Frosst: Fine-Tuned LLMs Generalize to Unseen Request Types
Nick Frosst · He built Cohere into a $5.5B AI startup; How to Win in AI; & Why LLMs won't lead to AGI. | Nick F... · PMF Show · Sep 9, 2024 · at 34:55
Nick Frosst, co-founder of Cohere, explains why instruction fine-tuning was unexpectedly effective and how models generalize to tasks not present in feedback data.
0:00 / 0:11exact quote · 11.5s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“If you train it on just those conversations, and you It's never been asked to do a summary. It will still be a, do a good job when you ask it to do a summary. So it generalizes to request is never asked, never been asked in the training data.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Nick Frosst
Frosst: Meta's Llama is a corporate giveaway for notoriety, not open source
“Lama, Lama three, like the stuff Meta is doing great models, great engineers, really cool stuff. But it's not open source the way Linux is open source. It's not open source the way, you know, I don't know, like Wikipedia's contributions are open source in that…”
Frosst: Large language models are not a clear path to AGI
“I don't think
It's a clear path to AGI.
I don't think we're going to be making digital gods anytime soon.”
Frosst: Creating human-like AGI is not an obvious human need
“It's not obvious to me that we want AGI. I don't wake up in the morning and say, gee, I wish my computer was a person. I wake up in the morning and I say, gee, I wish I could do my work faster. Like, gee, I wish I didn't have to do this boring thing. Gee, I wi…”
Frosst: Cohere is not chasing AGI, consumer bots, or public benchmarks
“So we're not interested in making like a public facing bot for consumers to talk to. We're not interested in chasing AGI. We're not interested in a lot of, there's a lot of like weird public benchmarks that exist trying to measure, you know, how close we are t…”
Frosst: AI enterprise value is in deployment and workflow, not weights
“When I think about Koshier's value, like it's not in, you know, the weights of any particular model. In fact, we we've released our model weights for researchers as well. So if researchers want to use our model, like they can download the weights from hugging …”
Frosst: AI agents will augment workers rather than replace full roles
“I think when you look at the work people do, there's actually a lot of edge cases they encounter all the time, and there's lots of things that they do that could not be done just on the world of the, just on text, just on the world of computers. They're actual…”