John Carmack talk at Upper Bound 2025 | Hacker News Reader

John Carmack talk at Upper Bound 2025

twitter.com

557 pointstosh360 comments

It's always a treat to watch a Carmack lecture or read anything he writes, and his notes here are no exception. He writes as an engineer, for engineers and documents all his thought processes and misteps in the exact detailed yet concise way you'd want a colleague to who was handing off some work.

One question I would have about the research direction is the emphasis on realtime. If I understand correctly he's doing online learning in realtime. Obviously makes for a cool demo and pulls on his optimisation background, and no doubt some great innovations will be required to make this work. But I guess the bitter lesson and recent history also tell us that some solutions may only emerge at compute levels beyond what is currently possible for realtime inference let alone learning. And the only example we have of entities solving Atari games is the human brain, of which we don't have a clear understanding of the compute capacity. In which case, why wouldn't it be better to focus purely on learning efficiency and relax the realtime requirement for now?

That's a genuine question by the way, definitely not an expert here and I'm sure there's a bunch of value to working within these constraints. I mean, jumping spiders solve reasonably complex problems with 100k neurons, so who knows.

It's because humans (and other animals) have enormous innate capacities and knowledge which makes learning new things much much simpler than if you start from scratch. It's not really because of human's computational capacity.

By innate do you mean evolved/instinctive? Surely even evolved behaviour must be expressed as brain function, and therefore would need a brain capable of handling that level of processing.

I don't think it's clear how much of a human brains function exists at birth though, I know it's theorised than even much of the sensory processing has to be learned.

I'm not arguing against computational theory of mind, I'm just saying that innate behaviours don't require the same level of scale as learnt ones.

Existing at birth is not the same thing as innate. Puberty is innate but it is not present at birth.

That's an interesting point. I can see that, as you say puberty and hormones impact brain function and hence behaviour, and those are inate and not learned. But at least superfically that would appear to be primarily broad behavioural effects, similar to what might be induced by medication. Rather than something that impacts pure abstract problem solving, which I guess is what the Atari games are supposed to represent?

This is obviously wrong from genetic defects that cause predictable development problems in specialized areas. They are innate but not present at birth.

> enormous innate capacities and knowledge

Hundreds of millions of years of trial-and-error biological pre-training where survival/propagation is the reward function

Yes but it remains innate to the individual.

There is just no reason to believe that we are born with some insanely big library of knowledge, and it sounds completely impossible. How would it be stored, and how would we even evolve it?

It just isn't needed. Just like you can find let's say kangaroos in the latent space of an image generator, so we learn abstract concepts and principles of how things work as a bonus of learning to process the senses.

Maybe a way to AGI could be figuring out how to combine a video generator with a LLM or something similar in a way that allows it to understand things intuitively, instead of doing just lots and lots of some statistical bullsit.

> There is just no reason to believe that we are born with some insanely big library of knowledge, and it sounds completely impossible. How would it be stored, and how would we even evolve it?

We do have that, ever felt fear of heights? That isn't learned, we are born with it. Same with fear of small moving objects like spiders or snakes.

Such things are learned/stored very different from memories, but its certainly there and we can see animals also have those. Like cats gets very scared of objects that are long and appear suddenly, like a cucumber, since their genetic instincts thinks its a snake.

Of course it is learned, and fear is triggered by anything unfamiliar, that causes a high reconstruction error. Because it means you don't understand it, and it could be dangerous. We are just not used to encoding anything so deep below the eye level, and it freaks us out.

Do you really think every single ant is learning all that on its own? And if ants can store that in their DNA, why don't you think other animals can? DNA works just fine as generic information storage, there are obviously a ton of behaviors and information encoded there from hundreds of millions of years of survival of the fittest.

Yeah that's why people are terrified by white noise.

> Like cats gets very scared of objects that are long and appear suddenly, like a cucumber, since their genetic instincts thinks its a snake.

After having raised four dozen kittens that a couple of feral sisters gave birth to in my garage, I’m certain that is nonsense. It’s an internet meme that became urban legend.

I don’t think they have ever even reacted to a cucumber, and I have run many experiments because my childhood cat loved cucumbers (we’d have to guard the basket of cucumbers after harvest, otherwise she’d bite every single one of them… just once).

>How would it be stored, and how would we even evolve it?

DNA and the same way anything else is evolved? The body is insanely complicated, I don't see why innate knowledge is so unbelievable.

From the notes:

"A reality check for people that think full embodied AGI is right around the corner is to ask your dancing humanoid robot to pick up a joystick and learn how to play an obscure video game."

This sounds like a problem that could be solved around the corner with a caveat.

Games generally are solvable for AI because they have feedback loops and a clear success or failure criteria. If the "picking up a Joystick" part is the limiting factor, sure. But why would we want robots to use an interface (especially a modern controller) heavily optimized for human hands; that seems like the definition of a horseless carriage.

I'm sure if you compared a monkey and a dolphins performance using a joystick you'd get results that aren't really correlated with their intelligence. I would guess that if you gave robots an R2D2 like port to jack into and play a game, that problem could be solved relatively quickly.

The point isn't about learning video games its about learning tasks unrelated to its specific competency generally.

No, the joystick part is really not the limiting factor. They’ve already done this with a direct software interface. Physical interface is a new challenge. But overall you are missing the point.

Just like OpenAI early on promised us an AGI and showed us how it "solved" Dota 2.

They also claimed it "learned" to play by playing itself only however it was clear that most of the advanced techniques were borrowed from existing AI and by observing humans.

No surprise they gave up on that project completely and I doubt they'll ever engage in anything like that again.

Money better spent on different marketing platforms.

It also wasn't even remotely close to learning Dota 2 proper. They ran a massively simplified version of the game where the AI and humans alternated between playing one of two pre-defined team compositions, meaning >90% of the games characters and >99.999999% of the possible compositions and matchups weren't even on the table, plus other standard mechanics were also changed or disabled altogether for the sake of the AI team.

Saying you've solved Dota after stripping out nearly all of its complexity is like saying you've solved Chess, but on a version where the back row is all Bishops.

Exactly. What I find surprising in this story though is not the OpenAI. It's investors not seeing through these blatant.. lets call them exaggerations of the reality and still trusting the company with their money. I know I wouldn't have. But then again, maybe that's why I'm poor.

In their hearts, startup investors are like Agent Mulder: they Want To Believe. Especially after they’ve already invested a little. They are willing to overlook obvious exaggerations up to and including fraud, because the alternative is admitting their judgment is not sound.

Look at how long Theranos went on! Miraculous product. Attractive young founder with all the right pedigree, credentials, and contacts, dressed in black trurtlenecks. Hell, she even talked like Steve Jobs! Investors never had a chance.

They already have 400 million daily users and a billion people using the product, with billions of consumer subscription revenue, faster than any company ever. They are also aggregating R&D talent at a density never before seen in Silicon Valley

That is what investors see. You seem to treat this as a purity contest where you define purity

I'm speaking about past events. Perhaps I didn't make it clear enough

Also apparently still not making a profit.

It was 6 years ago. I'm sure now there'd be no contest now if OpenAI dedicated resources to it, which it won't because it's busy with solving entirety of human language before others eat their lunch.

What do you base your certainty on? Were there any significant enough breakthroughs in the AGI?

ARC-AGI, while imagined as super hard for AI, was beaten enough that they had to come up with ARC-AGI-2.

"AI tend to be brittle and optimized for specific tasks, so we made a new specific task and then someone optimized for it" isn't some kind of gotcha. Once ARC puzzles became a benchmark they ceased to be meaningful WRT "AGI".

So if DOTA became a benchmark same way Chess or Go became earlier it would be promptly beaten. It just didn't stick before people moved to more useful "games".

Funnily enough, even dota2 has grown much more complex than it was 6 years ago, so it's a harder problem to solve today than it was back then

I agree that restricting the hero pool is a huge simplification. But they did play full 5v5 standard dota with just a restricted hero pool of 17 heroes and no illusions/control units according to theverge (https://www.theverge.com/2019/4/13/18309459/openai-five-dota...). It destroyed the professionals.

As an ex dota player, I don't think this is that far off from having full on, all heroes dota. Certainly not as far of as you are making it sound.

And dota is one of the most complex games, I expect for example that an AI would instantly solve CS since aim is such a large part of the game.

Another issue with the approach is that the model had direct access to game data, that is simply an unfair competitive advantage in dota, and it is obvious why that advantage would be unfair in CS.

It is certainly possible, but i won't be impressed by anything "playing CS" that isn't running a vision model on a display and moving a mouse, because that is the game. The game is not abstractly reacting to enemy positions and relocating the cursor, it's looking at a screen, seeing where the baddy is and then using this interface (the mouse) to get the cursor there as quickly as possible.

It would be like letting an AI plot its position on the field and what action its taking during a football match and then saying "Look, The AI would have scored dozens of times in this simulation, it is the greatest soccer player in the world!" No, sorry, the game actually requires you to locomote, abstractly describing your position may be fun but it's not the game

Did you read the paper? It had access to the dota 2 bot API, which is some gamestate but very far from all gamestate. It also had artifially limited reaction to something like 220ms, worse then professional gamers.

But then again, that is precisely the point. A chess bot also has access to gigabytes of perfect working memory. I don't see people complaining about that. It's perfectly valid to judge the best an AI can do vs the best a human can do. It's not really fair to take away exactly what a computer is good at from an AI and then say: "Look but the AI is now worse". Else you would also have to do it the other way around. How well could a human play dota if it only had access to the bot API. I don't think they would do well at all.

> But then again, that is precisely the point. A chess bot also has access to gigabytes of perfect working memory. I don't see people complaining about that.

There are ~86 billion neurons in the human brain. If we assume each neuron stores a single bit a human also has access to gigabytes of working memory. If we assume each synapse is a bit that's terabytes. Petabytes is not unreasonable assuming 1kb of storage per synapse. (And more than 1kb is also not unreasonable.)

The whole point of the exercise is figuring out how much memory compares to a human brain.

No human can, or would, flush their entire brain to use every single neuron as working memory for chess. By doing that you would even forget the rules of chess. At best a tiny subset of neurons could be used for that.

I wouldn't have expected for anyone to even attempt to argue a human can beat, or even approach, a computer on working memory. Wikipedia is just 24.05gb. You are somehow claiming here that a human can hold that in working memory. That is they read it once and have perfect recall. Not even the most extreme savants have shown such feats.

It depends on what you mean by "memory". Pure data recall? Sure, a computer has humans beat, but that's not really the purpose of human memory. I can freely reason about mathematical theorems I learned decades ago, and there are many mistakes that I've only made once, and will never make again.

> No human can, or would, flush their entire brain to use every single neuron as working memory for chess.

We don't know what this means. Each neuron connects to thousands of synapses. I would assume that there is some quantity of information encoded in each pairwise connection of synapse-paths through a neuron. I would assume this is more than a bit, and also that it has something that might be described as a lossy fractal compression with each pathway adding or subtracting from whatever structures store information so that each path can use the same physical things - though not with perfect fidelity.

But the nuts and bolts are somewhat beside the point. The point is that if you look at Leela zero it only needs like 3GB of RAM to run and we have no evidence this is more memory than human grandmasters use to play chess. Yes, humans have imperfect recall but that's not relevant because neural net based chess engines do not work based on perfect recall.

It's fine if the computer has access to gigabytes of working memory, it can use all of the "natural" advantages that it has to play the game, that's perfectly fair, but there is no comparison to make when you give models bespoke machine interfaces to play games whose core mechanics revolve around perception and physical coordination, it may be impressive, but they are playing a different game, something akin to HvH Counter Strike.

And you can try to play some game where you create disadvantages to try to balance out all of the advantages of the machine interface, but again, hard to reason about the edge cases, and easy to create a misleading headline like "artificially limited reaction time worse than professional gamers" while in practice being able to react to information much more quickly than a human player because of its exclusive interface to game state. All of that is fair and well, and doesn't take anything away from the very cool achievements of google et al., but when you change the core mechanics of the game to accommodate a uniquely challenged player, you're playing a different game! Chess is ~mostly not about physically moving the pieces on the board, but Counter Strike is about little more than that! (And dota is somewhere in between.)

> It destroyed the professionals.

Only the first time, later when it played better players it always lost. Players learned the faults of the AI after some time in game and the AI had very bad late game so they always won later.

Not on the last iteration.

To be fair humans have had quite a few million years across a growing population to gather all of the knowledge that we have.

As we're learning with LLMs, the dataset is what matters - and what's awesome is that you can see that in us, as well! I've read that our evolution is comparatively slow to the rate of knowledge accumulation in the information age - and that what this means is that you can essentially take a caveman, raise them in our modern environment and they'll be just as intelligent as the average human today.

But the core of our intelligence is logic/problem solving. We just have to solve higher order problems today, like figuring out how to make that chart in excel do the thing you want, but in days past it was figuring out how to keep the fire lit when it's raining. When you look at it, we've possessed the very core of that problem solving ability for quite a while now. I think that is the key to why we are human, and our close ancestors monkeys are...still just monkeys.

It's that problem solving ability that we need to figure out how to produce within ML models, then we'll be cooking with gas!

A human would learn it faster, and could immediately teach other humans.

AI clearly isn't at human level and it's OK to admit it.

> But why would we want robots to use an interface (especially a modern controller) heavily optimized for human hands; that seems like the definition of a horseless carriage.

Elon's response to this is that if we want these androids to replace human jobs then the lowest friction alternative is for the android to be able to do anything a human can do in a human amount of space. A specialized machine is faster and more efficient, but comes with engineering and integration costs that create a barrier to entry. Elon learned this lesson the hard way when he was building out the gigafactories and ended up having to hire a lot of people to do the work while they sorted out the issues with the robots. To someone like Elon a payroll is an ever growing parasite on a companies bottom line, far better if the entire thing is automated.

We don't really need AGI. We need better specialized AIs. Throw in a few specialized AIs and they will leave some impact in the society. That might not be that far away.

Specialized AIs have been making an impact on society since at least the 1960s. AI has long suffered from every time they come up with something new it gets renamed and becomes important (where it makes sense) without giving AI credit.

From what I can tell most in AI are currently hoping LLMs reach that point quick just because the hype is not helping AI at all.

Yeah I agree with it. There is a lot of hype, but there is some potentials there.

Yeah “AI” tools (such a loose term but largely applicable) have been involved in audio production for a very long time. They have actually made huge strides with noise removal/voice isolation, auto transcription/captioning, and “enhancement” in the last five years in particular.

I hate Adobe, I don’t like to give them credit for anything. But their audio enhance tool is actual sorcery. Every competitor isn’t even close. You can take garbage zoom audio and make it sound like it was borderline recorded in a treated room/studio. I’ve been in production for almost 15 years and it would take me half a day or more of tweaking a voice track with multiple tools that cost me hundreds of dollars to get it 50% as good as what they accomplish in a minute with the click of a button.

Bitter lesson applies here as well though. Generalized models will beat specialized models given enough time and compute. How much bespoke NLP is there anymore? Generalized foundational models will subsume all of it eventually.

Generalized models might be better but they are rarely more efficient.

You misunderstand the bitter lesson.

It's not about specialized vs generalized models - it's about how models are trained. The chess engine that beat Kasparov is a specialized model (it only plays chess), yet it's the bitter lesson's example for the smarter way to do AI.

Chess engines are better at chess than LLMs. It's not close. Perhaps eventually a superintelligence will surpass the engines, but that's far from assured.

Specialized AI are hardly obsolete and may never be. This hypothetical superintelligence may even decide not to waste resources trying to surpass the chess AI and instead use it as a tool.

I think your point that AI would refuse to play chess is interesting. To humans, chess is a strategic game. To a mathematician, chess is an exceedingly hard game, (pretty sure it is EXP complete, but I'm not fully familiar with Np/Exp completeness). To an AI, it seems like the AI will side with the mathematicians. AI is like "bro you can't even figure out if P=NP so how am I going to, you want me to waste power to solve an unsolvable problem?"

From Wikipedia, Garry Kasparov said it was a pleasure to watch AlphaZero play, especially since "its style was open and dynamic like his own".

People can't define AI because they don't want to consider AI as a subset of exponentially difficult algorithms, but they do want to consider AI as a generator of stylistic responses.

Yesterday my dad, in his late 70's, used Gemini with a video stream to program the thermostat. He then called me to tell me this, rather then call me to come stop by and program the thermostat.

You can call this hype, maybe it is all hype until LLMs can work on 10M LOC codebases, but recognize that LLMs are a shift that is totally incomparable to any previous AI advancement.

That is what open ai’s non-profit economic research arm has claimed. LLMs will fundamentally change how we interact with the world like the Internet did. It will take time like the Internet and a couple of hype cycle pops but it will change the way we do things.

It will help a single human do more in a white collar world.

https://arxiv.org/abs/2303.10130

There are clearly a lot of useful things about LLMs. However there is a lot of hype as well. It will take time to separate the two.

> He then called me to tell me this, rather then call me to come stop by and program the thermostat.

Sounds like AI robbed you of an opportunity to spend some time with your Dad, to me

Or maybe instead of spending time with your dad on a bs menial task, you could spent time fishing with him…

It's nice to think that but life and relationships are also composed of the little moments, which sometimes happen when someone asks you over to help with a "bs menial task"

It takes five minutes to program the thermostat, then you can have a beer on the patio if that's your speed and catch up for a bit

Life is little moments, not always the big commitments like taking a day to go fishing

That's the point of automating all of ourselves out of work, right? So we have more time to enjoy spending time with the people we love?

So isn't it kind of sad if we wind up automating those moments out of our lives instead?

For some of us that's a plus!

I'm there like twice a week don't worry. He knows about Gemini because I was showing him it two days before hah

Yeah. As a mediocre programmer I'm really scared about this. I don't think we are very far from AI replacing the mediocre programmers. Maybe a decade, at most.

I'd definitely like to improve my skills, but to be realistic, most of the programmers are not top-notch.

That is amazing. But I had a similar experience when I first taught my mum how to Google for computer problems. She called me up with delight to tell me how she fixed the printer problem herself, thanks to a Google search. In a way, LLMs are a refinement on search technology we already had.

Saying we don't "need" AGI is like saying we don't need electricity. Sure life existed before we had that capability, but it would be very transformative. Of course we can make specialized tools in the mean time.

Can you give an example how it would be transformative compared to specialized AI?

AGI is transformative in that it lets us replace knowledge workers completely, specialized AI requires knowledge workers to train them for new tasks while AGI doesn't.

Because it could very well exceed our capabilities beyond our wildest imaginations.

Because we evolved to get where we are, humans have all sorts of messy behaviours that aren't really compatible with a utopian society. Theft, violence, crime, greed - it's all completely unnecessary and yet most of us can't bring ourselves to solve these problems. And plenty are happy to live apathetically while billionaires become trillionaires...for what exactly? There's a whole industry of hyper-luxury goods now, because they make so much money even regular luxury is too cheap.

If we can produce AGI that exceeds the capabilities of our species, then my hope is that rather than the typical outcome of "they kill us all", that they will simply keep us in line. They will babysit us. They will force us all to get along, to ensure that we treat each other fairly.

As a parent teaches children to share by forcing them to break the cookie in half, perhaps AI will do the same for us.

Oh great, can't wait for our AI overlords to control us more! That's definitely compatible with a "utopian society"*.

Funnily enough, I still think some of the most interesting semi-recent writing on utopia was done ~15 years ago by... Eliezer Yudkowsky. You might be interested in the article on "Amputation of Destiny."

Link: https://www.lesswrong.com/posts/K4aGvLnHvYgX9pZHS/the-fun-th...

Who on earth has the resources to create true AGI and is interested in using it to create this sort of utopia for the masses?

If AGI is created it is most likely to be guided by someone like Altman or Musk, people whose interests couldn't be farther from what you describe. They want to make themselves gods and couldn't care less about random plebs.

If AGI is setting its own principles then I fail to see why it would care about us at all. Maybe we'll be amusing as pets but I expect a superhuman intelligence will treat us like we treat ants.

I still don’t see an issue of billionaires becoming trillionaires and being able to buy hyper luxury goods. Good for them and good for the people selling and manufacturing those goods. Meanwhile poverty is in all time lows and there’s a growing middle class at global level. Our middle class life conditions nowadays have a level of comfort that would get Kings from some centuries ago jealous.

If buying hyper luxury goods was the only thing that this wealth was good for, I'd agree.

But it's not: many of those billionaires are using their wealth to subvert democracy. They are quite literally enemies of the state.

The only lasting defense is to prevent this level of individual wealth accumulation.

I do haveba problem with that.

Basically, if we are going to allow single humans to have so much buying power, that needs to come with some clear expectations.

Risks must be taken. Pick something and advance society. Or, empower others in a reasonable, non discriminatory way, with no hooks.

There are a ton of things need doing.

Why not require a real effort?

We do not need more dynasties, and other enduring signs of wealth.

We do need those people to have earned it. (That is the no dynasty part again)

I know sounds harsh, but the fact is anyone can live very well for the rest of their lives on just a fraction of that money.

May well put it to work.

And if it were me personally? I got a list. Ready?

Why on earth would you want an AI that takes away our autonomy? It's wild to see someone actually advocate for this outcome.

Is this meant seriously? Do we really want something more intelligent than us to just force on us it's rules, logic and ways of living (or dying), which we may be too stupid to understand?

> humans have all sorts of messy behaviours that aren't really compatible with a utopian society. Theft, violence, crime, greed - it's all completely unnecessary and yet most of us can't bring ourselves to solve these problems.

Why would you think or claim those behaviors are unnecessary?

  > They will force us all to get along, to ensure that we treat each other fairly.

so, a dictatorship?

He he, too many comments to reply to. I see reactions in the same vein as those I get when I tell people it should be illegal to own a second/"holiday" home given the terrible housing markets preventing people from getting just one.

Well, I guess we shall see what the future brings, won't we.

The error in this argument is that electricity is real.

Indeed, and I'd go even further. In addition to existing, electricity is also usefully defined - which helps greatly in establishing its existence. Neither unicorns nor AGI currently exist but at least unicorns are well enough defined to establish whether an equine animal is or isn't one.

LLMs are being irresponsibly marketed as AGI to the general public and many do not seem to understand the difference. Clearly there are limits to how far current methods can be taken. They can only train on so much data and add more compute capacity before running into diminishing returns.

AGI is really something else entirely. If this was clarified, especially for Wall Street, we would see some changes in valuations.

At a certain point, it depends on how you define human reasoning capabilities. Maybe a collection of refined and specialized AI can get close. One counter argument could be that we are not far off from highly advanced robotics that have increasingly powerful reasoning abilities even without AGI.

After the automobile was created, global horse population fell dramatically. Who knows if the wealthy classes will need all these people to do the work for them in the near future.

Furthermore, we will be faced with it whether we want to or not, because others are making it happen.

What if AGI is just a bunch of specialized AIs put together?

It would seem our own generalized intelligence is an emergent property of many, _many_ specialized processes

I wonder if AI is the same

> It would seem our own generalized intelligence is an emergent property of many, _many_ specialized processes

You can say that about other animals, but about humans it is not so sure. No animal can be taught as general set of skills as a human can, they might have some better specialized skills but clearly there is something special that makes humans so much more versatile.

So it seems there was this simple little thing humans got that makes them general, while for example our very close relatives the monkeys are not.

Humans are the ceiling at the moment yes, but that doesn't mean the ceiling isn't higher.

Science is full of theories that are correct per our current knowledge and then subsequently disproven when research/methods/etc improves.

Humans aren't special, we are made from blood & bone, not magic. We will eventually build AGI if we keep at it. However unlike VCs with no real skills except having a lot of money™, I couldn't say whether this is gonna happen in 2 years or 2000.

Question was if cobbling together enough special intelligence creates general intelligence. Monkeys has a lot of special intelligence that our current AI models can't come close to, but still aren't seen as general intelligence like humans, so there is some little bit humans has that isn't just another special intelligence.

I guess the answer to that is that we don't know yet. I will be interesting to see if human intelligence/the way our brains work (which we certainly don't understand at all yet) is the only way to do intelligence or if models will eventually end up doing something novel.

It may be a property of (not only of?) humans that we can generate specialized inner processes. The hardcoded ones stay, the emergent ones come and go. Intelligence itself might be the ability to breed new specialized mental processes on demand.

Why not just hire like 100 of the smartest people across domains and give them SOTA AI, to keep the AI as accurate as possible?

Each of those 100 can hire teams or colleagues to make their domain better, so there’s always human expertise keeping the model updated.

"just"

They’re spending 10s of billions. Yes, just.

200 million to have dedicated top experts on hand is reasonable.

I think to many AI enthusiasts, we're already at the "specialized AIs" phase. The question is whether those will jump to AGI. I'm personally unconvinced but I'm not an ML researcher so my opinion is colored by what I use and what I read, not active research. I do think though that many specialized AIs is already enough to experience massive economic disruption.

This debate is exhausting because there's no coherent definition of AGI that people agree on.

I made a google form question for collecting AGI definitions cause I don't see anyone else doing it and I find it infinitely frustrating the range of definitions for this concept:

https://docs.google.com/forms/d/e/1FAIpQLScDF5_CMSjHZDDexHkc...

My concern is that people never get focused enough to care to define it - seems like the most likely case.

It doesn't really seem like there's much utility in defining it. It's like defining "heaven."

It's an ideal that some people believe in, and we're perpetually marching towards it

No, it’s never going to be precise but it’s important to have a good rough definition.

Can we just use Morris et al and move on with our lives?

Position: Levels of AGI for Operationalizing Progress on the Path to AGI: https://arxiv.org/html/2311.02462v4

There are generational policy and societal shifts that need to be addressed somewhere around true Competent AGI (50% of knowledge work tasks automatable). Just like climate change, we need a shared lexicon to refer to this continuum. You can argue for different values of X but the crucial point is if X% of knowledge work is automated within a decade, then there are obvious risks we need to think about.

So much of the discourse is stuck at “we will never get to X=99” when we could agree to disagree on that and move on to considering the x=25 case. Or predict our timelines for X and then actually be held accountable for our falsifiable predictions, instead of the current vide based discussions.

This is a great reply, thank you.

For me, I just zoom out a little further and say: at the rate AGI is approaching, what is the utility in trying to regulate it ahead of time?

Seems like advancement is slow enough that society can/will naturally regulate it based on what feels comfortable.

And it's a global phenomenon that can't have rules applied at the protocol level like the internet, because it's so culturally subjective.

Precedents need to be set first, and I think we'll only be able to call them when we see them.

It’s a good point. For epistemic hygiene I think it’s critical to actually have models of the growth rate and what is implied. Eg we are seeing exponential growth on many capability metrics (some with doubling-times of 7 months), but haven’t joined this up to economic growth numbers. In models where the growth continues you could imagine stuff getting crazy quickly, eg one year AI contributes 0.5% GDP only measurable in retrospect, next year 2%, year after 8%.

Personally I don’t think politicians are capable of adapting fast enough to this extreme scenario. So they need to start thinking about it (and building and debating legislation) long before it’s truly needed.

Of course if it turns out that we are living in one of the possible worlds where true economically meaningful capabilities are growing more slowly, or bottlenecks just happen to appear at this critical phase in the growth curve, then this line of preparation isn’t needed, but I’m more concerned about downside tail risk than the real but bounded costs of delaying progress by a couple years. (Though of course, we must ensure we don’t do to AI what we did to nuclear).

Finally I’ll note in agreement with your point, that there are a whole class of solutions that are mostly incomprehensible or inconceivable to most people at this time (ie currently fully outside the Overton Window). Eg radical abundance -> UBI might just solve the potential inequities of the tech, and therefore make premature job protection legislation vastly harmful on net. I mostly say “just full send it” when it comes to these mundane harms, it’s the existential ones (including non-death “loss of control” scenarios) that I feel warrant some careful thought. For that reason while I see where you are coming from, I somewhat disagree on your conclusion; I think we can meaningfully start acting on this as a society now.

This is great food for thought.

I like your idea of developing a new economic model as a proxy for possible futures; that at least can serve as a thinking platform.

Your comment inspired me to look at historical examples of this happening. Two trends emerged:

1. Rapid change always precedes policy. Couldn't find any examples of the reverse. That doesn't discount what you're saying at all, it reiterates that we probably need to be as vigilant and proactive as possible.

and related:

2. Things that seem impossible become normative. Electricity. The Industrial Revolution. Massive change turns into furniture. We adapt quickly as individuals even if societies collectively struggle to keep up. There will be many people that get caught in the margins, though.

Consider me fully convinced!

It is a marketing term. That's it. Trying to exhaustively define what AGI is or could be is like trying to explain what a Happy Meal is. At it's core, the Happy Meal was not invented to revolutionize food eating. It puts an attractive label on some mediocre food, a title that exists for the purpose of advertisement.

There is no point collecting definitions for AGI, it was not conceived as a description for something novel or provably existent. It is "Happy Meal marketing" but aimed for adults.

That’s historically inaccurate

My masters thesis advisor Ben Goertzel popularized the term and has been hosting the AGI conference since 2008:

https://agi-conference.org/

https://goertzel.org/agiri06/%5B1%5D%20Introduction_Nov15_PW...

I had lunch with Yoshua Bengio at AGI 2014 and it was most of the conversation that day

The name AGI (i.e. generalist AI) was originally intended to contrast with narrow AI which is only capable of one, or a few, specific narrow skills. A narrow AI might be able to play chess, or distinguish 20 breeds of dog, but wouldn't be able to play tic tac toe because it wasn't built for that. AGI would be able to learn to do anything, within reason.

The term AGI is obviously used very loosely with little agreement to it's precise definition, but I think a lot of people take it to mean not only generality, but specifically human-level generality, and human-level ability to learn from experience and solve problems.

A large part of the problem with AGI being poorly defined is that intelligence itself is poorly defined. Even if we choose to define AGI as meaning human-level intelligence, what does THAT mean? I think there is a simple reductionist definition of intelligence (as the word is used to refer to human/animal intelligence), but ultimately the meaning of words are derived from their usage, and the word "intelligence" is used in 100 different ways ...

> intended to contrast with narrow AI

I've thought for a while that the middle letter in AGI ('General' vs 'Specific') would be more useful and helpful if it were changed to Wide vs Narrow. All AIs can be evaluated on a scale of narrow to wide in terms of their abilities and I don't think that will change anytime soon.

Everyone understands that something is only wide or narrow in comparison to something else. While that's also true of the terms "general' and 'specific', those are less used that way in daily conversation these days. In science and tech we make distinctions about generalized vs specific but 'general' isn't a conversational term like 50 or 100 years ago. When I was a kid my grandparents would call the local supermarket, the 'general store' which I thought was an unusual usage even then.

I guess "general store" made more sense back then though. I grew up in the UK in the 60's and food shops were "narrow" - fishmonger, butcher, greengrocer (fruit & veg), bakery, etc. From that perspective a "general store" would have been noteworthy!

Generalization is a formal concept in machine learning and is measurable.

The Wikipedia article on AGI explains it well enough.

Researchers at Google have proposed a classification scheme with multiple levels of AGI. There are different opinions in the research community.

https://arxiv.org/abs/2311.02462

My whole point was to demonstrate that despite years of definition suggestions the term has never found broad agreement

Now that the term is in the general lexicon (which is crazy to me as an old guy doing this a long time) it’s morphing into something new

Like any good scientist i want to sample the population

In a way it sort of doesn't matter. If all we ever wind up with is unbelievably fancy chatbots, marketing teams will eventually draw their line in the sand and call it AGI.

There comes a theoretical point at which a definition is no longer relevant because it's obvious to everyone on an intuitive level. An easy lower bound for where this threshold might sit would be "when it can start and win wars unassisted under its own volition". Since, at that point no one on earth would have a need to debate it. It would simply be respected and understood for what it is without needing to define it.

Until such an obvious threshold is crossed, it will be whatever executives, product managers, and marketing teams say it is.

Is this supposed to be a gotcha? We know these systems are typically trained using RL and they are exceedingly good at learning games...

No it is not a “gotcha” and I don’t understand how you got that impression.

Carmack believes AGI systems should be able to learn new tasks in realtime alongside humans in the real world.

> the human brain, of which we don't have a clear understanding of the compute capacity

Neurons have finite (very low) speed of signal transfer, so just by measuring cognitive reaction time we can deduce upper bounds on how many _consecutive_ neuron connections are involved in reception, cognitive processing, and resulting reaction via muscles, even for very complex cognitive processes. And the number is just around 100 consecutive neurons involved one after another. So “the algorithm” could not be _that_ complex in the end (100x matmul+tanh?)

Granted, a lot of parallelism and feedback loops are involved, but overall it gives me (and many others) an impression that when the AGI algorithm is ever found, it’s “mini” version should be able to run on modest 2025 hardware in real time.

> (100x matmul+tanh?)

Biological neurons are way more complex than that. A single neuron has dentritic trees with subunits doing their own local computations. There are temporal dynamics in the firing sequences. There is so much more complexity in the biological networks. It's not comparable.

You could implement a Turing-machine with humans acting physically operating as logic gates. Then, every human is just a boolean function.

Neurons are stateful though, it is core to their function and how they learn.

This is exactly it. Biology is making massive use of hacked real time local network communication in ways we haven´t begun to explore.

The granted is doing a lot of work there. In fact, if you imagine a computer being able to do similar tasks as human brain can in around 100 steps, it becomes clear that considering parallelism is absolutely critical.

I'm sure there were offline rendering and 3D graphics workstation people saying the same about the comparatively crude work he was doing in the early 90s...

Obviously both Carmack and the rest of the world has changed since then, but it seems to me his main strength has always been in doing more with less (early id/Oculus, AA). When he's working in bigger orgs and/or with more established tech his output seems to suffer, at least in my view (possibly in his as well since he quit both Bethesda-id and Meta).

I don't know Carmack and can't claim to be anywhere close to his level, but as someone also mainly interested in realtime stuff I can imagine he also feels a slight disdain for the throw-more-compute-at-it approach of the current AI boom. I'm certainly glad he's not running around asking for investor money to train an LLM.

Best case scenario he teams up with some people who complement his skillset (akin to the game designers and artists at id back in the day) and comes up with a way to help bring some of the cutting edge to the masses, like with 3D graphics.

The thing about Carmack in the 90s... There was a lot of research going on around 3d graphics. Companies like SGI and Pixar were building specialized workstations for doing vector operations for 3d rendering. 3d was a thing. Game consoles with specialized 3d hardware would launch in 1994 with the Sega Saturn and the Sony Playstation (in Japan only for one year)

What Carmack did was basically get a 3d game running on existing COMMODITY hardware. The 386 chip that most people used for their excel spreadsheets did not do floating point operations well, so Carmack figured out how to do everything using integers.

May 1992 -> Wolfenstein 3d releases December 1993 -> Doom releases December 1994 -> Sony Playstation launches in Japan June 1996 -> Quake releases

So Wolfenstein and Doom were actually not really 3d games, but rather 2.5 games (you can't have rooms below other rooms). The first 3d game here is actually Quake which also eventually also got hardware acceleration support.

Carmack was the master of doing the seeminly impossible on super constrained hardware on virtually impossible timelines. If DOOM released in 1994 or 1995, would we still remember it in the same way?

> If DOOM released in 1994 or 1995, would we still remember it in the same way?

Maybe. One aspect of Wolfenstein and Doom's popularity is that it was years ahead of everyone else technically on PC hardware. The other aspect is that they were genre defining titles that set the standards for gameplay design. I think Doom Deathmatch would have caught on in 1995, as there really were very few (just Command and Conquer?) standout PC network multiplayer games released between 1993 and 1995.

I guess the thing about rapid change is... it's hard to imagine what kind of games would exist in a DOOMless world in an alternate 1995.

The first 3d console games started to come out that year, like Rayman. Star Wars Dark Forces with its own custom 3d engine also came out. Of course Dark Forces was, however, an overt clone of DOOM.

It's a bit ironic, but I think the gameplay innovation of DOOM tends to hold up more than the actual technical innovation. Things like BSP for level partitioning have slowly been phased out of game engines, we have ample floating point compute power and hardware acceleration ow, but even developers of the more recent DOOM games have started to realize that they should return to the original formula of "blast zombies in the face at high speed, and keep plot as window dressing"

Sort of in the middle, id games always felt tight. The engines were immersive not only because of graphics, but basic i/o was excellent.

> The first 3d console games started to come out that year, like Rayman.

Rayman was a 2D game.

It is. I have experienced others reference it as an early 3d Title too.

Great work by that team! Seriously. The graphics are great, and for the time period, excellent enough to be lumped in with the real 3d games.

I'm misremembering Rayman for the Sega Dreamcast. Looking at wikipedia, I now see that there weren't many games in 1995 even on the new consoles that really leveraged the 3d hardware. The PC actually had more such games despite many PCs lacking hardware acceleration for 3d rendering or even significant floating point capabilities. There's Sega Rally Racing for the Saturn, Virtua Fighter, Tekken...

Perhaps it's really 1996 that's the real turning point (with Mario 64 coming out), which makes DOOM about 3 years ahead of its time. And of course id shipped Quake that year....

> but even developers of the more recent DOOM games have started to realize that they should return to the original formula of "blast zombies in the face at high speed, and keep plot as window dressing"

There's still a lot of chatter breaking the continuity. In the original, the plot was entirely made up of what you were experiencing directly.

In the ending of the original game, you kill the demon spider brain robot thing and the demons kill your bunny rabbit. That's the plot

> Things like BSP for level partitioning have slowly been phased out of game engines

Hey, can you say more / do you have a link about this? I mean, for what reason are BSP trees phased out, and what are they replaced with? (quad/oct tree? AABB trees? or something entirely different?)

Read the quake pvs article linked to in this thread.

On top of my head as I remember it.. one of the reasons for Quake's use of a bsp was to allow back to front rendering of the world geometry without the use of a zbuffer. This was required to get decent performance with the software rasterizer.

I'm not 100% sure what's most commonly used these days, but for a large open world requiring data streaming, I could see the use for something like an octree and even portals.

The pipeline bottlenecks all changed in favor of bruteforcing the things that BSP had been solving with an elegant precomputed data structure - what BSP was extremely good at was eliminating overdraw and getting to where the scene could render exactly the number of pixels that were needed and no more. It's optimized around small, low-detail scenes that carefully manage occlusion.

More memory, bandwidth and cache means that more of your solutions are per-pixel instead of per-vertex and you can tolerate overdraw if it means you get to have higher polycount models. Likewise, the environment collision that was leveraged by the BSP process reduced the number of tests against walls, but introduced edge cases and hindered general-purpose physics features. Scaling physics leads in the direction of keeping the detailed collision tests at their original, per-poly detail, but doing things with sorting or tree structures to get a broadphase that filters the majority of tests against AABB or sphere bounds.

On a Wii(original Wii) 3D action game I helped ship, we just rendered the whole level at once, using only the most basic frustum culling technique; the hardware did the lifting, mostly through the z-buffer.

Adding to this, the nice thing about the bsp partitioning was you could also leverage it to make off screen monsters go to sleep or reduce their tick rate. Was helpful for optimizing AI as well as rendering. DOOM not only had some of the first pseudo 3d but also huge numbers of enemies... something that a lot of other games still cut down on

If DOOM released in 1994 or 1995, would we still remember it in the same way?

I think so, because the thing about DOOM is, it was an insanely good game. Yes, it pioneered fullscreen real-time perspective rendering on commodity hardware, instantly realigning the direction of much of the game industry, yadda yadda yadda, but at the end of the day it was a good-enough game for people to remember and respect even without considering the tech.

Minecraft would be a similar example. Minecraft looked like total ass, and games with similar rendering technology could have been (and were) made years earlier, but Minecraft was also good. And that was enough.

Hardware changes a lot in the time it takes to develop a game. When I read his plan files and interviews, I realized he seemed to spend a lot of time before developing the game thinking about what the next gen hardware was going to bring. Then design the best game they could think of whike targeting this not-yet-available hardware.

> So Wolfenstein and Doom were actually not really 3d games, but rather 2.5 games (you can't have rooms below other rooms). The first 3d game here is actually Quake

Ultima Underworld is a true 3D game from 1992. An incredibly impressive game, in more ways than one.

But also, he didn't do the technically hardest and most impressive part, Quake, on his own. IIUC he basically relied on Michael Abrash's help to get Quake done (in any reasonable amount of time).

Realizing that he needed Abrash (and aggressively recruiting him) could easily be seen as the most impressive thing he did to make Quake happen

I would say his multiple technical feats and phenomenal output are more impressive.

The world seems to have rewritten history, and forgotten Ultima Underworld, which shipped prior to Doom..

I think that's because it had such high system requirements that very few people could run it, unlike Wolfenstein 3D and Doom.

It was just about playable on a 25MHz 386. We used to put up with frame rates that would make your eyes bleed back in the early 90s.

https://www.youtube.com/watch?v=3VdRXgWoShM

Exactly, I played through underworld on a 386 16mhz. Being an rpg, the lower fps was much more tolerable than in Doom, which was in fact unplayable.

I'm sorry but that is not very playable even when it only renders to a quarter of the screen, especially compared to Wolfenstein 3D on similar hardware. It was also quite clunky in terms of interface. The guy on the video spends like 3 minutes trying to pick up a sack on the floor before just giving up.

Couple "3D" games shipped before Doom. Battlezone comes to mind.

The difference is that id owned the natural progression (from Wolf3D through Doom to Quake) and laid foundation to what we call today a FPS genre.

> his main strength has always been in doing more with less

Carmack builds his kingdom and then runs it well.

I makes me wonder how he would fare as an unknown Jr. developer with managers telling him "that's a neat idea, but for now we just need you to implement these Figma designs".

A key aspect of the Carmack approach (or similar 'smart hacker' unconventional career approach) is avoiding that situation in the first place. However, this also carries substantial career, financial and lifestyle risks & trade-offs - especially if you're not both talented enough and lucky enough to hit a sufficiently fertile oppty in the right time window on the first few tries.

Assuming one is willing to accept the risks and has the requisite high-talent plus strong work drive, the Carmack-like career pattern is to devote great care to evaluating and selecting opptys near the edges of newly emerging 'interesting things' which also: coincide with your interests/talents, are still at a point where a small team can plausibly generate meaningful traction, and have plausible potential to grow quickly and get big.

Carmack was fortunate that his strong interest in graphics and games overlapped a time period when Moore's Law was enabling quite capable CPU, RAM and GFX hardware to hit consumer prices. But we shouldn't dismiss Carmack's success as "luck". That kind of luck is an ever-present uncontrolled variable which must be factored into your approach - not ignored. Since Carmack has since shown he can get very interested in a variety of things, I assume he filtered his strong interests to pick the one with the most near-term growth potential which also matched his skills. I suspect the most fortunate "luck" Carmack had wasn't picking game graphics in the early 90s, it was that (for whatever reasons) he wasn't already employed in a more typical "well-paying job with a big, stable company, great benefits and career growth potential" so he was free to find the oppty in the first place.

I had a similarly unconventional career path which, fortunately, turned out very well for me (although not quite at Carmack's scale :-)). The best luck I had actually looked like 'bad luck' to me and everyone else. Due to my inability to succeed in a traditional educational context (and other personal shortcomings), I didn't have a college degree or resume sufficient to get a "good job", so I had little choice but to take the high-risk road and figure out the unconventional approach as best I could - which involved teaching myself, then hiring myself (because no one else would) and then repeatedly failing my way through learning startup entrepreneurship until I got good at it. I think the reality is that few who succeed on the 'unconventional approach' consciously chose that path at the beginning over lower risk, more comfortable alternatives - we simply never had those alternatives to 'bravely' reject in pursuit of our dreams :-).

  > "makes me wonder how he would fare as an unknown Jr. developer with managers telling him (...)"

he would probably write an open letter and left Meta. /s

Direct links:

https://docs.google.com/presentation/d/1GmGe9ref1nxEX_ekDuJX...

https://docs.google.com/document/d/1-Fqc6R6FdngRlxe9gi49PRvU...

Interesting reply from an openai insider: https://x.com/unixpickle/status/1925795730150527191

My bet is on Carmack.

I suspect Carmack in the Dancehall with the BFG.

What has he shipped in the last 20 years? Oculus is one thing, but that was firmly within his wheelhouse of graphics optimization. Abrash and co. handled the hardware side of things.

Carmack is a genius no doubt. But genius is the result of intense focused practice above and beyond anyone else in a particular area. Trying to extend that to other domains has been the downfall of so many others like him.

Ever since Romero departed the id Software had shipped *checks notes* Quake II, Quake III, Doom 3 and Quake 4.

Funnily enough Romero himself didn't ship much either. IMO it's one of the most iconic "duo breakups". The whole is greater than the sum of the parts.

Rage was Carmack's last big game at id Software before leaving.

Romero is credited on 27 games since he left id Software.

https://en.wikipedia.org/wiki/John_Romero#Games

None of them came close to the success of Quake, Doom or Commander Keen.

If you examine the list it includes games like "Gunman Taco Truck" by his 12yo sun, SIGIL I/II (Doom mods) and a remake of Dangerous Dave. Most of the money he made post-id came from Facebook farming games.

I'm not saying he's doing nothing. He's extremely talented and achieved more than most of us could ever dream of. I'm just pointing out that after he departed from id neither id nor him managed to replicate the earlier success. Who knows, maybe times had changed and it would be the same even if he stayed.

Their success with Doom and Quake was a confluence of things that cannot be replicated today. Carmack's programming talent gave them at least a year head start versus the competition. They introduced a new genre with no competition. Romero wrote game development tools that made them productive and able to deliver quickly. The artists and game designers created something innovative and fun to play, that stood the test of time.

Duke Nukem was released in 1996, then Unreal was released in 1998 and that's when they lost their technical advantage. The market became saturated with FPS.

Romero and Tom Hall founded Ion Storm which produced one successful game - Deus Ex. He gave up on AAA and went back to creating small games.

Carmack's licensed code was the basis of many successful games beyond the 90s, including Half Life 1 and 2 and the latest Doom games. We wouldn't have Half Life without id Software. Maybe Valve Software wouldn't exist.

"Graphics Carmack" is a genius but that doesn't mean that "AI Carmack" is too.

I wouldn't bet against him. "The Bitter Lesson" may imply an advantage to someone who historically has been at the tip of the spear for squeezing the most juice out of GPU hosted parallel computation.

Graphics rendering and AI live on the same pyramid of technology. A pyramid with a lot of bricks with the initials "JC" carved into them, as it turns out.

Only if computation is the bottleneck. GPT-4.5 shows it’s not.

I would be long carmack in the sense that I think he will have good judgement and taste running a business but I really don't see anything in common between AI and graphics.

Maybe someone better at aphorisms than me can say it better but I really don't see it. There are definitely mid-level low hanging fruits that would look like the kinds of things he did in graphics but the game just seems completely different.

I think people would do well to read about Philo Farnsworth in this context.

Carmack is always a genius, but like most people he requires luck, and like most people, the house always wins. Poor Armadillo Aerospace.

Exactly. I know him and like him. He is a genius programmer for sure BUT people forget that the last successful product that he released was Doom 3 over 20 years ago. Armadillo was a failure and Oculus went nowhere.

He's also admitted he doesn't have much of math chops, which you need if you want to make a dent in AI. (Although the same could have been said of 3D graphics when he did Wolfenstein and Doom, so perhaps he'll surprise us)

I wish him well TBH

Rage was released in 2011. His work at Meta produced highly optimized standalone VR. Whether you think it's successful or not, the tracking accuracy and latency is extremely competitive.

What has "Graphics Carmack" actually done since about 2001?

So, his initial tech was "Adaptive tile refresh" in Commander Keen, used to give it console style pixel-level scrolling. Turns out, they actually hampered themselves in Commander Keen 1 by not understanding the actual tech, and implemented "The Jolt", a feature that was not necessary. The actual hardware implemented scrolling the same way that consoles like the NES did, and did not need "the jolt", nor the limitations it imposed.

Then, Doom and Quake was mostly him writing really good optimizations of existing, known and documented algorithms and 3D techniques, usually by recognizing what assumptions they could make, what portions of the algorithm didn't need to be recalculated when, etc. Very talented at the time, but in the software development industry, making a good implementation of existing algorithms that utilize your specific requirements is called doing your job. This is still the height of his relative technical output IMO.

Fast Inverse Square Root was not invented by him, but was floating around in industry for a while. He still gets kudos for knowing about it and using it.

"Carmack's reverse" is a technique for doing stencil shadows that was a minor (but extremely clever) modification to the "standard" documented way of doing shadow buffers. There is evidence of the actual technique from a decade before Carmack put it in Doom 3 and it was outright patented by two different people the year before. There is no evidence that Carmack "stole" or anything this technique, it was independent discovery, but was clearly also just a topic in the industry at the time.

"Megatextures" from Rage didn't really go anywhere.

Did Carmack actually contribute anything to VR rendering while at Oculus?

People treat him like this programming god and I just don't understand. He was well read, had a good (maybe too good) work ethic, and was very talented at writing 386 era assembly code. These are all laudable, but doesn't in my mind imply that he's some sort of 10X programmer who could revolutionize random industries that he isn't familiar with. 3D graphics math isn't exactly difficult.

AI math isn't exactly difficult either.

Please read https://twobithistory.org/2019/11/06/doom-bsp.html and https://30fps.net/pages/pvs-portals-and-quake/

Also, I think most of the x86 magic was done by Abrash.

Appeal to authority is a logical fallacy. People often fall into the trap of thinking that because they are highly intelligent and an expert in one domain that this makes them an expert in one or more other domains. You see this all the time.

Expecting an expert in one thing to also be pretty good at other domains, especially when they're relatively related, isn't a fallacy.

Bayesian reasoning isn't a fallacy. A known expert in one domain is often correct about things in a related one. The post didn't claim that Carmack is right, just that that he's who they would bet on to be right, which seems perfectly reasonable to me.

> People often fall into the trap of thinking that because they are highly intelligent and an expert in one domain that this makes them an expert in one or more other domains.

While this is certainly true, I'm not aware of any evidence that Carmack thinks this way about himself. I think he's been successful enough that's he's personally 'post-economic' and is choosing to spend his time working on unsolved hard problems he thinks are extremely interesting and potentially tractable. In fact, he's actively sought out domain experts to work with him and accelerate his learning.

Funny, I was just commenting something similar here, see https://news.ycombinator.com/item?id=44071614

And I say this while most certainly not being as knowledgeable as this openai insider. So it even I can see this, then it's kinda bad, isn't it?

Can you explain which parts you think are bad and why?

Right? "Even I can see this" isn't exactly enlightening.

https://x.com/ID_AA_Carmack/status/1925973500327591979

  >> "they will learn the same lesson I did"

Which is what? Don't trust Altman? x)

Actually no, it's not interesting at all. Vague dismissal of an outsider is a pretty standard response by insecure academic types. It could have been interesting and/or helpful to the conversation if they went into specifics or explained anything at all. Since none of that's provided, it's "OpenAI insider" vs John Carmack AND Richard Sutton. I know who I would bet on.

He did go into specifics and explained his point. Or have you only read his first post?

It's a OpenAI researcher that's worked on some of their most successful projects, and I think the criticism in his X thread is very clear.

Systems that can learn to play Atari efficiently are exploiting the fact that the solutions to each game are simple to encode (compared to real world problems). Furthermore, you can nudge them towards those solutions using tricks that don't generalize to the real world.

That sounds like an extremely useful insight that makes this kind of research even more valuable.

Right, and the current state of tech - from accounts I’ve read, though not first hand experienced - is the “black box” methods of AI are absolutely questionable when delivering citations and factual basis for their conclusions. As in, the most real world challenge, in the basic sense, of getting facts right is still a bridge too far for OpenAI, ChatGPT, Grok, et al.

See also: specious ethics regarding the training of LLMs on copyright protected artistic works, not paying anything to the creators, and pocketing investor money while trying to legislate their way around decency in engineering as a science.

Carmack has a solid track record as an engineer, innovator, and above the board actor in the tech community. I cannot say the same for the AI cohort and I believe such a distinction is important when gauging the validity of critique or self-aggrandizement by the latter, especially at the expense of the former. I am an outlier in this community because of this perspective, but as a creator and knowledgeable enough about tech to see things through this lens, I am fine being in this position. 10 years from now will be a great time to look back on AI the way we’re looking back at Carmack’s game changing contributions 30 years ago.

Alex Nichol worked on "Gotta Learn Fast" in 2018 which Carmack mentions in his talk, he also worked on foundational deep learning methods like CLIP, DDPM, GLIDE, etc. Reducing him to a "seething openai insider" seems a bit unfair

It’s not vague, did you only see the first tweet or the entire thread?

Do you have an X account, if you're not logged in you'll only see the first post in the thread.

x.com/... -> xcancel.com/...

I use a Chrome extension to auto replace the string in the URL, works very well.

It seems that you’ve only read the first part of the message. X sometimes aggressively truncates content with no indication it’s done so. I’m not sure this is complete, but I’ve recovered this much:

> I read through these slides and felt like I was transported back to 2018.

> Having been in this spot years ago, thinking about what John & team are thinking about, I can't help but feel like they will learn the same lesson I did the hard way.

> The lesson: on a fundamental level, solutions to these games are low-dimensional. No matter how hard you hit them with from-scratch training, tiny models will work about as well as big ones. Why? Because there's just not that many bits to learn.

> If there's not that many bits to learn, then researcher input becomes non-negligible.

> "I found a trick that makes score go up!" -- yeah, you just hard-coded 100+ bits of information; a winning solution is probably only like 1000 bits. You see progress, but it's not the AI's.

> In this simplified RL setting, you don't see anything close to general intelligence. The neural networks aren't even that important.

> You won't see _real_ learning until you absorb a ton of bits into the model. The only way I really know to do this is with generative modeling.

> A classic example: why is frame stacking just as good as RNNs? John mentioned this in his slides. Shouldn't a better, more general architecture work better?

> YES, it should! But it doesn't, because these environments don't heavily encourage real intelligence.

I'm not sure what the moral is from this, but if Atari games are just too easy, at the same time the response of the machine-learning guys to the challenge of the NetHack Learning Environment seems to have mostly been to quietly give up. Why is generative modeling essential to finding harder challenges when NetHack is right there ...?

From a marketing perspective, this strikes me as a very predictable response.

I appreciate how they don't tell us what lesson they learned.

It is a thread. You may have only seen the first tweet because Twitter is a user-hostile trash fire.

“The lesson: on a fundamental level, solutions to these games are low-dimensional. No matter how hard you hit them with from-scratch training, tiny models will work about as well as big ones. Why? Because there's just not that many bits to learn.”

https://unrollnow.com/status/1925795730150527191

Thank you for clarifying. I don't have a Twitter account, and the linked tweet genuinely looks like a standalone object. Mea culpa.

Not your fault. They are the worst.

Each of these games is low-dimensional and require not the "intelligence" but more like "reflexes", I tend to agree.

However making a system that can beat an unknown game does require generalization. If not real a intelligence (whatever that means) but at the level of say "a wolf".

Whether it can arise from RL alone is not certain, but it's there somewhere.

I think some replies here are reading the full twitter thread, while others (not logged in?) see only the first tweet. The first tweet alone does come off as a dismissal with no insight.

indeed, this is pure walled garden sh*t

Carmack replied to that https://x.com/ID_AA_Carmack/status/1925973500327591979

I was really excited when I heard Carmack was focusing on AI and am really looking forward to watching this when the video is up - but just from looking at the slides it seems like he tried to build a system that can play the Atari? Seems like a fun project, but curious what will come out of it or if there is an associated paper being released.

Atari games are widely used in Reinforcement Learning (RL) research as a standard benchmark.

https://github.com/Farama-Foundation/Arcade-Learning-Environ...

The goal is to develop algorithms that generalize to other tasks.

They were highly used. OpenAI even included them in their RL Gym library back in the old days when they were still doing open research. But if you look at this leaderboard from 7 (yes, seven!) years ago [1], most of them were already solved way beyond human capabilities. But we didn't get a really useful general purpose algorithm out of it. As an AI researcher, I always considered Atari a fun academic exercise, but nothing more. Similar to how recognising characters using convnets was cool in the nineties and early 00s, but didn't give us general purpose image understanding. Only modern GPUs and massive training datasets did. Nowadays most cutting-edge RL game research focuses on much more advanced games like Minecraft which is thought to be better suited. But I'm pretty sure it's still not enough. Even role-playing GTA VI won't be. We probably need a pretty advanced physical simulation of the real world before we can get agents to handle the real world. But that means solving the problem of generating such an environment first, because you can't train on the actual real world due to the sample inefficiency of all current algorithms. Nvidia is doing some really interesting research in this direction by combining physics simulation and image generation models to simulate an environment, while getting accuracy and diversity at the same time into training data. But it still feels like some key ingredient is missing.

[1]https://github.com/cshenton/atari-leaderboard

> But it still feels like some key ingredient is missing.

Continuous training is the key ingredient. Humans can use existing knowledge and apply it to new scenarios, and so can most AI. But AI cannot permanently remember the result of its actions in the real world, and so its body of knowledge cannot expand.

Take a toddler and an oven. The toddler has no concept of what an oven is other than maybe that it smells nice. The toddler will touch the oven, notice that it experiences pain (because the oven is hot) and learn that oven = danger. Place a current AI in a droid toddler body? It will never learn and keep touching the oven as soon as the information of "oven = danger" is out of the context window.

For some cases this inability to learn is actually desirable. You don't want anyone and everyone to be able to train ChatGPT unsupervised, otherwise you get 4chan flooding it with offensive crap like they did to Tay [1], but for AI that physically interacts with the meatspace, constant evaluation and learning is all but mandatory if it is to safely interact with its surroundings. "Dumb" robots run regular calibration cycles for their limbs to make sure they are still aligned to compensate for random deviations, and so will AI robots.

[1] https://en.wikipedia.org/wiki/Tay_(chatbot)

This kind of context management is not that hard, even when building LLMs. Especially when you have huge windows like we do today. Look at how ChatGPT can remember things permanently after you said them once using a function call to edit the permanent memory section inside the context. You can also see that in Anthropic's latest post on Claude 4 where it learns to play Pokemon. The only remaining issue here is maybe how to diffuse explicit knowledge from the stored context into the weights. Andrej Karpathy wrote a good piece on this recently. But personally I believe this might not even be necessary if you can manage your context well enough and see it more like RAM while the LLM is the CPU. For your example you can then always just fetch such information from a permanent storage like a VDB and load it into context once you enter an area in the real world.

> This kind of context management is not that hard, even when building LLMs.

It is, at least if you wish to be in the meatspace, that's my point. Every day has 86400 seconds during which a human brain constantly adapts to and learns from external input - either directly as it's being awake or indirectly during nighttime cleanup processes.

On top of that, humans have built-in filters for training. Basically, we see some drunkard shouting about the Hollow Earth on the sidewalk... our brain knows that this is a drunkard and that Hollow Earth is absolutely crackpot material, so if it stores anything at all then the fact that there is a drunkard on that street and one might take another route next time, but the drunkard's rambling is forgotten maybe five minutes later.

AI, in contrast, needs to be hand-held by humans during training that annotate, "grade" or weigh information during the compilation of the training dataset, in order that the AI knows what is written in "Mein Kampf" so it can answer questions upon it, but that it also knows (or at least: won't openly regurgitate) that the solution to economic problems isn't to just deport Jews.

And huge context windows aren't the answer either. My wife says me, she would like to have a fruit cake for her next birthday. I'll probably remember that piece of information (or at the very least I'll write it down)... but an AI butler? I'd be really surprised if this is still in its context space in a year, and even if it is, I would not be surprised if it weren't able to recall that fact.

And the final thing is prompts... also not the answer. We've seen it just a few days ago with Grok - someone messed with the system prompt so it randomly interjected "white genocide" claims into completely unrelated conversation [1] despite hopefully being trained on a ... more civilised dataset, and to the contrary, we've also seen Grok reply to Twitter questions in a way that suggest that it is aware its training data is biased.

[1] https://www.reuters.com/business/musks-xai-updates-grok-chat...

>Every day has 86400 seconds during which a human brain constantly adapts to and learns from external

That's not even remotely true. At least not in the sense that it is for context in transformer models. Or can you tell me all the visual and auditory inputs you experienced yesterday at the 45232nd second? You only learn permanently and effectively from particular stimulation coupled with surprise. That has a sample rate which is orders of magnitude lower. And it's exactly the kind of sampling that can be replicated with a run-of-the-mill persistent memory system for an LLM. I would wager that you could fit most people's core experiences and memories that they can randomly access at any moment into a 1000 page book - something that fits well into state of the art context windows. For deeper more detailed things you can always fall back to another system.

Your definition of "learning" is incomplete because you're applying LLM concepts to how human brains work. An LLM only "learns" during training. From that point forward all it has is its context and vector DBs. If an LLM and vector DB is not actively interacted with, nothing happens to it. However for the brain, experiencing IS learning. And the brain NEVER stops experiencing.

Just because I don't remember my experiences at second 45232 on May 22, doesn't mean that my brain was not actively adapting to my experiences at that moment. The brain does a lot more learning than just what is conscious. And then when I went to sleep the brain continued pruning and organizing my unconscious learning for the day.

Seeing if someone can go from token to freeform physical usefulness will be interesting. I'm of the belief that LLMs are too verbose and energy intensive to go from language regurgitation machines to moving in the real world according to free form prompting. It may be accomplishable with the vast amount of hype investment, but I think the energy requirements and latency will make an LLM-based approach economically infeasible.

> I'm of the belief that LLMs are too verbose and energy intensive to go from language regurgitation machines to moving in the real world according to free form prompting.

This is not just possible, it is already happening. It just gets drowned in the media noise about chatbots. Look at some current research in this area (e.g. by Nvidia last year).

> You only learn permanently and effectively from particular stimulation coupled with surprise.

This is just, not true. A single 2min conversation with emotional or intellectual resonance can significantly alter a human’s thought process for years. There are some topics where every time they come up directly or analogously I can recall something a teacher told me in high school that “stuck” with me for whatever reason. And it isn’t even a “core” experience, just something that instantly clicked for my brain and altered my problem solving. At the time, there’s no heuristic that could predict how or why that particular interaction should have that kind of staying power.

Not to mention, experiences that subtly alter thinking or behavior just by virtue of providing some baseline familiarity instead of blank slate problem solving or routine. Like how you subtly adjust how you interact with coworkers based on the culture of your current company over time vs the last without any “flash” of insight required.

But that resonance is a form of surprise. You are just using different words for the same context. At the same time: You are correct in the sense, that this "surprise" is completely ignored by today's LLMs. They only use this in training mode and not for continuous learning. Whether one can find a sufficiently useful definition of "surprise" to use auxiliary "learning systems" (vector DBs or system prompts) has yet to be shown

You are just rephrasing things without using terms commonly used in research. I used "surprise" because it is a) true (if not complete) and b) easy to understand for people outside of the field. The correct term you are looking for is "arousal" (not necessarily sexual). There is tons of research on the fact that arousal enhances memory formation that would otherwise need many, many repetitions. But it also inhibits memorisation of nearby events. So you either need a very particular emotional state to remember a specific thing or massive repetitions to remember many new things. There's no way to cheat the sample inefficiency of your own brain. And for LLMs we have only figured out the first one, at least without external algorithms.

[1] https://www.nature.com/articles/nrn1052

[2] https://pubmed.ncbi.nlm.nih.gov/26151918/

Having used vectorDBs before, we're very much not there yet. We don't have any appreciable amounts of context for any reasonable real-life memory. It works if that is the most recent thing you did. Have you talked to an LLM for a day? Stuff is gone before the first hour. You have to use every trick currently in the book, treat context like it's your precious pet

VectorDBs are basically just one excuse of many to make up for a part of the system that is lacking capability due to technical limitations. I'm currently at 50:50 if the problems will be overcome directly by the models or by such support systems. Used to be 80:20 but models have grown in usefulness much faster than all the tools we built around them.

Big context windows are a poor substitute for updating the weights. Its like keeping a journal because your memory is failing.

It reminds me of the movie Memento.

it's more than that. Our understanding from space and time could be stemming from continuous training. Every time we look at something, there seems to be a background process that is categorizing items that are on the retinal image.

This is a continuous process.

> Humans can use existing knowledge and apply it to new scenarios, and so can most AI

Doesn't the article states that this is not true? AI cannot apply to B what it learned about A.

Well, ChatGPT knows about the 90s Balkan wars, a topic to which LWT hasn't made an episode that I'm aware of, and yet I can ask it to write a script for a Balkan wars episode that reads surprisingly like John Oliver while being reasonably correct.

Essentially Carmack pointed in the slides that teaching AI to play game a, b or c didn't improve AI at all at learning game d from scratch.

That's essentially what we're looking for when we talk about general intelligence, the capability to adapting what we know to what we know nothing about.

> Continuous training is the key ingredient. Humans can use existing knowledge and apply it to new scenarios, and so can most AI. But AI cannot permanently remember the result of its actions in the real world, and so its body of knowledge cannot expand.

I think it depends on how you look at it. I don't want to torture the analogy too much, but I see the pre-training (getting model weights out of an enormous corpus of text) as more akin to the billions of years of evolution that led to the modern human brain. The brain still has a lot to learn once you're born, but it already also has lots of structures (e.g. to handle visual input, language, etc) and built-in knowledge (instincts). And you can't change that over the course of your life.

I wouldn't be surprised if we ended up in a "pre-train / RAG / context window" architecture of AI, analogously to "evolution / long term memory / short term memory" in humans.

Being highly used in the past is good, it's a benchmark to compare against.

I watched the talk live. I felt that his main argument was that Atari _looks_ solved, but there's still plenty of value that could be gained by revisiting these "solved" games. For one, learning how to play games through a physical interface is a way to start engaging with the kinds of problems that make robotics hard (e.g., latency). They're also a good environment to study catastrophic forgetting: an hour of training on one game shouldn't erase a model's ability to play other games.

I think we could eventually saturate Atari, but for now it looks like it's still a good source of problems that are just out of reach of current methods.

Is a highly specialized bespoke robot for a Atari controller really that different? If anyone cared about latency they could have added it to the emulated controls and video with random noise.

I think it is. Latency was just one of the problems he described. A physical controller sometimes adds "phantom inputs" as the joystick transitions between two inputs. Physical actuators also slow down with wear. A physical Atari-playing robot needs to learn qualitatively different strategies that are somewhat more robust to these problems. Emulators also let the bot take as much time as it needs between frames, which is much easier than playing in real time. To me, all of this makes a physical robot seem like a decent way to start engaging with problems that come up in robotics but not simulated games.

My impression is that Atari was 80% solved, and then researchers and companies moved on.

A company solves self-driving 80% of the way and makes a lot of VC cash along the way. Then they solve intelligent chatbots 80% of the way and make a lot of VC cash along the way. Now they're working on solving humanoid robotics 80% of the way... I wonder why?

In the end, we have technology that can do some neat tricks, but can't be relied upon.

There are probably still some very hard problems in certain Atari games. Only the brave dare tackle these problems, because failure comes sharp and fast. Whereas, throwing more compute at a bigger LLM might not really accomplish anything, but we can make people think it accomplished something, and thus failure is not really possible.

DeepMind’s original demos were also of Atari gameplay.

He says they will open source it which is cool. I agree that I don't understand what's novel here. Playing with a physical controller and camera on a laptop GPU in real time is cool, and maybe that hasn't specifically been done before, but it doesn't seem surprising that it is possible.

If it is substantially more sample efficient, or generalizable, than prior work then that would be exciting. But I'm not sure if it is?

Maybe that's exactly his goal: not to come up with something that beats the competition, but play with the architecture, get a feel for what works & what doesn't, how various aspects affect the output, and improve on that. Design more efficient architectures, or come up with something that has unique features compared to other models.

If so, scaling up may be more of a distraction rather than helpful (besides wasting resources).

I hope he succeeds in whatever he's aiming for.

You will find many agents that solved (e.g., finished, reached high score) atari games, but there is still so much more work to do in the field. I wrote my Master's thesis on how to learn from few interactions with the game, so that if the algorithm is ported to actual robots they don't need to walk and fall for centuries before learning behaviors. I think there is more research to do on higher levels of generalization: when you know how to play a few video games, you quickly understand how to play a new one intuitively, and I haven't seen thorough research on that.

I can tell you right now without any research that video game designers reuse interface patterns and game mechanics that were already known when making new games. Those patterns and mechanics are also often analogies for real life allowing humans to intuitively play the games. If people can't play your game intuitively, they might say it's a bad game.

So why can't AI learn those and reapply the same understanding to new games?

His goal was not just to solve Atari games. That was already done.

His goal is to develop generic methods. So you could work with more complex games or the physical world for that, as that is what you want in the end. However, his insight is, you can even modify the Atari setting to test this, e.g. to work in realtime, and the added complexity by more complex games doesn't really give you any new additional insights at this point.

But how is this different to what NVIDIA have already done? They have robots that can achieve arbitrary and fluid actions in the real world by training NNs in very accurate GPU simulated environments using physics engines. Moving a little Atari stick around seems like not much compared to sorting through your groceries etc.

The approach NVIDIA are using (and other labs) clearly works. It's not going to be more than a year or two now before robotics is as solved as NLP and chatbots are today.

I think he argues that they would not be able to play Atari games this way (I don't know; maybe I also misunderstood).

But also, he argues a lot about sample efficiency. He wants to develop algorithms/methods/models which can learn much faster / with much fewer data.

I want smarter NPCs in games.

Here's what they built, https://x.com/ID_AA_Carmack/status/1925243539543265286

Quite exciting. Without diminishing the amazing value of LLMs, I don't think that path goes all the way to AGI. No idea if Carmack has the answer, but some good things will come out of that small research group, for sure.

Isn't that what Deepmind did 12 years ago?

Playing Atari games makes it easy to benchmark and compare and contrast his future research with Deepmind and more recent efforts.

IIRC Deepmind (and OpenAI and ...) have done this on software-only setups (emulators, TAS, etc); while this one has live input and actuators in the loop, so, kind of the same thing but operating in the physical realm.

I do agree that it is not particularly groundbreaking, but it's a nice "hey, here's our first update".

He points that out in his notes and says DeepMind needed specialized training and/or 200M frames of training just to kinda play one game.

What deepmind accomplished with suicidal Mario was so much more than you probably ever will know from outside the company.

Do tell if you can. Were you there?

This is a rather useless comment without elaborating.

I wish he did this with VR environment instead like they mention at the start of the slides. A VR environment with a JPEG camera filter, physics sim, noise, robot simulation. If anyone could program that well its him.

Using real life robots is going to be a huge bottleneck for training hours no matter what they do.

I still don’t think we have a clear enough idea of what a concept is to be able to think about AGI. And then being able to use concepts from one area to translate into another area, what is the process by which the brain combines and abstracts ideas into something new?

Known entities are recurring patterns (we give names to things that occur more than once, in the world or in our thoughts). Concepts are recurring thought patterns. Abstractions, relations, metaphors, are all ways of finding and transferring patterns from one domain to another.

Sure, I understand what the terminology means but I don't believe we get to AGI without some ability to translate the learning of say using a mouse to using a trackpad in a simple way. Humans make these translations all the while, you know how to use a new room and the items in it automatically but I personally see the systems we have built are currently very brittle when they see new environments because they can't simplify everything to its fundamentals and then extrapolate back to more complex tasks. You could train a human on using an Android phone and give them an iPhone and they would do pretty well, if you did this with modern machine learning systems you will get an extremely high error rate. Or say you train an model on how to use a sword, I'm not convinced it would know how to use and ax or pair of crutches as a weapon.

Maybe it will turn out to simply be enough artificial neurons and everything works. But I don't believe that.

Another thought experiment - if OpenAI AGI was right around the corner, why are they wasting time/money/energy buying a product-less vanity hardware startup run by Ive?

Why not tackle robotics if anything. Or really just be the best AGI and everyone will be knocking on your door to license it in their hardware/software stacks, you will print infinite money.

Or have your AGI design products

All the more reason not to acquihire Ive for $6.5B, if true

Does AGI necessarily mean super-genius? Was KITT AGI? I'm not sure he could design products?

Is VC really funding a trillion dollars of GPU purchases to replace labor that could instead be bid out to developing world mechanical turks for $1/hr?

No the term for that is ASI...artifical super intelligence. People in AI community have different timelines for that than AGI.

This line of thought doesn't work, because any company approaching AGI might be actively trying to hide that information from regulators and the military. Being the 1st AGI company is actually pretty risky.

VCs are far too conditioned as hype men to hide the ball like that.

After generations of boastful over-promising, do you really believe THIS time they are underpromising?

However their actual actions resemble companies who know AGI isn’t even on the horizon and moreso they are acting as exactly as if they believe the AI hype bubble is coming to an end and they need to dump the stuff into the public markets asap.

There really isn’t any other way to interpret OpenAI’s actions for the last few months.

Sure it could all be a feint to hide their amazing progress. Or it could be what it looks like.

Given the hype cycles of the last 20 years, I’m going with the second.

Name any other company acting like that besides OpenAI. Or any person besides Sam Altman, the guy who screwed up OpenAI's structure/board/funding. You just have a narrative you want to be true, and one company half fits into that narrative.

AGI is not enough. Seriously, imagine if they had an AGI in their ChatGPT interface. It’s not enough to do anything truly meaningful. It’s like a genius in the woods somewhere. For AGI to have an impact it needs to be everywhere.

Robotics to navigate the physical world seems more impactful than some pin/glasses product to provide a passive audio/visual interface to the chatbot doesn't seem so earth shattering either though.

What would you do with a 10x or 100x smarter Siri/Alexa? I still don't see my life changing.

Give me a robot that can legitimately do household errands like the dishes, laundry, etc.. now we are talking.

and build cities ...

First we need the agi government to do zoning & permitting reform

I think a big problem with AGI discussion is sometimes people aren’t being creative enough. A 100x smarter ChatGPT means biology immortality

> Seriously, imagine if they had an AGI in their ChatGPT interface. It’s not enough to do anything truly meaningful

If they had that people would make agents with it and then it can do tons of truly meaningful things.

People try to make agents with the current one but its really difficult since its not AGI.

Once AGI is accomplished I can’t imagine what else it would do but bootstrap itself up which, depending on compute, could scale quite far. OpenAI would only need to feed it compute for the most part.

I don’t think AGI is close, but once it happens it’s hard to imagine it not “escaping” (whenever we want to define that as).

They have a large robotics research group.

If Ive hype brings more investment valuation than it costs to acquire his company, and the time costs etc., it's a win for current investors regardless of if it is a good investment for the investors in the new round.

What Carmack is doing is right. More people need to get away from training their models just with words. AI need the physicality.

>AI need the physicality.

which i found interesting, because i remember Carmack saying simulated environments are way forward and physical environments are too impractical for developing AI

Yeah in that way this demo seemed gimmicky like he acknowledged. He said in the past he would almost count people out if they weren’t training RL in a virtual environment. I agree, still happy he’s staying on the path of online continual learning though

Nvidia seems to think the same thing. Here's Jim Fan talking about a "physical Turing test" and how embodied AI is the way forward.

https://www.youtube.com/watch?v=_2NijXqBESI

He also talks needing large amounts of compute to run the virtual environments where you'll be training embodied AI. Very much worth watching.

> More people need to get away from training their models just with words.

They started doing that a couple of years ago. The frontier "language" models are natively multimodal, trained on audio, text, video, images. That is all in the same model, not separate models stitched together. The inputs are tokenized and mapped into a shared embedding space.

Gemini, GPT-4o, Grok 3, Claude 3, Llama 4. These are all multimodal, not just "language models".

(If you know) how does that work?

Are the audio/video/images tokenized the same way as text and then fed in as a stream? Or is the training objective different than "predict next token"?

If the former, do you think there are limitations to "stream of tokens"? Or is that essentially how humans work? (Like I think of our input as many-dimensional. But maybe it is compressed to a stream of tokens in part of our perception layer.)

Ask Gemini to explain how it was trained

https://g.co/gemini/share/f64c3358d9fa

"... Since I am new to the research community, I made an effort" This means they've probably submitted a paper too.

It states it's a research, not a product company.

To be fair, OpenAI is also a "research lab" rather than "product company" and they still sell products for $200/month, not sure the distinction matters in practice much today as long as the entity is incorporated somehow.

That's what I said

Why would AGI choose to be embodied? We talk about creating a superior intelligence and having it drive our cars and clean our homes. The scenario in Dan Simmons' Hyperion seems much more plausible: we invent AGI and it disappears into the cloud and largely ignores us.

It doesn't need to be permanent. If humans could escape from their embodiment temporarily they would certainly do so. Being permanently bounded to a physical interface is definitely a disadvantage.

Looking at other examples in sci-fi, perhaps to stop my body from pressing its off-switch?

With distributed backups in place, AIs will be much less worried about self-preservation than we are.

anywhere we can watch the presentation ? the slides alone are great, but if he's saying stuff alongside, I'd be interested in that too :)

> It is worth trying out one of the many web based reaction time testers – you will find that you average over 160 milliseconds.

TIL JC has elite reflexes

And nerves of steel

I'm with OpenAI folks on this one: Atari just won't cut it for AGI. My layman intuition is that RL works well when rewards give good signal all the time. Until it does RL is basically random search. That's where massive data diversity like we have in text comes in handy.

In a game there might be a level with a door and a key, and because there's no reward for getting the key closer to the door, bridging this gap requires random search in a massive state space. But in the vast sea of scenarios that you can find in Common Crawl there's probably one, where you are 1 step from the key, and the key is 1 step from the door, so you get the reward signal from it without having to search an enormous state space.

You might say "but you have to search through the giant Common Crawl". Well yes, but while doing so you will get reward signal not just for the key and door problem, but for nearly every problem in the world.

The point is: pretraining teaches models to extract signal that can be used to explore solutions to hard search problems, and if you don't do that you are wasting your time enumerating giant state spaces.

You can actually easily test and overcome this by training a model simultaneously on a massive of text and Atari while carefully balancing learning rates between the two.

> Fundamentally, I believe in the importance of learning from a stream of interactive experience, as humans and animals do, which is quite different from the throw-everything-in-a-blender approach of pretraining an LLM. The blender approach can still be world-changingly valuable, but there are plenty of people advancing the state of the art there.

It's a shame that pretrained approach leads to such good enough result. The learning-from-experience, or what should be the "right" approach, will stagnate. I might be wrong, but it seems that aside from Carmack and a small team, "the world" is just not looking/investing on that side of the AI anymore.

However, I find it funny that Carmack is now researching for such approach. At the end of the day, he was the one who invented Portals, an algorithm to circumvent the need to reproduce the whole 3D world and therefore making 3D games computationally possible.

As a side note, I wonder what models are to come once we see the latest state of the art AI Video training technologies, in synch with the joystick movements from a real player. Maybe the results are so astonishing that even Carmack changes his mind on the subject.

EDIT::grammar & typos

i think you're overstating this. Yann LeCun (chief scientist at Meta) is firmly in this camp, and i think most companies trying to bring AI into the real world via some sort of robotics technology are thinking about and testing this approach.

Thank you. You are right, most likely the ones working in the field haven't switched. But the truth is that big bucks are in pretrained technologies. As Carmack himself said, "there are plenty of people advancing the state of the art there".

Humans had 500 million * 8670 hours of Pre-Training.

I don't get why Carmack would say things should be learned in hours or upper bounds it to human lifetime.

I think there's a difference between "should be learned" and "should be able to be learned" here.

I’m not necessarily convinced despite my human bias that it’s a superior mechanism. Humans work the way they do and learn the way they do in no small part because of biological limitations and a physical reality. It’s not clear that a virtual entity needs to face the same limitations, although clearly learning from feedback that’s not available to an AI is important. It is true though that humans are more energy efficient learners, but letting the AI experiment with the real world and get feedback that’s way may be the only missing piece rather than a problem with the “blender” approach.

> It's a shame that pretrained approach leads to such good enough result. The learning-from-experience, or what should be the "right" approach, will stagnate.

We’ll see. I’m skeptical that you’ll ever get novel theories like special and general relativity out of LLMs. For stuff like that I suspect you need the interactive learning approach, and perhaps more importantly, the ability to reject the current best theories and invent a replacement.

I'm surprised there's as much interest in looking at the structure/behavior of the biological brain, and less interest in considering the behavior of our vision system. Our brains are not CPUs, and our eyes are definitely not a grid of pixels with a fixed framerate.

Bro went his whole career and managed to somehow create a gig for himself where he gets the AI money while playing Atari. It's hard to increase the respect for someone who you already maxed out on but there we go. Carmack is a cool guy.

Although Carmack is the quintessential not-a-brogrammer.

I've never actually seen a brogrammer though, I've seen people who program only because they get money for it, I thought for a while those where it, but I'm not sure if I think they qualify either.

It means “programmer I don’t like”. Very versatile insult and vague enough that it’s impossible to defend against.

No, I'm pretty sure it's just a programmer that understands the world in terms of bros.

"in the terms of bros" what does that even mean? I think bro is a term that's used pretty widely, for different things, in different cultures and contexts, I call my brother bro.. I've heard people call their friends bro.. I've heard someone tell a police officer "don't tase me, bro"..

My co-worker's 7 year old daughter calls him "bruh"

I call my 4yo daughter bro

From watching on the wall I've seen brogrammer used in various contexts (this is not an exhaustive list):

- Someone who is a programmer but follows a hypermasculine cliche and makes sure everyone knows about it.

- An insult used by other developers for someone who is more physically fit or interested in their health than themselves.

- An insult used by engineers or other people who are not happy with the over representation of men in the industry. So everyone is lumped in the category.

- Someone who is obsessed with the technology and trying to grind their skills on it to an excessive level.

> Someone who is obsessed with the technology and trying to grind their skills on it to an excessive level

sounds like a person who respects their own profession though

> someone who is more physically fit or interested in their health than themselves

Isn't "interested in their health" a signal that they are interested in themselves, rather than the opposite?

Disambigation: I believe "themselves" refers to the one insulting, not the one interested in their health.

It tripped me up too, to be fair.

Apologies, grammar is hard.

A. A male programmer who uses the word "bro" unironically in conversation.

B. A person who is physically and culturally indistinguishable from A

What physical or cultural characteristics would make a person "indistinguishable from A"?

Brogrammer here. I recognize that I spend 12 hours a day writing code (and loving it) as a fun thing but also a danger if I do it all sitting down. I stay busy and incorporate workouts into my day.

I don’t take it as a pejorative, it’s an acknowledgement of my efforts to be even considered in this category. For those wondering I have a family, and have healthy activities otherwise. No cool diets or bioscience, just code, physical activity and coffee/water.

This isn’t a lifestyle I’m saying everyone should do, only that people should do what makes them happiest and most fulfilled for their set of goals.

I always think of “bro” as being a hyper version of “dude.” It’s generically applied to any random person, but it’s also exclusively male. So using it implies “this ingroup is assumed to be 100% male.”

On the other hand, I’ve seen and heard “dude” and “guy” used by and applied to women by other women. Not common but it happens. But I’ve never heard “bro” used that way.

Expression of endearment in this case.

Honestly, having gone through the slides, it's a bit painful to see Carmack "rediscover" stuff I've learned in a reinforcement learning lecture like ten years ago.

But don't get me wrong! Since this is a long-term research endeavor of his, I believe really starting from the basics is good for him and will empower him to bring something new to the table eventually.

I'm surprised though that he "only" came so far as of now. Maybe my slight idolization of Carmack made me kinda of blind to the fact that this kind of research is a mean beast after all and there is a reason that huuuuge research labs dump countless of man-decades into this kind of stuff with no guaranteed breakthroughs.

https://x.com/unixpickle/status/1925795730150527191

I was just going to answer https://news.ycombinator.com/item?id=44071595 who mentioned exactly the same tweet.

I'm nowhere as good at my craft as someone who works for openai, which the author of that tweet seems to be, but if even I can see this, then it's bad, isn't it?

I find it interesting that he dismisses LLMs.

I would argue that if he wants to do AGI through RL, a LLM could be a perfect teacher or oracle.

After all i'm not walking around as a human and not having guidance. It should/could make RL a lot faster leveraging this.

My logical part / RL part does need the 'database'/fact part and my facts are trying to be as logical as possible but its just not.

IMO, he's right. LLMs can't be AI because they don't create a model of observations to predict things, they just imitate observations based on their likeness to each other. When you play Quake, you create a simple model of the game physics and use that fast model to navigate the game. Your equivalent of LLM has a role too: it's a fuzzy detector of things you encounter in the game, sounds, images and symbols, but once detected, those things are fed into the fast and rigid physics model.

Yes but the LLM could tell the physics system that it is physics related.

Hey look you see a stone falling

A lot of the problems John mentioned (camera jpeg, latency, real time decisions) have been worked on by comma.ai for many years. He could have just used their stack and build on it the general learning parts that comma is not focusing on.

And plenty of other people.

It's still a lot better to really learn and discover it yourself to really get it.

Also it's hard to determine how much time someone spent on particular topic.

Reuse is not always necessary - sometimes things are just done for fun and exploration, not for appeasing thirsty VCs and grabbing that market share.

Reinventing the exact same thing and shouting from the rooftops about it is exactly how you appease thirsty VCs and grab that market share.

Might I interest you in my new startup, that is a bus, but with technology?

Why would John Carmack who is so rich that he does things for shits, giggles and personal development, give a hoot what a VC cares about?

This is Carmack, who builds new things. He built one of the first 3D game engines based on the hard math.

Carmack himself has done a lot of work on end to end latency, during his time at oculus.

I feel top level AI creation is beyond his skill set.

He’s a AAA software engineer but the prerequisites to build out cutting edge AI require deep formal math that is beyond his education and years at this point.

Nothing to stop him playing around with AI models though.

There is no deep formal math in AI. It is a game of numbers.

All deep formal math is a boundary to a thing.

The formal math takes a few months to learn. He is more than smart enough to figure that out.

I think you overestimate the level of math required in AI and at the same time I think you underestimate the math skills of John. AI runs on GPUs, Quake 2 engine was one of the first to optimized for GPUs (OpenGL).

I'm pretty excited to see him in this domain. I think he'll focus on some DeepSeek style improvements.

this.

Having JC focusing on, say, writing a performant OSS CUDA replacement could be bigger than any of the last 20 announcements from openai/goggle/deepmind/etc

This feels like an understatement. At the time, young me had the impression Carmack came first, then the industry created 3dfx/OpenGL to run his games better. I still have nothing but respect for his skills decades later.

GLQuake was released 11 months before Quake 2.

What do you mean "beyond his skill set"? He effectively invented 3D gaming, which led to major leaps and investments into graphics cards which are now used for cryptocurrency and AI. He also did significant contributions into VR.

He's probably one of the most qualified people around.

He definitely did not invent 3D, he just pioneered it for PC gaming and got PC gaming up to speed to console gaming

The math around machine learning is very manageable, and a lot of research in that area is throwing heuristics at a wall to see what sticks

John Carmack:

So I asked Ilya, their chief scientist, for a reading list. This is my path, my way of doing things: give me a stack of all the stuff I need to know to actually be relevant in this space.

And he gave me a list of like 40 research papers and said, 'If you really learn all of these, you'll know 90% of what matters today! And I did. I plowed through all those things and it all started sorting out in my head.

What this misses is that research is a competitive endeavor. To succeed as a researcher you don’t just need to know the bare minimum required to do research in your field. You need to be able to do it better than most of the people you’re competing against. I know that HN as a collective has near-unlimited faith in Carmack’s abilities (and he is no doubt Very Smart). But he’s competing with other Very Smart people who have decades more experience of AI research.

To put it another way, the idea that John Carmack is going to do groundbreaking research in AI is roughly as plausible as the idea that Yann LeCun is going to make a successful AAA video game. Stranger things have happened, but I won’t be holding my breath.

Why does it need to be competitive? Maybe the guy has enough money to let him do whatever he wants regardless of the outcome and he chose AI because it's interesting to him.

In research you have to succeed before your competitors. It’s not research if it’s already been done.

You're forgetting that a whole string of breakthroughs are all fairly recent (like in the last decade). Everyone, including the pro's, is treading new ground.

In that context anyone can make progress in the field, as long as they understand what they're dealing with.

Better regard mr. Carmack as an X factor. Maybe the experts will leave him in the dust. Or maybe he'll come up with something that none of the experts cared to look into.

Lots of scientific fields have seen breakthroughs in the past decade. Doesn’t mean that any random smart person can jump in and start doing groundbreaking research.

But a random smart person will jump in and make groundbreaking research.

The difference is Carmack is literally a T-shaped dude -- hell, he's a T-shaped dude with lots of vertical lines :P

I believe all his in-depth experience in other areas will heavily unlock him to bring about another breakthrough. He's that good.

There are lots of less famous researchers toiling away who are also very smart and who have much more detailed specialist knowledge. I don’t doubt John Carmack’s intellectual ability, but he’s famous because he made some cool video games, not because he won the America’s Smartest Person competition. This idea that he’s the AI messiah seems almost cultish.

I'd love to have a copy of that list. Just to see how much I've yet to absorb.

https://aman.ai/primers/ai/top-30-papers/

AI creation is more like Alchemy than science, and breakthroughs come not from math background, but from intuition and a bit of math skills. Transformers behind the chatbots isn't a rocket science and were discovered almost by accident. The next breakthrough will come a similar way. I'd frankly bet on someone like Carmack than on some theoretical researcher who is churning out papers.

Reply on news.ycombinator.com