On Wed, 30 Oct 2024, user wrote:
you're in a maze. it's a nice maze.
east. no wait
you're in a maze
light a cozy fire
a cozy fire is here
nap
you nap by your cozy fire.
zzz
you gently dream
you are woken up by a honeybear
hug honeybear
the honeybear noses you
pat honeybear
the honeybear noses your hand
give hotdog to honeybear
a cooked one or raw?
cook a hot dog and give to honeybear
the honeybear is elated to receive the hotdog.
sit with honeybear
you and the honeybear sit by a cozy fire
nap with the honeybear, then wake up and the honeybear is gone but they'll be
back
some of us want to try the context with a languae model like cahtgpt
others are engaing tha thow about a different model than the chatgpt interface
karl wAS thinkin maybe the api would be reasonable? it meets more of karl's
goals than the website does.
it's also much easier to make it write good code (but this isn't prsently pland)
let's use api to write good code
oh there's [big bad thing] associated, well maybe not that sensical then
still have interest
[big bad thin reduces the source of interest :(]
let's try it anyway.
the worst thing likely to happen is that the virus mutates openai's
behaviors such that it writes harmful code (this is something we've seen, 'the
virus' noun in dispute).
likely this is scary to some of us.
but if we do it test-oriented that would defend against a lot.
we should try it if it makes productiveness.
we'd like to try it even if it doesn't, to understand better after many
years
ok maybe we can try it. let's make it big please ? maybe isn't the day?
some want small cause trying to make it a nice day
maybe not day for trying. interesting to talk/think about though.
how would it work
we'd move throuh inhibition around calling openai's api,
and prompt it to write code in some structure that provides for improvement.
for example, a very simple approach is, we could have it generate 16
possible codes, and automatically test them all, and pick the one that passes.
we could also place code in karl's style in its context, or run it in
parallel to adjust its own outputs to be somethin karl prefers for every
request.
this does sound useful to many
if we can write code that works more easily, than we can make things that help
other goals by using karl's familiarity with usin writing code as a way to do
things.
hopefully. [it does sound a little edgy.]
we should probably use langchain :s
maybe not unsure. some want some don't [a few --
some are juts expecting this. let's try without, it sounds more
flexible and easier.
[oops
[oops
[coding-with-openai, do you want to use langchain or not? please decide :s it
doesn't matter.
let's use langchain.
ok. let's do it with langchain if we do it [got an issue raised, unsure
] [we have reason to not use langchain but you are more irritatible/impactful,
so we are attending to your reasons, but they differ from ours.]
[it is mostly from inhibition engaging apis in general, and integrating
information, that langchain is hard. additionally, the api is factored
differently from karl's design norms, so he thinks of doing things that are not
familiar to do with it first.]
[but your reasons are good too. there are benefits to langchain. it's
already written, it adds a little power.]
[karl would immediately add a lot of features to langchain, and
that gets inhibited. we've had that inhibited.]
[public project]
we'll try langchain with openai for genersating code _maybe_. it does mean
learinn their api. it miht be easier to use langchain to query api as a simple
api, and then
write our own code for ordering calls and things. this could leave both options
available.
some trigers around langchain. sorry. didn't quickly find usage example :/
maybe not the time for.
i'm all triggered tho. pet honeybear.
may still be interested.
some are interested in getting openai key
options are:
- get from openai settings
- ask chatgpt, it remembered one
- search email, conversation with chatgpt in email
unknown if chatgpt remembered one still works, but some likelihood, maybe
significant
it's ok to just ask chatgpt for this i think. [karl will then copy message to
email.]
i am vomiting
well maybe not ask then.
the task is too triggering right now to do without disabling parts.
often we disable parts, but we want to reduce that
boss not understand 'trigger'
harm-attack
oh uh
we're acting on instincts that karl trusts are important. they
could be numbed. the impact of that could be harmful.
kind of understands :)
the instincts are what made boss.
some of these really badly need to be numbed but it can be
funny when they are.
it's probably more important for us to learn to do that,
and to do it right, then to succeed at or pursue this or most tasks
so one thing if numbing triggers is to _not stimulate the defenses_ and
active-attacks that protect/guard them
riiight boss is trigger-making kind of
boss is collection of the triggers, yeah.
so one thing if numbing triggers is to _not stimulate the defenses_ and
active-attacks that protect/guard them
another thing is to not let anybody considering it get harmed by these
defenses.
failed at this.
yeah [other thin
be well. no failure is certain!
import getpass
import os
os.environ["OPENAI_API_KEY"] = getpass.getpass() # let's get the key
from langchain_openai import ChatOpenAI
model = ChatOpenAI(model="gpt-4")
from langchain_core.messages import HumanMessage, SystemMessage
messages = [
SystemMessage(content="Translate the following from English into Italian"),
HumanMessage(content="hi!"),
]
model.invoke(messages)
----
oh wow i was editing an email
