colin-fraser.net
Driven by industry progress, inspired by provocative leadership, plus don't mind a good pair of shoes or a great @PennStateFball scoreboard either.
1,108 posts
1,675 followers
134 following
Regular Contributor
Active Commenter
comment in response to
post
The presentation that they spontaneously decided to communicate by touch tones because it’s a better way for AIs to talk is fake yeah
comment in response to
post
with one exception
comment in response to
post
honestly I'm surprised they gave him the chance to delete it
comment in response to
post
there is so much psychology in these four paragraphs
comment in response to
post
or I mean, do, but know that they might get mad about it!
comment in response to
post
yeah, lots of psychology going on in all directions here
comment in response to
post
just don't tweet about your job!
comment in response to
post
you gotta just delete the tweet
comment in response to
post
donald = {"personality": ["dark", "sarcastic", "flirtatious"], "situation": ["haunted by ghosts of past", "in need of distraction"]}
comment in response to
post
I think that’s true, but I also think this is a bit different. The Deep Research product is supposed to be able to search the Internet and synthesize information from multiple sources and so on. It ought to be able to just accurately report the facts as they are presented in the sources.
comment in response to
post
My favourite part is the output is it says RJ Barrett (who was traded from the Knicks to the Raptors over a year ago) started on both the Knicks and also the Raptors, and to account for this apparent inconsistency it bafflingly claims that Barrett is “on loan” for the Raptors
comment in response to
post
My favourite part is the output is it says RJ Barrett (who was traded from the Knicks to the Raptors over a year ago) started on both the Knicks and also the Raptors, and to account for this apparent inconsistency it bafflingly claims that Barrett is “on loan” for the Raptors
comment in response to
post
I basically have no idea what you’re supposed to do with it. It doesn’t seem useful to me.
comment in response to
post
Yeah. Getting it to do what you want by getting it pumped up
comment in response to
post
comment in response to
post
I did not tell it to add these little sections at the end
comment in response to
post
This is my fault for being problematic and generalizing over all rich white woman executives with brown hair but I thought she'd be bringing more of a Sheryl energy to the operation
comment in response to
post
www.cbsnews.com/news/trump-a...
comment in response to
post
I just read through the paper that is cited as the model that actually extracts the Big 5 and it turns out it's trained exclusively on Russian selfies www.nature.com/articles/s41...
comment in response to
post
if we just draw the rest of the fucking owl then we'll have AI software engineers etc
comment in response to
post
related: bsky.app/profile/did:...
comment in response to
post
does he think that the reason it's "universally accepted as true" that there is no largest prime number is that the CEO of math said so?
comment in response to
post
I don't come on here to see my work emails
comment in response to
post
You may need to acquire some negative beliefs
comment in response to
post
Yeah, for years they’ve put up with these annoying demands because the site can’t just reliability engineer itself, but we are just a year or two away from the site reliability engineering itself so fuck it, let them quit
comment in response to
post
Yeah I mean that’s basically what RLHF is for, right? It’s a way of introducing the goal of pleasing the raters into the loss function
comment in response to
post
Well, I’m not sure about that. Obviously it depends on how you define goals. But as I get to later in the thread, I’m perfectly happy to say that Dijkstra’s algorithm “has a goal” of finding the shortest path in a graph, for example.
comment in response to
post
I’ve written about a lot of these same ideas on medium medium.com/@colin.frase...
comment in response to
post
Ok well, again, the method in this paper uses human raters extensively
comment in response to
post
Note as well that you are mistaken about the point of this paper. It’s not a description of what they do with Claude in production. That’s a trade secret; they’re not going to publish that. It’s a proof of concept for an experimental post-training method. No one knows if and how much they use it.
comment in response to
post
It also says very explicitly in the paper that they’re not even trying to eliminate raters
“Although here we largely eliminate direct human supervision for harmlessness, rather than removing human supervision, in the longer term our goal is to make human supervision as efficacious as possible.”
comment in response to
post
They train a “harmless model” using AI-only ratings, but still use human labels for “helpfulness”. As you’ll notice if you read all the way to the end of the abstract, this allows them to accomplish the same goal as RLHF “with far fewer human labels”
comment in response to
post
Duly noted