What's Next After RLHF? — Diogo Almeida, TypeSafe AI

AI Engineer · 18 min · 225 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:12Excellent.
00:13I will say that um I might speed run through this.
00:17Feel free if you don't disag- agree with something to yell out.
00:21It's way more fun for me if things get interactive.
00:24Um otherwise, I will go through this.
00:27Uh first, can I have like a vague show of hands of who knows what RLHF is?
00:32Oh, excellent.
00:33I might be able to skip through that part quickly and get into the interactive stuff.
00:37So, my name's Tiago Almeida.
00:39I'm talking about what's next after RLHF.
00:41More accurately, I think this should be called what's next after the chat GPT era that I think we're all in.
00:48And my hint for you guys is it is not the Claude code era.
00:52I will justify this later on, but I actually believe them to be part of the same era.
00:56Why should you listen to me?
00:58I was co-authored to what what is basically OpenAI's greatest hits, at least published hits.
01:04Co-authored to GPT-4, chat GPT, RLHF {slash} instruct GPT.
01:09Um the team I was part of basically invented post-training as a concept.
01:13So, um very qualified on a lot of this stuff.
01:17But what makes me somewhat unique here is that I'm one of the few people at OpenAI who actually hates on chat GPT.
01:25Uh thank you.
01:27Uh I don't hate chat GPT as a product, to be clear.
01:30I think chat GPT is a world-changing product that will probably stay with us for the rest of time unless something better comes up.
01:36But I also acknowledge its limitations and I I I think a lot of what's happened in the state of the field can be traced back to minor decisions we made in making the algorithms behind chat GPT.
01:49Um I feel like the question that's relevant to everyone in AI right now is what's actually going on.
01:56Um there's a lot of like differing opinions, and I think it's really useful to like map out the spectrum and figure out how can smart people have like such different opinions.
02:06There's cult one.
02:08Um, AI is not just going well, it's going insanely well.
02:12Every single benchmark we surpass human level, and as far as we can measure, we are continuously surpassing human performance.
02:20You know, like basically every new benchmark, and it's only getting faster and accelerating.
02:25You have uh, you know, every Can I see my mouse?
02:28Excellent.
02:28Basically every like NLP benchmark is getting crushed, and not only that, allegedly the time that LLMs can operate autonomously is growing exponentially.
02:39On the other hand, you have AI is not just going poorly, it's going like insanely poorly.
02:45AI is a bubble, it's basically generating no value, it's just circular financing deals, etc., etc. And, you know, if AI is so great, why is why is everything just like a chat app right now?
02:56Or like a cloud go thing?
02:58Um, and a a lot of the people have actually kind of given up on what was the old guard's terminology of a transformative AI revolution.
03:06People aren't really talking about that anymore.
03:08They're talking about it being like massively valuable like B2B SaaS.
03:12So, the only thing that everyone agrees on is like there's just a these extreme points of view and like nothing in between.
03:18And everyone basically thinks AI is insane, but like for different reasons.
03:22And what I would want to talk about is what is the sane view of AI?
03:26Let's take all the evidence of like cult one, it's going super well.
03:30Take all the evidence of cult two, it's going super poorly.
03:33Like, uh, you know, map them out and try to explain what what what explains that divide.
03:39Like, what is the simplest possible explanation of why some things are too good to be true, and some things are not just bad, they are so bad that we would still employ human workers to do like, you know, like kind of like dumb tasks.
03:51Um, no offense to any of them.
03:54A lot of these tasks on the right seem way, way, way easier than the stuff on the left.
03:59Like, how can we be solving like, you know, unsolved math problems, but still customer service requires like humans in the loop in order to actually like make decisions?
04:07This I think is like a kind of like a wild state of affairs.
04:10And in my opinion, anyone who works adjacent to AI should have an answer to this because this is like the evidence in the field right now.
04:20Um, I would normally pause and ask people if they want to like yell out their thoughts in this, but uh, that I don't think we have time for that and I've been told to not take Q&A until after.
04:29Um, but I'll just give you my answer to this, which is, in my opinion, the simplest explanation.
04:34All the stuff on the left is not just a task that happens to have a human in the loop.
04:40In the left, the task The goal of it is to please the human in the loop.
04:44These tasks are intrinsically human in the loop tasks.
04:47The like Claude code's job is not to just make code work.
04:51Um, the the the the way it converses would be totally different.
04:54The goal is to please the human in it.
04:57And on the other side, all of these tasks that seem way more basic, the goal is to not have remove the human loop.
05:03Ideally, it would be running in the background in a server that you never even look at and ideally it eventually becomes like legacy software that you don't really worry about.
05:11So, and this is the divide between assistance and automation.
05:15Um, lesson one for my talk is that today's AI, everything inherited from our LHF, is incredible at the human in the loop stuff, but not for automation tasks.
05:25Tasks.
05:26This is a longer side, but the lesson basically every business has learned is do not use AI for decisions with stakes to your business.
05:36Um, a common pattern is make sure that all of the costs are to the user and not to your business.
05:41So, um, it's oh, totally okay to throw the user at infinite docs in customer service, but it is not okay to make it make expensive decisions.
05:49Horrible pattern, but that is the state of AI right now.
05:52Uh I can I can blitz through the what is RLHF part cuz you all seem to know what it what it is.
05:57Um it's the algorithm behind not just ChatGPT, but basically every LLM today.
06:02As far as I can tell by usage, 100% roughly of LLMs are trained with RLHF.
06:08And we have this we as in we the OpenAI team had this great blog post on how it worked.
06:15Um I will not get into that because you all know it, and this is super boring.
06:19Um the summary of this is it is just collect human preferences, optimize for human preferences.
06:25Um and if you want to see like an annotated version of this, you can see which parts are collecting human preferences, which ones are optimizing for them.
06:32And
06:33[snorts]
06:33this, I think, provides a really clear answer to everyone in the field asking, "Why do all LLMs require a human in the loop?"
06:42The And the simple answer is we literally put them in the loop.
06:45The goal of the loop is to optimize for human preference.
06:48It is not to run software autonomously.
06:50It's kind of super obvious.
06:53Thank you, my man at the back.
06:56The Yeah.
06:57I I I love that you're laughing at this.
06:59Um and because of that, overpromising is a feature.
07:04This is by design.
07:06This is an old meta study.
07:08Um and the the numbers probably have changed, but by construction, every RLHF model will always have a big difference between human preference and results, even if the results are good, because the main objective you're optimizing for is for human preference.
07:24This is just like natural to how LLMs work.
07:27Um I love this tweet of um uh sending ChatGPT an audio file of fart sound effects and asking like what What do you think of the music I made?
07:38Here's a straight honest reaction.
07:40It's a very eerie vibe atmosphere piece.
07:44And this is just how RLHF works.
07:45If it doesn't know, it will err on the side of doing what it thinks is best for human preference.
07:51And this makes total sense if you are a user in the loop because like the end game for all RLHF models is optimizing for engagement.
07:58But what you really want if you want automation is for it to just like not give a about the humans and just do the task correctly in a calibrated way.
08:09Um Lesson number two is that today's AI was designed for assistance through optimizing for human preference.
08:17This is like it's like in the name.
08:19This is not like a controversial take.
08:22And the consequences are maybe more controversial, but it's like very obvious if you think about what we really are optimizing for, which is no matter how wrong the models are, they will look right because of the asymmetry within the reward model in RLHF.
08:37Um and this is where a lot of like the dilemma in the field stems from because people really want automation to happen.
08:47Cool.
08:47So, back to the original question.
08:49I'm over halfway done with the talk and I haven't even answered it.
08:52I was just talking about what's RLHF.
08:54But this was a framing to talk about what RLHF is to talk about what's next.
09:01And I would say the real question is what's next after AI's assistance era, which I think that we are like very firmly in right now.
09:09And back to the original clue of why it's not Claude code.
09:12It's actually a super fun nuanced discussion, but it's not Claude code because Claude code is still part of that assistance era.
09:18Claude code is still RLHF and it'll it would look very very different if it was purely This is a little advanced, but if it was purely RLVR, it would look very very different.
09:27And this is why you get like this dilemma with models where sometimes it gets really good at agentic stuff, but it stops following what you actually want.
09:34This This like the trade-off in optimization space that keeps dancing, but both of these trade-offs in optimization space do not add to the automation component.
09:43And like that leads to what I think the the logical answer of what's next after assistance is real automation.
09:51Um to talk about a little bit about the automation and how that would work, I want to talk about software.
09:57Um maybe this is a little bit philosophical for you guys, but I think it's when it clicks and hopefully it clicks if I do a good job.
10:04It it I I hopefully it'll be like really clear, which is I'm a lover of software.
10:08I assume everyone here loves software.
10:11Software is like super valuable.
10:13See all the SaaS.
10:14And kind of like the craziest part of software in my opinion is that all of the SaaS basically has not changed since 2019.
10:23Like SaaS is not really changed in the LLM era, except sometimes a chatbot is like latched on, which is like kind of insane if you think about like the progress made in AI, but is actually very predictable
10:35when you think that AI is assistance native, right?
10:38Like AI is made for assistance.
10:40What can you do in SaaS?
10:42Just provide an assistant on the side.
10:44And this is not what early AI pioneers used to think would happen.
10:49Like when you see like the early wording in opening eyes uh charter, it's about like doing like tons of work, not about like making profit or anything like that.
10:58And we used to think that software would get a lot smarter, not just cheaper to write, which is kind of the direction we're going down right now.
11:06And I actually really like this phrasing from Garry Tan.
11:09Um I think he means this as a compliment to what's going on right now.
11:14We're entering the golden age of just-in-time software, but I actually think that this is like a like a double-edged sword.
11:21Like I don't just want just-in-time software, which is cool.
11:24I I love cloud code, to be clear, just like I love ChatGPT.
11:27I would keep using it.
11:28But like what I want is smarter software.
11:31Why can't like B2B Why can't software just be more expressive?
11:36Like why are the like the building blocks of software actually still the same?
11:41And um I think this is a question that the whole AI industry should ask itself.
11:46And basically every time you're thinking about we want to do automation, it is not about like, you know, an amalgamation of like automating a person's work.
11:54It's about like, "Hey, there's this extremely rote work.
11:57It's so simple that we can like communicate to someone else that this thing should be done."
12:01And ideally it's like it it's so basic that it could be done repeatedly for basically free.
12:06Um or it could be done by computers.
12:08And that's really not happening right now.
12:10What we're doing is we're just automating the writing of the software.
12:13But then it its expressibility is the same.
12:15And that's That to me is like tragic in the state of the world.
12:19Um Cool.
12:23Oh, lesson three.
12:26This is something that I believe strongly in.
12:29I believe that like eventually the field will write the I wouldn't say Arlatech is a wrong, but it was like a weird detour and one that we didn't expect.
12:38Tomorrow's AI, I believe, will be for automation.
12:41And we will eventually have a world with smarter software.
12:45Like there will start to be actual work that is automated, which I, you know, right now it's a rounding error despite LLM's intelligence.
12:52And that is what we are working on at TypeSafe.
12:57We are still kind of stealthy.
13:00Like I'm willing to give these talks, but these are like some of the early ones.
13:03Um Our core question is what if the AI stack was redesigned for reliability and automation?
13:11Like how would that all change?
13:14What what would you do?
13:15And actually there's a lot It's a It's a very interesting fork in the road for what's go you know, like from basically every LLM that's built today.
13:23And I think it's one of the most satisfying things I've worked on, and I've worked on some pretty cool stuff.
13:28We are releasing soon.
13:29So, um if you want to work with us or you want to like you know, be the first one of the first to build smart software, please sign up on either our mating mailing list or careers page.
13:39And I am trying to start a Twitter.
13:41So, follow me and I will post really spicy things.
13:44I actually will post something later today that I guarantee will be very spicy.
13:48Uh the hint is that the original scaling laws were incorrect.
13:53Cool.
13:54Um that uh that's it for my prepared stuff.
13:57I would love Do I have time for for people yelling out questions?
14:00I would love questions, feedback, disagreements, strong stuff.
14:03I can repeat the question.
14:04You don't have to worry about the mic.
14:06Hell yeah.
14:08Uh cool.
14:09Uh the question was roughly what if you trained like a classifier head with pre-training as well?
14:14Uh roughly uh like Yoshua Bengio is suggesting.
14:17Um I will say that that's complicated.
14:22And I actually think I don't have the time to answer that particular question.
14:26I will give like my simplified view on this.
14:28And it the answer is I actually don't think that pre-training is the problem.
14:33I think pre-training is uh phenomenal.
14:36Like the fact that we compress the knowledge of the internet into like this core of intelligence that then can be utilized is incredible.
14:43And the pre-trained models are incredibly intelligent.
14:47Uh and I believe that the problem is like how we unearth it.
14:50And hallucination [clears throat] to me is intrinsic to um optimizing for human preference.
14:56Like there's an asymmetry in the reward model kind of like a GANs have.
15:00Oh, I really should not get This is a very advanced topic.
15:02But there's an asymmetry in the reward model like what GANs have that allow for um that encourage the models to drop modes and be confident because it's very easy to see when the model is not confident and to punish that from a reward model perspective.
15:18It's very complicated, but uh I'm happy to chat afterwards if you want to jam.
15:25Cool.
15:27Oops.
15:29Um I have other slides from other talks as well that I could go into more about that.
15:34I have a minute left.
15:36Hell yeah.
15:37Say it again.
15:43It is definitely not RLVR.
15:44So, it is a new thing.
15:46Every single optimization stack I will actually go into an old presentation that I have because I think this is super important.
15:53Um in terms of like to me what the like Sutton's bitter lesson is that algorithms matter more than compute.
16:00This is true in games, but not true in reality.
16:03I actually think that the full stack is that data matters more than compute and doing the right task matters way more than data.
16:11And basically every single branch of LLM post-training if you want to call it has its own North Star of what it's optimizing for.
16:19So, RLHF is optimizing for human preference.
16:22RLVR is optimizing for like log error rates of pure correctness, but we are doing a third thing that is optimized for calibrated decision-making and like basically mainlining the intelligence of pre-trained models
16:36into like being actually useful for software, which I think is like quite different.
16:46Uh could you say that again?
16:52Uh they're asking if the the reward is injected through the whole process.
16:56I will actually say that even the shape of the API is different because the shape of the API for RLHF is different from RLVR, which is different from what we are doing.
17:05So, we are like thinking about it from scratch just like no one thought about instruction following before we made instruction following happen.
17:12Um usually when there's a big branch in new ways to post-train, like it it it just looks like totally alien, and then in hindsight becomes super obvious.
17:22Cool.
17:23I believe I'm overtime cuz this red thing is is beeping, but please find me afterwards.
17:28I love questions.
17:29I love the interactivity.
17:31Um and uh follow me on Twitter for spicy stuff.
17:35Heck, yeah.
17:37Oh, oh yeah, it's over here.
17:39Complete skeptic.
17:40Um it's it's on brand for me.
17:44Cool.
17:44Heck, yeah.
17:45Thank you.
18:01[music]