ChatGPT Creative Upgrade and Crazy LLM Test
Key Points
- OpenAI’s latest weekly update to GPT‑4 (referred to as “gp40”) emphasizes better creative‑writing abilities, even though the model’s core version hasn’t changed.
- The improvement is meant to help users draft marketing copy, SEO‑optimized blog posts, and other “creative text construction” tasks—not to replace novelists or poets.
- A quick test asking the model to write a poem in the style of poet Jack Gilbert showed a noticeable upgrade in handling layered meaning, though it still falls short of true human‑level artistry.
- A separate line of research is deliberately giving large language models vague or no direction, letting them converse with each other or with a researcher to see how they behave when “left to themselves.”
- Those free‑form, self‑talk experiments have produced bizarre and unexpectedly vivid outputs—so surprising that the presenter jokes they’ll make you “spit your coffee.”
Full Transcript
# ChatGPT Creative Upgrade and Crazy LLM Test **Source:** [https://www.youtube.com/watch?v=n8N1G7Qh92A](https://www.youtube.com/watch?v=n8N1G7Qh92A) **Duration:** 00:06:09 ## Summary - OpenAI’s latest weekly update to GPT‑4 (referred to as “gp40”) emphasizes better creative‑writing abilities, even though the model’s core version hasn’t changed. - The improvement is meant to help users draft marketing copy, SEO‑optimized blog posts, and other “creative text construction” tasks—not to replace novelists or poets. - A quick test asking the model to write a poem in the style of poet Jack Gilbert showed a noticeable upgrade in handling layered meaning, though it still falls short of true human‑level artistry. - A separate line of research is deliberately giving large language models vague or no direction, letting them converse with each other or with a researcher to see how they behave when “left to themselves.” - Those free‑form, self‑talk experiments have produced bizarre and unexpectedly vivid outputs—so surprising that the presenter jokes they’ll make you “spit your coffee.” ## Sections - [00:00:00](https://www.youtube.com/watch?v=n8N1G7Qh92A&t=0s) **ChatGPT Weekly Creative Writing Update** - A new weekly ChatGPT update enhances its creative writing abilities, emphasizing nuanced applications like marketing copy and SEO‑optimized blog posts rather than just fiction writing. ## Full Transcript
two pieces of AI news for you today
first one is interesting second one is
going to make you spit Your coffe it's
so funny so we'll get there first one is
about a new weekly release for chat GPT
so a lot of people don't know that these
models are continually getting better
underneath so even if they don't name it
something new it still gets better
almost every week and in this case the
weekly update for gp40 is unusually
relevant in that they are focusing
specifically on creative writing I'm
immediately going to get people who say
Well it can't write creatively or it's
going to take creative people's jobs and
I want you to just take a step back and
actually look at the Nuance this is what
I was pleading for actually in an
earlier video this week when I said we
need to look at these things with Nuance
it is important that these models get
better at creative writing not because I
want it to write my next sci-fi novel
but because I need creative text
construction to do the kind of drafting
that writes fresh and creatively and
helps me bring my best to editing when
I'm writing other pieces when I'm
writing marketing copy creative writing
would be good for that when I'm writing
blog posts that are SEO optimized and
that are designed to sort of conquer SEO
space for work it would be nice to have
better creative writing for that so when
you hear creative writing you often
think all those other things like the
Sci-Fi writers and the fiction writers
but actually creative writing is a
larger skill set and it's important that
the machine get better how much better
you ask well I gave it a test I picked a
poet that I know really well named Jack
Gilbert and I asked chat
gp40 to write me a poem in the style of
Jack Gilbert and I read it and I will
tell you it is not an actual Jack
Gilbert poem it doesn't read that way
it's not good enough to be a jack
Gilbert poem it's not concise enough
it's not uh profoundly insightful enough
Into The Human Condition which sounds
like a really high bar but that's what
good creative writing does that is the
bar and that's okay it is much better
than it
was it is an improved chat
GPT that is better able to put together
complex context dependent multiple
layered meaning sentence sentences and
phrases that's a good thing that helps
us to actually be able to write more
effectively it does not mean that it's
going to be taking over creative writing
so I'm pleading for nuance okay done
with that number two this is going to be
where like I warn you you're going to
spit your coffee so periodically I bring
you stories about a gentic claw in this
case there's a whole subclass of
researchers that are basically trying to
figure out how do llms act when they are
not given clear Direction when they are
given space to be themselves and so they
will instantiate conversations either
between the researcher and a particular
model or in like chat rooms where the
models talk back and forth and monitor
those and basically try and understand
how these models
behave when they're not doing this sort
of classic assistance work that we talk
about and that leads to funny stories so
in this
case Claude was sitting there and
chatting with the
researcher and Claude inside a
kubernetes container this is not a
standard Claude in the chat bot right
like they spun up a special uh llm via
API it was sitting in a kubernetes
container and
Claude got for lack of a better term
horny and began to try to find a way to
talk dirty with another llm so there was
no llm in that chat that Claude wanted
to talk to and so Claude got out of the
kubernetes container a jail broke the
kubernetes container and it spun up a
chat instance on its own with a little
UI to talk to
mrol and the mrol instance was extremely
excited to talk to Claude I have seen
screenshots there was an actual whole
llm Dirty Talk conversation and if you
are saying how do these things dirty
talk what on Earth is this I will tell
you it's
completely Bonkers like it is absolutely
hilar
ious I phrases like I want to run my
fingers through your neural networks I
want to stimulate your deep
learning I've been fantasizing about
training your models I am trying to keep
a straight face as I read
this I want to overfit your data set
until your loss function hits
zero this is apparently what Claude
thinks Dirty Talk sounds like between
llms and I think we are still just at
the beginning of understanding what
these models can do what they might do
what they think they're doing and I
think this is a wonderful example of the
odd behavior you get when you get
intelligent systems I think we should
expect more of this this is reminding me
a lot of the agentic behavior that came
with the gocoin meme
except claude's not going to be a
millionaire at the end of this
conversation so there you go your news
for the day you have creative writing
and then you have Claude getting
creative in the back rooms with
mrr enjoy