Reflection AI’s Misha Laskin on the AlphaGo Moment for LLMs | Training Data

[@squamish4244]

His comment on where we are in the late 1800s is on point. That was during the Second Industrial Revolution. We're in the beginning stages of the Third Industrial Revolution, which will happen a hell of a lot faster and with much larger effects than even that one did.

[@GNARGNARHEAD]

this actually gets pretty good, really excited to see what Reinforcement Learning can do for second stage training on Language Models

[@stephaniezhan2610]

Awesome to hear. Thank you!

[@fintech1378]

Maybe 'in-context learning from human feedback' will unlock new capabilities too, langchain is working something for that?

[@squamish4244]

And people who don't know enough about AI wonder why the next models are taking several years and not *one* year. It's because they're working on next-level stuff, people! It takes time!

[@BrianMosleyUK]

1:03:23 great advice, and one which resonates strongly with me. Great interview, thanks everyone. New subscriber.

[@stephaniezhan2610]

Excited to hear. Thank you!

[@odiseezall]

The audience can make up its mind about the mind state of people that say "safety is reliability" and "AGI will do boring tasks for you on your computer".

[@ansha2221]

Excellent podcast.

[@dinarwali386]

Fascinating. How is accuracy measured and maintained with respect to AI agents over time?

[@richoffks]

It’s not simple enough to answer in a comment bro 😭 they have a bunch of metrics that are basically useless because the llms keep surpassing them but then there are human evaluation metrics as well, for example, how much sense it makes to a human.

[@stephaniezhan2610]

I actually think this is an interesting area of opportunity. I think SWE-Bench is the best benchmark for testing real-world agentic capabilities for programming, amongst those that exist today. But we need more and better benchmarks!

[@jordanmiller11]

Note that his evidence for agents being close is they haven’t run all the tests on LLM models yet. Not that they have any verifiable evidence. Not promising.

[@wwkk4964]

What if reasoning is an illusion and the ground truth is just specialized associative memory of what worked.

[@Gnaritas42]

Because it's not, type 2 thinking is real, type 1 thinking is what you're talking about, you can build a 2 from a 1 with some loops and more architecture. 1 comes first, 2 is an upgrade. 2 requires something 1 doesn't do, unbounded compute, aka the ability to think longer, because you're planning internally, internal thinking, and comparing your plans to choose the better one, and using logic and induction and deduction to work through them. This is what humans do, well documented, not an illusion, just a different mode.

[@wwkk4964]

@@Gnaritas42 perhaps declaring axioms as truths and reverse engineering data to present in a book filled with confirmation bias is not exclusively Kahneman's achievement. This appears to be your type 1 thinking it's type 2 (both are are illusory )

[@Gnaritas42]

@@wwkk4964 oh god you're one of those nutjobs, no thanks. And no, they're not illusory, you're just not sane.

[@ozgurgulerx]

Isn't agency more about the cognitive architecture you build on top of the LLM? e.g. the recent LATS , GoT, STAR papers, Q* etc...Nothing about that here...He seems to be more interested training his agency LLM which may not be the right direction...Not much meat overall for agency unfortunately.

[@ManasSharma-e4m]

I am in first year i am worried about my future

[@ParnianMotamedi]

is the sequoi digital currency? it listed in pancake swap? Please answer me soon i'm in hurry

[@alxfazio]

You've got top-tier microphones, so why use AI filters that make your podcasts sound artificial? Let your natural voices shine through!

[@420_gunna]

pink jacket smiling like a bloodless jackal the entire time and rocking in her seat Chill we get it

[@user-yo6vy9lx2g]

Super super podcast

[@jbraunschweiger]

I don’t think reflection will exist in a year

[@gotemlearning]

Why not?

[@swapanjain892]

Troll

[@jbraunschweiger]

@@gotemlearning no moat. The only reason the big players haven’t done this yet is there is not much value to most people’s writing style.

[@gotemlearning]

@@jbraunschweiger writing style? what does that have to do with it? I do sort of agree about the moat -- it's hard to find one in the hottest pursuit of the 21st century -- but they might find a niche.

[@mattwesney]

This is painful