#59: Anthropic CEO Says Human-Level AI Is 2-3 Years Away, How Hackers Are Trying to Make AI Go Rogue, and Fake AI-Generated Books on Amazon

It’s been another interesting week in the world of AI…with a few
things we need to keep our eyes on. Paul and Mike break it all
down—and then some—on this week’s episode of The Marketing AI Show.
Anthropic CEO joins the Dwarkesh podcast to talk about the future
of AI. Dario Amodei, CEO and co-founder of Anthropic (maker of the
Claude 2 large language model released in July of this year), just
gave a wide-ranging interview on the future of AI. The interview
took place on a recent episode of the Dwarkesh Podcast (linked in
the show notes). It’s a must-listen, primarily because these types
of interviews aren’t all that common. Due to considerations around
competition and security, the heads of major AI outfits don’t
always share in-depth their views on the industry and where it’s
going. Not to mention, Amodei himself has a relatively small
footprint online, so hearing from him is even less common. We’d
encourage you to listen to the entire episode, but on our podcast,
Paul and Mike call out some big highlights that have us thinking a
little differently about the future of AI. Red-teaming at DEF CON
finding flaws and exploits in chatbots If you aren’t familiar with
“red-teaming,” it’s a critical aspect of making generative AI
models that are as safe and aligned as possible. For example, GPT-4
was red-teamed for 6 months before its release in March 2023. This
week, top hackers from around the world have converged at DEF CON
in Vegas to find flaws and exploits in the latest chatbots from
OpenAI, Google, Anthropic, and Stability. The teams working on red
teaming often find these exploits by trying to “break” these
systems in novel ways and by imagining creative, though nefarious,
ways in which AI tools can be misused. The Washington Post shares
examples of red-teaming: “AI red teams are studying a variety of
potential exploits, including “prompt attacks” that override a
language model’s built-in instructions and “data poisoning”
campaigns that manipulate the model’s training data to change its
outputs.” The results of the competition will be kept under wraps
for a few months, so companies have time to address any issues
highlighted by the red teaming efforts. Award-winning author finds
AI-generated books written in her name Author Jane Friedman, who
has written multiple books and was named a “Publishing Commentator
of the Year” for her work in the publishing industry, woke up to a
nightmare this week. A reader emailed her about her new book which
just hit Amazon. The nightmare wasn’t due to getting a terrible
review from a reader, it was due to the fact Friedman hadn’t
written a new book at all. Friedman quickly discovered that half a
dozen books had been published under her name that she didn’t
write—and the books were AI-generated. The fake titles have since
been removed from Amazon, but not before Friedman met resistance
from the company. Paul and Mike explain the situation…and the
implications. There are rapid-fire topics to be discussed,
including Zoom backtracking since last week’s episode, and much,
much more. Tune in!