I love you so jevving much

The recording opens as the speaker begins; the MC's introduction is not captured. The talk ran past its slot (14:05–14:23): the recording runs 14:07–14:30.

Transcribed locally with Qwen3-ASR 1.7B (timings per ~5 s segment), prompted with the speaker names and affiliations of overlapping sessions. "Diogo Almeida" at 00:00:03 and 00:00:14 and "TypeSafe" at 00:19:42 may come from that prompt: the Voice Memos transcript has "Diego Almeda" and "type safe" there. Other names are kept as transcribed: the model is "Jeff", "Jeb", "JEP", "Jetting" and "Java", RLHF is also "R LHF" and "Arlo", and the KV cache is also "KB cache" and "KDC". The Voice Memos transcript is kept beside it as transcript.txt.

00:00:00 — Introduction

All right. Give me a sec to make sure this is working. Heck yeah, I'm Diogo Almeida, and this talk is I love you so, Jetting much. Glass animals reference. We're not just weird. Like I said, I'm Diogo Almeida. I'm known as the Dev Guy, sometimes Devy Pinkman. Also co-created ChatGPT, RLHF, and InstructGPT, but. I assume that for the crowd today, y'all know who I am at least mostly.

I have like a bunch of AIE specific stuff I want to get to, but I like going into my favorite topics in AI.

00:00:37 — Where is all the automation?

First one is, what the fuck is actually going on, and where is all the automation? If you've heard me say this before, I'm sorry, but it's such an important thing to beat into the field. Because people still do not get it, and I need to yell this until. Everyone really understands what the fuck is going on, and if you're in AI, you should have a really good answer about why.

AI can do so many awesome things, and yet still be so useless for a bunch of other ones. My relationship to ChatGPT is. Somewhat unique. One of the few people who hates on ChatGPT, who made it. I don't hate ChatGPT. To be clear, I love using agents. I love using coding agents. I love chatbots. They're just quite limited, and in particular, I think it's a little tragic.

How much they've bifurcated the field into overpromise, underdeliver, really, really terrible stuff for the world, in my opinion.

00:01:26 — The two cults of AI

And I like to talk about like the two cults in AI and how we can resolve the cognitive dissonance. That everyone I think should be having, unless they have a very good answer. There's cult one, which is you know the cult of like EA's. They look at all the benchmarks and they say, like, not only are LLMs. Surpassing human level on basically every benchmark, they have been for two years.

We've been announcing AGI three months away for two years now. Even since like the O3 days, do people remember that? No one even care cares about that model anymore. It was that was like peak internal to OpenAI AI hype. And on the other hand, you look at the non benchmarks, and you look at. The reality, and you see the rest of the world, especially if you guys go outside and leave San Francisco.

A lot of the world is like really unchanged, despite AI being. So fucking cool, and I think AI is transformative. It's a rare. Point of view to actually think AI is cool and transformative, and also like underwhelming the world right now. And the tricky part is, this is the same field, and we are all seeing the same facts. So, what gives?

00:02:29 — Why is AI jagged?

You know why is AI jagged, so to speak? Like why can it be so good at some things and so not good at other things? And what is the Occam's Razor explanation for that? I will be blasting through my slides fast. Sorry, I need to slow down for people taking pictures. But I'm noticing that this is gonna go fast. Is all I'm saying. So I think that.

We as a field should have an Occam's razor explanation for why are there some things that are just. Not just good; they are unbelievably good. Superhuman, like Millennium Prize problems in math. But on the other hand, they're not just bad. They're so useless. They can't even be used to touch a counting. That sounds way easier than million prize problems in math, right? I could do the latter. I definitely could not do the former, and I was a mathlete.

So, what gives? What is the reason for this? And I normally love this as an interactive thing, but I will instead just tell you my answer. Sorry for the power dynamic.

00:03:24 — Human in the loop, by design

Everything on the left side is intrinsically a human-in-the-loop task. Like deep research, I could not do better than deep research. The models are so fucking good at that. But if gun to my head or gun to your head, you had to pick between me and Deep Research. And the answer had to be correct, not well written. You probably pick me every time. And that's the distinction between like optimizing for automation, which tends to be a lot easier and a lot simpler.

Yet AI keeps failing this. Every year, there's like an infamous failure of AI failing to even automate drive-throughs. And that should be kind of shocking to people. And my claim is actually for these tests pleasing a human. Is nothing to do with doing the task. What you want to do is basically be. An abstraction away. You want the decision to be made, and you want to get back into doing automation.

And code is the language of automation. Therefore, you need to integrate with. This is a lot, but I will summarize again.

00:04:15 — RLHF and the assistance trap

I get the irony. That you know, I co-created RLHF, and I do believe RLHF is part of the problem. And Arlach Jeff in particular is why all AI today does require a human in the loop. With an asterisk, unless the task is like very trivially verifiable and can run on its own. But still requires a human in the loop to verify afterwards. It is because we literally put it in.

That is what the HF is in RLHF. And my claim is that there's a better way, you know. My summary of the state is 100% of LMs today. Are optimized for assistance. You might talk about RLVR, but the RLVR models are very bad to interact with if they've not been RLHF. The result is LMs become too good to be true at assistance. They're just so fucking good. Like you should never pick me over ChatGPT for your ChatGPT queries.

But for actually doing like. You know, run in the background cognitive work. You should not trust LMs at all. Because they've not been optimized for that, which is why they're so bad at the tasks of automation. And I would also guess that automation is a lot more valuable than. Because you can do way cooler, deeper. I'm going to get into this later, but I think there's going to be so much stuff that we can do.

And we are just scratching the surface of that right now, and that's the tragic thing about AI, the state of AI.

00:05:30 — Rethinking how AI is done

Let's talk about the evolution. I can't tell if this is appropriate, but it's too late. Oh God! Don't take pictures of this one. Cool. Oh, getting a thumbs up. Yeah, excellent. So we instead of doing RLHF, we just have to do something else. We need to think from first principles. How do we optimize for machine native AI? You guys hopefully know the interface of JEP, and that is what we are doing. More important than optimizing for its interface.

We are rethinking how AI should be done because this is what we did again for Arlo.

00:06:05 — The bitterest lesson: doing the right task

For Sutton's bitter lesson, I'm going to loosely paraphrase to make my point easier to say, is that. Scaling and compute matters more than output. This is why people become like scaling filled gheronies. And I would argue that there's an even bitterer lesson than that. Which is data always matters more than compute. We are learning that quite a bit in the field now, but it has been obvious.

Since the beginning of time, I'll tell you that. This is the one you should be photographing. I would argue that the bitterest lesson is doing the right task. And doing the right task is like a weird thing to think about, especially for ML people, because they like algorithms. They want number goes down or up or whichever is good direction. They don't want to think about the real world.

But this is always how you make gigantic gains in performance. This is how the big breakthroughs happen. Hundred percent of the time, it comes from this. And this is true for classical RL games. This is why Sutton had his bitter lesson. Which is data is derived by compute when you're doing rollouts. Like this. Why would you need off-policy data when playing chess or something like that? And doing the right task is defined by the rules. So actually, when you're doing RL, you don't even need to think about the right task.

For pre-training, we actually, as a field, had to do a lot of thinking about the right task. I don't know if people remember the birthdays with masked language modeling. But it took a while, and it was very unclear to figure out that the next token prediction was super fucking cool. Not because it was better at mathematical modeling, you know the objectives on their own aren't special. It's about what you can do with them afterwards.

And it turns out, next token prediction is incredible at generating strings, which is what these pre-trained models were so good at. So no level of mass language modeling, even if you had today's compute and algorithms, would make it as interesting as GPT-3 was. Because it is just doing a different task, and. Data, you know, the internet is obviously a giant bottleneck. Now we're like, you know, burning books or whatever in order to get more data, but.

Data obviously matters a lot more than computing. People keep like rediscovering this and figuring this out. It's extra true for making LLMs useful.

00:08:04 — What instruction following showed

Before R LHF even existed, the data for R LHF didn't exist, and you could. Like it is night and day. I was doing the next slide. This is the probably the relevant part. This is what this was the work in the RL stuff they were. And the high level was that even with the dumbest algorithms in the world, even with a GPT-2 sized model. You got something like GPT nine level performance, like real GPT nine, not today's like fake GPT numbers.

Like the numbers used to mean like oh, when hundred x scaling and effective compute. For pretraining, and they've given up on that because it just can't scale anymore. And still, even with that, and you take something like GPT seven to nine. To match GPT 2, the model from like 2018 or whatever it was, at the task of instruction following. That is kind of wild because you think big LMs are so important for this, and they are, but that is not the most important thing.

To even the gigantic jump of doing the right.

00:08:57 — Reliable intelligence

And spoilers. This is what we did for you, right? Like we made this gigantic leap in the field that no one could realize could be coming. Because we were optimizing for a task, and very few people think about the right task. This is why, Jeb has the Jebnessique. That's a really smart pun, I think. A little bit of a pity laugh. But thank you. You get what you optimize for. I'm looking for a better word than reliable. So if you guys have suggestions.

Let me know, but there's this reliability of intelligence that's really missing in LLMs. I often call it you get Einstein ninety five percent of the time, but Mr Bean five. If you can't tell which is which, how the fuck is it automated? And instead, instead of like focusing on like sometimes occasional smarts. That you can then verify. What if you focused on extreme, extreme narrow focused reliability?

So you can get that smarts every time, and this doesn't show up in benchmarks. It is the reason why people love Jeff. And this is also I realize with Mike's like we obviously did prepare for a ChatGPT moment for. Developers, but it happened because people in the AI space just. Overpromise, underdeliver so fucking much that having a model that actually doesn't do that was a surprise. And it is not the speed, cost, or even interface.

This is something that you know my team tells me I shouldn't be telling people so that we can have a long term advantage, and I care too much. But it's reliable intelligence. Make sick models, please. Even if we, you know, if it doesn't involve our models.

00:10:25 — Smart software

We see that Java is like a foundation for smart software. I'm going to skip all of this, but I think software is going to be. So fucking bad. I think the narrative of SaaS apocalypse is obviously not going to happen. But I think that. Even calling it a sassa palooza is like an understatement, and I'll get into that bit. And you guys ideally are the perfect crowd for this.

I'm so excited for that part because that's the part that's new to me. I want to make a distinction between smart software and just-in-time software. The crazy part is AI is so transformative, and yet. Somehow, despite AI being so awesome, software has essentially unchanged in the last decade. Except being cheaper to write, isn't that what weird? Isn't it weird that the best we can do in software?

Is having a chatbot to the side that sometimes interacts with the software, but cannot take expensive decisions for you. That is. Kind of like a weird thing if AI was really that transformative. And I think the problem is we're focusing too much on coding agents. And not enough about putting the intelligence into the inner loop, which I'm going to get more into a bit, and. People are doing. There's been tons of successes. We don't need to flex this. There's so many cool things.

00:11:32 — LLM Stockholm syndrome

But my AIE secret is going to be that the field isn't getting hard enough. And this is what Sutik said that I can make a call to action for you guys that's unique, and hopefully I have enough time for this. I think the world still has LLM Stockholm syndrome. We are still thinking from a really LLM built place. Which is, I think, tragic. You know, like. I love the software saying the tools you use change the way you think, and I think that it's changed the way we think way too much.

And we're not really rethinking software like I think the field needs. I think we could just be doing so much more. Hopefully, I can try to do my best for my remaining time to convince you guys of that. So let's jet to the future.

00:12:12 — The blue LED

I an analogy I love. Bear with me is about LEDs. Before there was the blue LED, there was red and green LEDs. These were really limited in capability, and they were used for like simple indicators, and you know like. Clocks and stuff like that, and when the blue LED was made discovered, I actually don't really know how. Things outside of computer science work, but when it existed, I guess.

It's enabled not just like 50 percent more use cases, which lets you. About like changing colors from two to three, it enabled so many more. Like now, LEDs could be used for white light. It could be used for like the displays that we all use right now, and it created a revolution in lighting. And it's it's it's unfathomable how much bigger. The capabilities of LEDs were just.

You know, adding a new primitive that composes together with old primitives. You can see where I'm going here, maybe, and. I think it's a similar story for software. Software is so fucking incredible. Ignore AI for the last you know half decade. But software has done so many insanely cool things. We've done a lot of stuff, you know. Like, do I? Love using SaaS, not really, but it doesn't solve problems.

Absolutely, and it's incredible how much we've done with three fundamental logics. You know, like we are doing this with like simple binary operations that can do things like greater than, less than, equal. It's kind of nuts. How much software has been built from such simple primitives? And my claim is that what we are doing with AI now, when we're making it our intelligence machine native. This is kind of conceptually adding a fourth semantic logic gate.

What if we can program with common sense, with intention, with human understanding? Like, do we still need regexes with like crazy things, or can we just say like we want a Gmail email address or a real looking one? There's so much that we can do that we're going to need to push the you know push the frontiers of. Cheap, smart, fast, in order to get done. But I think there's going to be like so many things here, and I think just.

Like the blue LED, it's gonna be way bigger than just. Increasing the representational ability of software by 33% because it's adding effectively another. Kind of sort of logic gate, and this sci-fi is the stuff that I think will be impossible to predict. Just like who could predict how powerful the internet would have been? Because it requires, it's like much more grassroots and decentralized, and it requires people to be empowered to do insane stuff.

00:14:32 — Software research

I call this software research. That's not a joke, but. I think that there's going to be like an explosion of what software can do. And we will want people to be exploring those boundaries of software. Not necessarily because it's valuable, but because they can, because we have like new frontiers to explore. And this is what the early internet was like. Should all of those websites existed? Probably not. But like, it was very.

Cool, and we found a lot of awesome stuff from that, and I think we you know we have another opportunity for that, so we should not just keep living. In this little, you know, fish tank that we are the victims of openai tropic on whatever they ship. And we just have to be like Ellen wrappers. So I'm going to pitch. Software research is the direction here. The theme is going to be real software engineering, not just white coding.

It's gonna be the kind of stuff that everyone has known if you've been a software engineer. Decomposition, breaking tasks down. We need to think about stages, ideally ones that can be verified and have real encapsulation. And we don't need to actually do state management. Instead of assuming everything is global state within a KV cache, and the theme is just going to be like real, talking engineering. By people who care a lot about the end result, it doesn't have to be valuable. Ideally, it's sick as fuck.

And you know we should be pushing the limits.

00:15:50 — Ideas: rethinking agents

I'm gonna go into a couple of like my favorite ideas, and I want to figure out how to encourage these in the world. So this is my first step at this. I'm just going to talk about it. We have raised a bunch of money, so I would like to figure out how to fund this. But we don't have the operations help to do that. But here are just some ideas out there.

I think agents are, they're very cool, but like they're just. They're so inflexible in design because of the tyranny of the KB cache. I don't have the time for this. If this timer is to be believed, but believe me that there's so much that. People have tried to do, but fail that because it ends up becoming inefficient because fights against the KB cache. But I think there's a lot more to do. Why is memory open, right? Like, why can't we have all of our history of coding agents?

Always included in every new call. That. Is how continual learning works, right? And we have the memory available. It's just that we don't have the smarts. To map, produce over the memory to find out like what you should what you should remember what you should. I think there's many ways to do log in or even better lookups. Why, when you're making many agents at once, why can't they not share state? Why do they need to rediscover the same patterns of the codebase?

Over and over and over again. And why can't non-naive routing and subagents work? Like even naive routing doesn't work. But like, why can't we do these things? And why does like the you know the big labs keep going back and forth on whether or not subagents work? I think that, like from an engineering perspective, there's elegance to it, but they fight against the KDC. And I think that if we design agents around that, we could do so much more.

00:17:20 — Ideas: principal-agent problems

This one's probably the most vague of my directions, but I feel very strongly about it. The world is full of principal-agent problems. This is when you have like you know ulterior motives between this is not a coding agent but like any agent. The classic example is. Citizens versus elected officials. I don't want to make. A too political claim here, but like the misalignment in all levels of society.

And this results in like waste, fraud, bad things, and it results in for a society that wants fairness. You need to overemphasize the letter over the spirit. Which then results in loopholes and pay to win dynamics in the world. And I think the root of that problem is we can't encode the spirit. Just like with software, we're constrained to encoding bits instead of what we actually meant in that situation, and I think so many.

Organization, society can be transformed by really encoding the spirit all over the place.

00:18:11 — Ideas: real-time security, types and interfaces

Obviously, real-time intelligence security. Can we prevent the OpenAI Hugging Face sandbox? Blah blah. Why did why couldn't they prevent it? It is not because they didn't have the intelligence. I would argue it's a lot easier to figure out they're doing a naughty thing. Than to actually do the naughty thing, it is because it was not. Fast, cheap, and reliable enough to actually run real time, and what happens when?

Every corporation has their own real-time security that's actually smart. How cool would that fucking be? I can't talk about generative UIs because people are going to be talking about it afterwards. I don't want to steal their thunder. But it's coolest chips, trust me. Like semantic dependent typing, why can't That we have like a type system that actually is semantic, right? Like, why can't we? Like declare what local properties we want, so we don't have to read a chain of dependencies deep down.

These all should be possible, and imagine if you could like run a GPT-like model. In your lint checker to check here, it's all type safe and it could uphold semantic properties. It would be so fucking cool. There was a great Twitter video on. Like a voice interface. In general, I think new interfaces are so exciting. This is to me the return of like this is going to show by agent software like do what I mean.

This used to be like a popular thing, but it never really worked because. It was too many hard-coded rules to really figure out what that means. This stopping slot, obviously a big theme of here, but like we definitely believe this. This is my props. I have to bet. We have a stop GenAI baseball bat for TypeSafe. It is very awesome.

00:19:45 — The root of slop

I want to talk about the root of slop. The root of slop is that AI is unreliable. Therefore, you as a company, or you as you know, like someone who's doing stuff, it can't be used for decisions that costs you stuff. But it's actually totally fine to push the costs into other people. Which is horrible, you know, externalized costs, and then this is what ends up becoming stock.

You know, if people actually read the stuff and filter it and like care about the quality, it becomes not slop. But now the world is full of this. Because there's an asymmetric advantage for people who are like bad actors and don't give a shit. So I think that we can save the world from at least in bad hell among many other things, and that would be very cool.

00:20:25 — Cursed software and automating everything

And even something I think is super sick is Curse Software. Like why? If you guys ever seen some really bad UI baffles. It is so fucking cool. Even though they make horrible interfaces, like this whole game of like, would you like to delete your account becomes. You know a random thing, and I actually think. Doing things because we can is what software engineering is about. Because who knows? Maybe one percent of these ideas actually lead to something.

Super awesome, and I think like really cultivating even bad ideas is kind of what being like this you know playful hacker is all about. And generally, just automating everything. This is what the follow up. This is this is slop an inch, but our logo is a little bit wrong. But don't stare at that. That's a little weird. This is what a lot of the follow-up talks are going to be about.

00:21:07 — What couldn't we automate?

Thank you. If you're a software engineer, you almost certainly think in the world of. If you know anything about software, and you've interacted with software. You know that there are things software can do and software cannot do because that is just only obvious. Right, like there's some things that you know you need to do because that is not compatible with software. But what if Those constraints were no longer true.

Maybe not today, but maybe in like six months or a year. What couldn't we automate? And that's kind of where I want to lead this whole thing. Because I think there is very likely to be a future where it's not a question of whether or not we can automate something. It's going to be a question of like how much work is it, which is how it is right now for the stuff that can be automated, and that would be cool as shit.

So hell yeah!

00:21:57 — Thanks

Thank you. Woo! Deep overtime, but thank you. Early believers, you know, we had no idea so many people would resonate with this. We thought we were the insane ones, but it turns out like we were actually the majority of the world who realized AI could be so much more. Thank you to like the people who are like really realistic about reliability. Everyone knows I hate benchmarks, and a lot of other people are really Screaming out like benchmarks fucking suck.

They're actually. They hurt the people who do not have time for anything other than benchmarks. Most, it's tragic. And for the people who are pushing the limits of software, like fuck yes. And we want to be there with you, helping you all do that. The dreamers and the people who don't want to be limited by Everything that has happened so far, we. Want to really fight for you, and we want this to be as reliable as a fourth logic gate.

And that is a very high bar, and we have like a lot of work to do. So, thank you, thank you.