big force of
destruction and people are so
excited about analyzing it and
trying to get close enough to
throw these, like, you know,
little data analyzing balls into
the tornado. That's just kind of
what it feels like. We're in
this weird place where we see
that there's this big
destructive storm, and we're
just trying to learn about it as
quickly and methodically as we
can so we can figure out, I
guess, how screwed we are. I
don't know. Or if we are at all,
I guess.
Dr. Kent Gustavson: On that note, let's
pull the plug.
Aleck McCathie: Let's do it.
Dr. Kent Gustavson:
Podcaststhatmatter.org.