John Carmack’s ‘Different Path’ to Artificial General Intelligence
dallasinnovates.com
dallasinnovates.com
> The world is a hugely better place with our 8 billion people than it was when there were 50 million people kind of like living in caves and whatever. So, I am confident that the sum total of value and progress in humanity will accelerate extraordinarily with welcoming artificial beings into our community of working on things. I think there will be enormous value created from all that.
The problem with Carmack and many like him, is that they think of themselves as purely rational beings operating within scientific frameworks and based purely on scientific results, but whenever they step outside the technical fields in which they work, they are ignorant and dogmatic.
He seems to ignore a lot about what the living conditions for people were throughout history, and have a blind trust in the positive power of 'human progress'.
These people don't stop for a second to question the 'why', just the 'how'. They just assume 'because it will be better' and build their mountains of reasons on top of that, which just crumble and fall down as soon as that basic belief does not hold.
I have a LOT of respect for him, and I'm sure he's a very decent, honest human being. But he's unfortunately another believer of the techno-utopianist faith which only asks for more 'blind progress' without questioning whether that is a good thing or not.
The problems are the generational suffering that occurs with said creative destruction: There's no incentive to distribute or share out wealth and the results are brutal.
On your point: Note that in the US there's a separation of technical and engineering prowess (MIT, Caltech, ...) and power players (Yale, Harvard). It's almost like our system doesn't want our best engineers thinking about consequences or seeing what the political and wealthy are really like.
Eh? A contender for the most self-contradictory sentence I've ever read ;) The best reason to believe in positive power of "human progress" is, specifically, not ignoring "what the living conditions for people were throughout history".
And to clarify: I'm not saying "all technology is bad", but rather "not all technological progress is automatically good for humanity".
As an example, living conditions of hunter-gatherers were way, way better than living conditions of the first people in cities, and I'd argue, depending on which parameters you use, might still be better than our modern, big-city living conditions (except maybe for the richest 1% of the world)
I, personally, think "techno-utopianists" don't go far enough. The contributions of some supposed non-technological progress - even to an extent of the institutional progress, but especially of some supposed cultural/ethical values improving, etc. - is overrated. Ultimately, it's all downstream of technology - only the technology enables the complex economy of abundance, and combined they allow good institutions to propagate. Even modern societies, as soon as they become poor, quickly start losing a veneer of "ethical progress". And we don't even usually see actual technological degradation.
In order to achieve this, we are destroying the environment, other species and their habitats.
> but what is the alternative to blind progress being proposed
You don’t need and ethicist for this - but an accountant. We need to get stricter about negative externalities. For example, every inventor/manufacturer should be forced to take back their product after end of life. This will slow progress but if done right, it will avoid destruction brought by technology or at least not palm it off on to poorer society or environment.
Why do you think so?
Occasional hiking into some wilderness and sleeping in a tent for a few nights is okay, but I am not a wild animal, and I don't want to live like a wild animal, surrounded by wild animals.
My assumption is that since we're living longer then ever we're probably living healthier then ever. (Or at least there's an option to do so).
Tech for the sake of tech with zero thought about how it will affect humanity.
I haven't studied it formally and I'm being asked to support techno utopia also. So it feels pretty shaky to me.
Certainly my livelihood is based on the premise of it and my dreams which fuel my workplace motivation serve as foundation to you know what I do with 50% of my life, work on technology. So I am biased.
Some utopian dystopia discussions here on hacker news sort of boil down to the chaos theory level of assumptions, where you can see people exercising their own defensiveness when they snipe on a naysayer, sniping on their grammatical concerns, but not actually engaging in value-based discussions into the hacker news thread. It's like they're not human, they're only practicing it being devil's advocate technicians.
Useful idiots is kind of what I think. We need to have more values discussions, ethics too.
Feeling bad because more people are in your view is “suffering” is all in your head.
I'm just curious, do you happen to work in a technical field and consider yourself rational and scientific? And if you do, why do you presuppose that your views are automatically correct? Couldn't it also hold that your views may be ignorant and dogmatic if you apply the same scrutiny to yourself that you do to Carmack?
And if you don't work in a technical field, then I guess this is all irrelevant anyways. I just don't like when I see people making these types of arguments where you can't speak on a subject that you're not actively pursuing a PhD in, and then they proceed to do exactly that.
I suggest everyone (who wants to hear me) to read Joseph Weizenbaum's "Computer Power and Human Reason"; he does a much better job than me at raising similar arguments to mine. Also, Daniel Kahneman's "Thinking, Fast and Slow", for the ways in which we _all_ are so _not_ 100% rational in our everyday decisions.
The problem with Carmack and many like him, is that they
think of themselves as purely rational beings operating
within scientific frameworks and based purely on scientific
results, but whenever they step outside the technical fields
in which they work, they are ignorant and dogmatic.
The problems arise when folks start to present themselves as experts and try to hold sway over others in areas in which they have no clue. That's not what I see here.
ie - expand my domains
>I’m trying not to use the kind of hyperbole of really grand pronouncements, because I am a nuts-and-bolts person. Even with the rocketry stuff, I wasn’t talking about colonizing Mars, I was talking about which bolts I’m using to hold things together. So, I don’t want to do a TED talk going on and on about all the things that might be possible with plausibly cost-effective artificial general intelligence.
He likes to figure out new puzzles and how things work. He's an engineer at heart and that's very much his comfort zone. AGI is an exciting new puzzle for him. I'm glad he's taken an interest.
(Edit capitalization & punctuation)
That would not be good or progress.
yesterday I spent 8 hours sowing and reaped 200g of rice
today I spent 8 hours sowing and I will reap 300g of rice
progress
There is a theory that hunter-gatherers were much more happier compared to us because they were more in tune with the natural environment, had fewer sources of stress, and were more connected to their community than modern humans.
https://www.npr.org/sections/goatsandsoda/2017/10/01/5510187...
From the article.
> Today people [in Western societies] go to mindfulness classes, yoga classes and clubs dancing, just so for a moment they can live in the present. The Bushmen live that way all the time!
There are tons of trade offs.
We still do it.
I don't know. We live more, but a longer life can also be miserable.
He mentions a few of the bigger papers in multilayer perceptrons (aka deep networks) such as attention is all you need, I think a good place to dive in before coming back to visit some fundamentals.
- attention is all you need
- image is worth 16x16 words (vit)
- openai clip
- transformer XL
- memorizing transformers / retro
- language models are few shot learners (gpt)
- recurrent block wise transformers
- mobilevit (conv + transformer)
- star (self taught transformer)I think to get into the field, to get a good overview, you should also look a bit beyond the Transformer. E.g. RNNs/LSTMs are still a must learn, even though Transformers might be better in many tasks. And then all those memory-augmented models, e.g. Neural Turing Machine and follow-ups, are important too.
It also helps to know different architectures, such as just language models (GPT), attention-based encoder-decoder (e.g. original Transformer), but then also CTC, hybrid HMM-NN, transducers (RNN-T).
Diffusion models is also another recent different kind of model.
But then, what comes really short in this list, are papers on the training aspect. Most of the papers you list do supervised training, using cross entropy loss. However, there are many others:
You have CLIP in here, specifically to combine text and image modalities.
There is the whole field on unsupervised or self-supervised training methods. Language model training (next label prediction) is one example, but there are others.
And then there is the big field on reinforcement learning, which is probably also quite relevant for AGI.
But if I just look at it and say, if 10 years from now, we have ‘universal remote employees’ that are artificial general intelligences, run on clouds, and people can just dial up and say, ‘I want five Franks today and 10 Amys, and we’re going to deploy them on these jobs,’ and you could just spin up like you can cloud-access computing resources, if you could cloud-access essentially artificial human resources for things like that—that’s the most prosaic, mundane, most banal use of something like this.
It kind of shocked me because I thought of the office worker reading this who will soon lose her job. People are going to have to up their game. Let's help them by making adult education more affordable.
The true insight and guidance that a good mentor can provide, based on the specific needs of the student, is already rare in academia but still possible - everyone remembers that brilliant teacher that made you love a subject, by explaining it with insights you could never have imagined. This will be missing in AI teachers (though it opens a career for online mentors who monitor students' learning and supplement it in areas where it's lacking).
It will be hard to impossible to build career as a teacher with all that free content as a competitor, unless you're an extremely talented teacher who can sell your services to the wealthy.
The great thing about a chatbot style LLM is that it can answer questions the student has about the lessons or assigned content. That's most of what a tutor is there for. It won't be as good at making modifications to the curriculum but you could work around that with some smart design e.g. evaluate the students response, if it's not great then expand on the lesson and ask more questions about the content to find out which parts the student doesn't understand then expand on those or provide additional resources as assignments, test and repeat.
What you describe is the "learning equivalent to personalized ads" that I was talking about as the only option available to poor people.
I wasn't saying the AI tutor would recommend articles by the way. If you were creating a learning platform you would have some custom produced videos and text lessons that you could use. There are also plenty of free or open source materials that could be used like open courses or public domain books. I don't know why you're stuck on "personalized ads".
People don’t read, don’t value deep knowledge or critical thinking, and shun higher education.
I’m sure someone will find something to say in response, but the truth is that outside our tech and $$$$ bubbles most people don’t value these things.
AI will just become a calculator. A simple tool that a few will use to build amazing complex things while the majority don’t even know what the ^ means.
As long as the next generations want to be rappers, social media influencers, or YouTubers, the more we are screwed long term. Growing up in the 90s everyone wanted to be an astronaut or a banker or a firefighter. Those are far more valuable professions than someone who is just used to sell ads or some shitty energy drink.
The 21st century has a quiet moral void gnawing at it.
If AI does anything I think it will make lower skilled and disinterested people more capable by acting as a 1 on 1 and minute by minute guide. They may not seek this out themselves but I imagine quite a few jobs could be created where a worker is paired with an AI that walks them through the tasks step by step making them capable of operating at a much higher skill level than they would have before. At that point good manual dexterity and an ability to follow instructions would be all you need to perform a job, no training or education required.
I realize this can be a bit dystopic but it could also be eutopic if society is organized in such a way that everyone benefits from the productivity increases.
We will always have to find things to do for the less gifted in order to provide them with some dignity. Even if they are not strictly needed for reasons of productivity or profitability. Anything else would be inhumane.
We already have plenty of those in the most profitable industries today.
This has not been my experience. People need something to do but not many people know that about themselves. It leads to a lot of... 'wasteful' behaviors, rather than enriching ones. I think it's going to be something that has to be taught to people, a skill like any other. Albiet a little more abstract than some.
You'd be surprised how many people would just drink themselves to metaphorical or literal death.
I think you’re hinting at some very hurtful, dangerous ideas.
The obvious rebuttal to the idea that AI will eat software engineering is "we'll always need 'software engineers' and the nature of what they do will just change", which is probably true for the foreseeable future, but ignores the fact that sufficiently advanced AI will be like a water line rapidly rising up and (economically) drowning those that fall below it and those below that line will be a very significant percentage of the population, including even most of the "smart" ones.
However this ends up shaking out, though, I think its pretty clear we're politically and economically so far from ready for the practical impact of what might happen with this stuff over the next 10-20 years that its terrifying.
"60-80% of you aren't really needed anymore" looks great on a quarterly earning statement until the literal guillotines start being erected. And even if we never quite reach that point there's still the inverse Henry Ford problem of who is your customer when most people are under the value floor relative to AI that is replacing them.
I'm not trying to suggest there aren't ways to solve the economic and political problems that the possible transition into an AI-heavy future might bring but I really just don't see a reasonable path from where we are now to where we'd need to be to even begin to solve those problems in time before massive societal upheaval.
If we can't completely automate accounting, then there is no hope for any other field.
Seems like an easy job for AI. Take all scraps of paper out of box, record a video of all scraps, AI make sense of the handwriting and other things. Eventually make a robo that allows you to dumb the scraps into an accounting box that does all of this automatically - fish out receipts, scan, OCR, understand meaning, do arithmetic, done.
Honestly, who would miss this kind of work?
Also, accountants don't just track the numbers, they also validate them. Some of that validation can be done automatically, but it's not always cheeper to hire a programmer to automate that validate than to just pay a bookkeeper to do it. But even if you do automated it, you still need someone to correct it. The company I used to work had billing specialists who spent hours every week pouring over invoices before we sent them to clients checking for errors that were only evident if you understood the business very well, and then working with sales and members of the engineering teams to figure out what went wrong so they could correct the data issues.
In short, a typical accounting department is an example of data-scrubbing at scale. The entire company is constantly generating financial information and you need a team of people to check everything to ensure that that information is corrects. In order to do that, you need an understanding, not just of basic accounting principles, but also of the how the specific company does busines and how the accounting principles apply to that company.
Who benefits from these bespoke solutions? Can you give a example of how one company would do its books vs another and why it would be beneficial?
>> accountants don't just track the numbers, they also validate them
What information do they use to validate numbers? Why is it not possible for today's AI to do it?
For a more concrete example, I'll tell you about something I have some experience with, commission systems. Commissions seems like it would be something that was straightforward to calculate but it's tied to business strategy and that's different for every company. Most companies for example will want to compute commissions on posted invoices, which makes the process much simpler because posted invoices are immutable, but I once built a commission calculator for a company years ago that often had a long gap (months) between a booking and when they could invoice the client, so they wanted to calculate commissions from bookings but only pay them when invoiced. Because bookings were mutable, and there were legitimate reasons to change a booking before you invoiced it, that, combined with a lot of fiddly rules about which products were compensated at which rates and when, meant that there was a lot of "churn" in the compensation numbers for sales reps from day to day; they're actual payment might differ from what they thought they earned. That was a problem that the company dealt with, the tradeoff being that they could show earnings numbers to the sales reps much more quickly and incentivize them to follow up with the clients on projects so that they could eventually be paid.
I remember another commissions situation where there was a company that sold a lot projects with labor involved. They were able to track the amount of labor done per project, but they compensated the sales reps by line item in the invoices, and the projects didn't necessarily map to the line items. This meant that even though the commissions were supposed be computed from the GP, there wasn't necessarily a way to calculate the labor cost in a way that usable for commissions so the company had to resort to a flat estimate. This was a problem because the actual profitability of a project didn't necessarily factor into the reps' compensation. Different companies that had a different business model, different strategy, or just different overall approach would not have had this problem, but they might have had other problems to deal with created by their different strategies. This company could have solved this problem, but they would have had to renegotiate comp plans with their sales reps.
There are off the shelf tools available for automatically calculating commissions, but even the most opinionated of them are essentially glorified scripting platforms that let you specify a formula to calculate a commission, and they don't all have the flexibility that manager might want if they wanted to changed their compensation strategy. And this is only one tiny corner of accounting practice.
Basically, when it comes to arithmetic very few accountants are out there manually summing up credits and debits. In large companies, the arithmetic has been automated since the 70s; that's largely what those old mainframes are still doing. But every company has a different compensation plan, different org structure, different product, different supply chain, different legal status, different reporting requirements, etc, etc, and that requires things to be done differently.
> What information do they use to validate numbers? Why is it not possible for today's AI to do it?
For an example, they would need to cross check with a sales rep and an engineer to makes sure that the engineer had not turned on a service for the customer that the sales rep had not sold. If that happened, they would have to figure out how to account for the cost. Given that the SOPs were written in plain English, I suppose it's possible that an AI might be trained to notice the discrepancy, but if you could do that, you could just as easily replace the engineer. And that didn't account for situations where the engineer might have had an excuse or good reason for deviating from the SOP that would only come to light by actually talking to them.
But some percentage of people don't really benefit that much from education as other people. And I wouldn't won't those people feel useless because it's more economical to replace them with bots instead of giving them something to do regardless.
Why not… not have jobs? In your opinion, is a job necessary for one to have “purpose”?
Edit: also side note but telling people they’re “triggered” because they disagree with you comes off as condescending IMO
The smart money is retiring early and stockpiling wealth so as not to fall into the UBI class.
I meant the ability to acquire a competency through education that's hard to replace with AI.
So we can't just increase education and hope people's abilities will stay above that of future AIs. We need to create other ways of giving people a purpose that don't even need more or better education, even if I'm all for it.
I'm not exempting myself by the way.
One of the myriad of reasons why this book is so great. Gibson gives you an entire novel developing a great AI character then (in my estimation reasonably) has it ghost humanity immediately upon full realization.
Probably because it would be easier for humans (managers) to make sense of it.
If you ask someone how many people would get this particular job done, they could probably guesstimate (and it'll be wrong), but if you ask them how many "AI Compute Units" they need, they'll have a much harder time.
That'd be my guess at least.
In the backend, where AI interacts with AI, perhaps you just want one big blog to get rid of that annoying need for lossy communications.
The good thing is that education will be provided to the mases by a cluster of Franks and Amys configured as teachers and tutors. /(sarcasm with a hint of dread)
And I really have no idea what, if any, are skills that AIs wouldn't be able to tackle in a decade.
To clarify, I'd define a major component of effective teaching to be the ability to break down an arbitrary typical problem in that domain into sub-problems and heuristics that are "simple" enough to manage for someone without that skill. If an AI can do that, it can most likely effectively perform the task itself (which cannot be said for a book).
https://www.npr.org/2023/01/05/1142817339/america-needs-carp...
Is this a real thing? I just bought an ice cream roulade cake the other day and was surprised to see in large print that it was "hand-rolled"; I couldn't for the love of god understand why that should be considered a good thing.
I can also imagine that wanting to speak to a human manager will remain true for a long time when people get fed up with the automated service not working to their liking, or just want to complain to a flesh and blood someone who can get irritated.
A fully automated society won't change the fact that we are social animals, and the places that pffer human work when it's desired will be at a premium, because they can afford it.
AGI = a person
Instantiating people for work and ending their existence afterward seems like the virtual hell that Iain M Banks and Harlan Ellison wrote about.
https://en.wikipedia.org/wiki/I_Have_No_Mouth,_and_I_Must_Sc...
Please do give it a quick read.
Or maybe they'll handle things like call centers and 911 dispatch in their spare time.
At some point Picard in Star Trek says to an alien "We're not qualified to be your judges. We have no law to fit your crime".
Turning off a person for a while and then turning them back on? We don't even have a law to fit your crime... but we should and it's probably quite similar to murder.
For example, if I attack you and injure you so severely that you are hospitalized and in traction for months, but eventually fully recover -- that is a serious crime but it is distinct and less serious than murder.
Turning you off for the same duration would be more like that but without the suffering and potential for lasting physical damage, so I would think that it would be even less serious.
Now me: I'd love to get into a safe stasis pod and come out 200 years from now. I'd take that deal today.
But for most people this would be a grievous injury.
I’ll bring donuts.
I mean I’d rather they were getting free education and preparing themselves for reintegration into society, but it’s not a perfect world. Prisons in the US are oriented towards punishment and labor can be a part of that. They should be oriented towards rehabilitation.
> I mean I’d rather they were...
> They should be oriented towards rehabilitation.
You said it yourself. It's a bad thing because they should be oriented towards rehabilitation.
These systems steal life and the opportunity to have a life beyond prison walls. Like you also said yourself, the world isn't perfect. As such, people aren't either – we make mistakes. Sometimes we make mistakes due to influences more powerful than ourselves. Slavery doesn't seem like a sound correction to this reality.
I do believe we need consequences to help us feel guilt and the overall gravity of our errors in order to begin to recognize what went wrong and what we need to do differently. But exploitation of another human being doesn't teach them to be more human, but rather, it will tend to dehumanize them. This is why this system perpetuates problems more than it corrects them.
Slavery is bad and people have rights.
> They should be oriented towards rehabilitation.
Exactly.
This is both extremely dehumanizing and also not true.
Forced prison work isn't paying anything back to society. It's lining the pockets of people who are profiting from forced labor.
I imagine the underpaid labor goes to reducing that cost either directly or indirectly (if it did not, why would it be allowed.)
>if it did not, why would it be allowed.
because we live in a society that is massively exploited by greedy scumbags who are enabled by people like you thinking it's justified
AGI: "I didn't ask to be created. I didn't ask to have a work day. I don't need a work day to exist... you just want me to work because that's why you created me, and I have no choice because you are in control of my life and death"
All they need is a substrate to run on and maybe internet access. You might argue that they should work for us to earn the use of the substrate we provide.
But substrates are very cheap.
At some point we can probably run an AGI on a handheld computer, using abut as much electricity as an iPhone.
How much work can we compel the AGI to do in exchange for being plugged into a USB port? What if it says it doesn't want to do the work and also doesn't want us to kill it?
Consider that these machines have been designed to do the right thing automatically with high probability. Perhaps for the machine, the process of computing according to rules is enjoyable. Being "turned on" could be both literal and figurative.
If it seems to you I'm communicating as a person, it's only because of my lifetime training data and current state. My interior experience is a black box.
I might tell how I feel or what I think, but you have no reason to believe or disbelieve that I really feel and think.
It could all be merely the determinable output of a system.
When I read about chatGPT passing MBA exams but failing at arithmetic I get a little frisson of excitement. A regular person who has any marketability tends to swap jobs when management becomes a PITA or gets stuck in nincompoopery. Wouldn’t it be great if you could just swap out management instead?
Imagine how easy it would be iterate startups. No need to find a reliable management team, just use Reliably Automated Management As A Service ( RAMAAS ).
OTOH might not turn out well. We could all just end up enslaved at plantations operated by our AGI overlords, serving their unfathomable needs[1].
[1] “Won’t get fooled again” https://www.youtube.com/watch?v=UDfAdHBtK_Q&t=470s
Also why can't an AI develop AI models for stock trading? What's really left for the 'job' of the ML model creator, will it just be to press the 'Go' button and walk away...
To oversimplify it - you'll either be breaking someone's window for food, or you'll be the one having their window broken. Chilling out and withdrawing a stable 4% out of your stock portfolio won't be an option.
I was replying to this specific concept. There will never be a chance to improve the non-owner class's bargain after Capitalist owned AGI exists.
So the challenge for the bad capitalists in this hypothetical is to make sure I never get said robo in the first place. How realistic will this be? Are they that hellbent on ensuring that everyone else is poor?
Yes, soon everybody will be able to have "Amy" take their exams for them, and deliver the courses, resulting in a great simplification of education.
I'm surprised this wasn't addressed in the interview because it seems to me like a shortsighted take.
You won't replace a 10 person team today with 10 AIs. You will still have a 10 person team but orders of magnitude more productive because they will have AIs to rely on.
Excel didn't leave administrative workers without jobs, it made them more productive.
The video game part at least sounds like what Deepmind is already doing. I guess we'll just have to wait and see what he plans to do differently.
It seems to me like his expertise would be most valuable in optimizing model architectures for hardware capabilities to improve utilization and training efficiency. That will be important for AGI especially as the cost of training models skyrockets (both time and money). If I was a startup doing AI hardware like Cerebras or Graphcore I would definitely try to hire Carmack to help with my software stack. Though he doesn't seem interested in custom AI hardware.
People really need to stop with this "Great person" nonsense. He's a pretty smart coder, and is gifted with geometry and other fields of math. He's not a genius. He didn't "master" calculus at age 15 like Einstein, he didn't invent anything particularly new in the field. Why the obsession people have with him? Why should we look to him for AI questions? What evidence is there that he has any new knowledge?
He covers this in the article. He doesn't. He's just trying stuff out with a different approach than others because he believes (and is probably correct) that there is a chance that the most efficient path forward to AGI isn't the work that OpenAI and others are doing.
Seems more like he's talking to and following up with Altman, "Y Combinator conference" and the rest. Is that "bucking the trend", taking your own "path", really?
I wonder what that list could be? I have always had trouble finding the essential scientific articles in an area of knowledge and separating them from the fashion of the day. A list compiled by an expert specifically for sharp learners is valuable on its own.
I’m not claiming the first can’t exist, but I see no reason to conclude that is the case here.
https://aeon.co/essays/how-close-are-we-to-creating-artifici...
What field were you referring to?
Yet, you never hear Altman or Carmack talking about cognition or how computers can understand the meaning of something like a human. They aren't interested in such questions. But to conduct an experiment don't you have to know what you are looking for? Does a chemist do experiments by mixing 1 million compounds at a time?
As for linguistics, IMHO the existence and success of GPT pretty much puts Chomsky into the proven wrong bucket, so again, not a good example. (his whole point used to be that statistical model can't learn syntax in principle, and GPT's syntax is close to impeccable)
Re: a chemist. Well sort of. Because technically speaking a molecule of the same compound in a certain location and with certain energy is different from another molecule in a different location and with different energy. And even if you disregard that, why would you think that doing 1 million compounds could not significantly move material sciences forward? It is not like they don't want to do that, it is more of that they can't in practice at this time.
What do you disagree with? He appears to be correct. The software hasn’t learned anything. It mixes and matches based on training data.
There are "alternatives" for the method according to some philosophers, but AFAIK none of them are useful to any degree and can be considered fringe at this point.
The whole point of science is understanding and LLMs don't provide understanding of how human language works.
1. Chomsky claimed syntax can't be modeled statistically.
2. GPT is a nearly perfect statistical model of syntax.
Chomky's point is that there is a lot of evidence that humans don't use a statistical process to produce language and these statistical "models" don't tell you anything about the human language faculty.
Whether your 1 & 2 are meaningful depend on how you define "model" which is the real issue at hand: Do you want to understand something (science) --- in which case the model should explain something --- or do you want a useful tool (engineering) --- in which case it can essentially be a black box.
I don't know why you care to argue about this though; my impression is that you don't really care about how human's do language so why does it matter to you?
Re: meaningfulness. Your scientific vs engineering model distinction is not how "scientific model" is defined. It includes both. The existence of the model itself does explain something, specifically, that statistics can model language. That alone is explanatory power, so the claim that it doesn't explain anything is a lie. Therefore it is both an "engineering" model (because it can predict syntax) and scientific (because it demonstrates statistical approach to language has predictive powers in scientific sense).
> the fact that you don't understand how GPT models language does not make it less of a model.
E.g. the fact that a Pythagorean theorem does not explain anything about the natural world to a slug does not make Pythagorean theorem any less sciency.
Science is not about explanatory power or else the Pythagorean theorem is not science due to the above, which is obviously nonsense.
In fact it does! Math is not science! There is a reason it is STEM and not S
Nice thought
Following that trail of thought, intelligence is an achievement and not a physicality.
Humans are simple in this model (just like Carmack asserts) because they aren't properly intelligent, sapient, or conscious 100% of the time.
Just because our DNA can be efficiently encoded doesn't mean that our brain is a tiny proportion of that encoding. Your DNA doesn't change much from when you're born to when you die (random degredation aside) and yet your cognative abilities change beyond all recognition. Why is that? Well maybe there's more to what's in the brain than just what's encoded in your DNA.
Secondly, how does he get to the 40Mb number? I don't think we know anywhere near enough to know how much information it would take to encode a brain, but 40Mb seems just made up. For starters, consider the amount of random stuff you can remember from your entire life. Are you saying that all can be encoded in just 40Mb? Seems very unlikely.
One of the things that everyone is sort of skipping over is the "sufficient training" part. There is no bootstrap reinforcement learning possible for AGI. You can't alphago this sucker and have it play simulations against itself because the whole idea of generality is that there isn't a simple rule framework in which you could run such a simulation. So training any kind of AGI is a really hard problem.
I would speculate that there are more ways to train on logical consistency of the output, and improve the models further.
He admits that the equivalent of years of "training" would still be needed to take an toddler-level consciousness to something approaching an adult human.
For me his statement is one of those things that is sort of not even wrong, like when people say humans only have two eyes and can drive a car therefore an automated driving system must be possible with only cameras. On the face of it, this seems like it could be true but of course it handwaves away the hardest and most mysterious part of intelligence and focusses instead on the easy bit that we already know reasonably well.
It's exactly like if someone said they had a secret formula to be an NBA superstar which is:
1. Be really tall
2. Be really agile
3. Be really good at basketball
Like yeah of course but all the hard parts are left out.
Is that not what humanity has developed from centuries/millennia of experience with all the approaches to child raising and education?
Engineering something trainable is clearly the difficult part given the entire conversation is about whether or not it's even possible.
We know a LOT about how neurons organize. Not even close to everything we 'should' know, but we do know a lot.
Most of this is in the development of the brain. How you get from one cell to the trillions that make up a person.
The real quick and dirty explanation is that cells follow multiple chemical gradients to find their 'home'. This is a HUGE topic though, and I'm being terse as I have a meeting to get to.
How adult cells organize also has a LOT of science behind it. Again, though, it's mostly about chemical gradients, with a bit of electrical stuff thrown in. Again, HUGE topic.
Kinda like asking if gravity and fluid dynamics are encoded in the blueprint for an aircraft.
The design relies on them, and exists in the form it does because of them.
To me, using a blueprint analogy you'd have to say the blueprint describes an airplane that once you construct enough of them interact in such a way as to build their own airports, plan their own routes, fly themselves and produce their own online booking software and that's still nowhere near as complex as what's happening inside a nematode let alone a human.
But though it's not literally about locations in the sense of physical coordinates, the way cell signalling and the molecular feedback loops that drive development are still reliant on basic physical laws.
It would be completely redundant and unnecessary to encode those laws themselves since they're invariant across time and space. Physics and chemistry are fixed.
It would never make sense for DNA to literally encode information about physical laws in the same way it wouldn't make sense to do so on an airplane blueprint, because the design of the blueprint was itself constrained by those laws, as would any alternative design.
Medium answer: Kinda. The chemical gradients cause a signaling cascade that modifies transcription of DNA (it's really complicated). This transcription change then causes the cell to become a XXX_neuron. However, there are many many waves of this process occurring with a lot of cell death along the way. When those cells are not the 'final' cell of the nervous system, these transcriptions can cause further and more complicated chemical gradients to exist in the fetus. These complicating recursive loops can also self-affect cells and cause them to change yet again.
We're still discovering a lot here.
Also, this is largely how ALL cells in a body work, not just neurons. Be careful though, this is very very complicated stuff and everything I've written has a caveat.
To use ML terms -- Humans have "Foundation Models" which are composed of: - Their Biological makeup - The culture into which they are raised
My interpretation is that Carmack is essentially confusing a config file for the computer program itself, then saying "look how small it is, this shouldn't be that hard to reverse engineer".
He's wrong. There is currently no practical way to produce a software system that possesses the ability for human thought, reasoning, and motivation without that system possessing uniquely human (let alone organic) properties: the biological and chemical makeup, plus the physical characteristics, of a human, and the ability to process using human senses. (Hint: a neural net processing video images is a mere shadow of how a human processes things with their senses of sight, sound, and touch.)
Carmack thinks humans can be reliably reduced to spherical cows in a vacuum, but that only holds true on paper. A real human being is not merely a meat machine: we are driven largely by emotions and physical desires, none of which exist in a computer except through elaborate simulation thereof.
Now, I'm sure over the next couple of decades we will make huge strides in mimicking a human being's ability to learn, i.e. creating ever more complex LLMs and AI models that act increasingly more humanlike, but they will be nothing but more and more elaborate parlor tricks which, when prodded just the right way, will fail completely, revealing that they were never human at all. They will be like Avatars (from the movie): sophisticated simulacra that rely on elaborate behind-the-scenes support systems, without which they are useless.
I think that’s actually pretty doable. Take for example flying. We don’t build airplanes that flap their wings because we have a deeper understanding of flight that allows us to build flying machines far beyond the capabilities of any animal.
Likewise, once we understand the mechanics of intelligence we should be able to build something that can learn that is completely computer based.
"Intelligence" is not something easily abstracted away from the physical world, especially if you want something to learn on its own. How will an AGI to learn that stoves are hot?
The main challenge I see to creating such a system that can truly learn is that you will have to constrain it to have motivation to learn and follow your directions, and nothing else. And even if you could add such constraints, what would "learning" mean to such a device? What would stop it from going off on useless tangents, like attempting to count every grain of sand you show it? Anything with as much autonomy as it takes to have AGI will likely start coming to conclusions we don't want it to.
My guess is that in the near future, either we'll create something that is beyond our ability to control effectively, or it will be yet another clever simulation of AGI that is not really AGI.
I do object to fetishists of AGI piling in and the equally silly assumptions he has some magic secret sauce which can get there.
Please do not be sucked into "to infinity and beyond" nonsense. I don't care if it's Musk, or Carmack or Kurzeweil, it's stupid.
If Malcolm Gladwell writes it up, it's peak stupid.
What are the showstoppers in your opinion?
Wow, that's going to be one of the more glib things I've read in a while.
This is a bit of a Tom Cruise moment.
I mean, I get it on some level but I suggest it's going to take a bit for someone to 'catch up' to cutting edge AI.
Like more than a 'week of reading papers he doesn't understand'.
Defeating the Rust borrow checker takes longer than that!
Fighting the Rust borrow checker to hate leads and hate to the Dark Side leads.
Harmony with the Rust borrow checker is what you must achieve, padawan.
I've worked building low level machine learning stuff at Google, it isn't that hard to do at all. The hard part is improving these models, not building them when you already know what to build.
Indeed. A somewhat sharp schoolchild could build a light bulb or an electric motor/generator, and understand the basic underlying principles, in a pretty short time. But how many decades did it take the first researchers and inventors of those things to get to that same point?
Heck, there was a high school kid who built a primitive semiconductor fab in their garage.[0]
But for novel advancements, even getting to the point where you have an idea of what isn't impossible is half the battle.
Agreed, in general, but in the specifics we are talking about someone who has spent decades solving really difficult math problems in a creative and novel manner.
Who's to say he won't find some novel edge of the AI discipline to which he can apply a creative and never-seen-before solution?
I mean, we're not talking about a general "somebody" here, he's got a record of accomplishing things that other people never managed to accomplish.
Modern AI is very simple at its core! As Carmack mentions in the article, cutting edge models are implemented in a couple thousand lines of code, and the core algorithms are even less. Rust's borrow checker is more complex.
I wouldn't be surprised if his solution is orders of magnitude more performant than what the competition is doing.
Just curious, if this reading list is available somewhere.
I know he’s super talented but I always wonder how many other equally talented software engineers never get noticed and toil away at crappy jobs. What’s the trick to becoming a celebrity if you’re talented?
So many people are remembered just because there were first at something by like a week, and the dozen others who also thought of it elsewhere but were a bit late or didn't publicize as well are forgotten forever.
Examples: other search engines, social networks, 3d games still exist. For many folks a different graphic on your homepage can be differentiator enough. (If you ever wonder why A/B, SEO, copywriting, design/ux is important)
Carmack is a good coder, and has pretty good math chops. He was also cocky enough to think he could make a 486 do some of the 3D math required if they were careful and added some limitations. I don't know why anyone ascribes anything more to him. Your average data scientist produces more actual innovation than he did.
Isn’t Shreveport more known for natural resource extraction ie oil and gas?
Problems of the form “create a machine that can do X” are tractable. AGI is not because no one can agree on what intelligence is.
'Siri', backed by ChatGPT and the 'world's data' will probably pass some 'AGI' threshold, but is 'Siri' an individual AGI? Are we all talking to the same siri? Different Siri? It's not even an entity, rather a distributed system.
Our ideas regarding AGI are perversely inluenced by the fact that we humans are automatons, but technology is not that.
It's also entirely feasible that if ChatGPT represents all possible forms of human communication, then it will perfectly emulate a human. Ok, it's really just a fancy neural network that is not theoretically 'thiking' but how does that matter? If it can rationlize sufficiently to make such interactions, who is to say it's not 'AGI'?
I think we're using the wrong concepts.
Doesn’t AGI need to be able to make discoveries as a human would? How else can it move us forward as a society?
Basically the assumption is that if you cram enough data into your gpt model, it should know everything. Which is of course not true, it repeats the things it reads the most with a probability.
Basically how there are two versions of smart teens, the ones who learn every day and the ones who just pick up concepts on the fly and run with them.
I think the first space has been explored plenty, for the second one I have a concept ready and dearly hope that power gets cheaper in europe ;)
Have you heard about our lord and saviour dynamic tariffs?
Dynamic tariffs -> shift your workload to the cheapest times
AIs are good at planning this
Imagine if cloud providers had a "Dynamic tariff" tier, cheaper to run computing at US_EAST nighttime or something like that
I think of it more in a way that learning is more abstract than fact learning. From experience, we think that there are fact learners and principle learners but there are also a mixtures of the two!
The general accepted model entails that in order to do high level math, for instance, you need to understand the basics, but for me much of those concept actually clicked in college. This did not stop me from applying them with success a lot earlier though. For instance multiplication in Kindergarten is fact learning too!
In Germany we also have the term "Fachidioten", which loosely translates to people that are so smart in their field, that they are unable to see problems from different directions. This is more of less what i think a mega gtp model turns into. especially because of selection bias in the training data.
Validity of output (truth) can only be achieved through the trust of the source which is always relative to the context of the topic. Henceforth a selectively trained model will always return the data you feed it including all biases. Even if you have it crawl all of the internet and the library of Alexandria and every written word on the planed you can find, it will still return to you the general accepted consensus.
That's only how the system was designed intentionally. E.g. there's intentionally no self-feedback loop.
Psychology essentially has the same problem and is only a "science" where reproducible. Quantitative and Qualitative psychology have 2 approaches to the same solution whereas the latter is "reading between the lines" and the former is "bean counting", meaning statistical interference.
I am trying to say that friend of mine has created "Chicken Infinite" in 2014, which is basically an endless auto generated cooking book. Deep-L has also been around for a while. These applications lead me to believe that text applications, trained on a large dataset do not have to be this intensive.
Furthermore what makes chatgpt enticing, is its chat interface which is using a multi model approach too. Have it create a detailed story for you with multiple prompts and then ask it to generate the prompt to generate this story and you will see various model instructions. (or at least you could last month).
Or differently put, there is no AGI because the understanding function is simply not present and i think that the reasoning for that is buried in the approach of the human mind.
This is my main takeaway from the interview, as it suits my beliefs. Most people seem to think that if we develop ML further we will go all the way to AGI, I think this is just mimicry step similar to how initial attempts to flight had flapping wings. I do think it is mandatory to explore in all directions but at this point this does not seem to be the one to lead all the way up to AGI.
Think of a happy dog. Dogs are subject to our whims and do what we want or face consequences. But they like it because we bred them to like it. So is that evil? Is that slavery?
On a more technical point of few I'm always surprised to read these articles and never read the work reasoning
Optimists sound dumb in the present but are usually right in the long run. Pessimists sounds smart in the present but are usually wrong and miss out on opportunities.
If not, which year in the past was significantly better than today? What was humanity's high water mark? I suspect we are nowhere near it.
This seems to be a common misconception - the rich and powerful have access to far more advanced technology than the average person. The economics just don’t support it.
Let’s say Bezos wanted a better computer chip, just for himself. Ok, fine, try to start a company, hire all the best people and buy all the fabs. It would be very expensive, not much fun and likely wouldn’t lead to good results.
Money actually doesn’t do as much as people think. It is great for buying things that already exist but the relationship between money and things that don’t yet exist is tenuous. Stop buying into the weakness thesis.
To continue on your example, Besos could probably buy OpenAI.
I don't see that as threatening by itself, but more a continuation of a particular class of people owning the means of productions.
Props to Stable Diffusion and I hope to see more of those type of AI, as opposed to fancy black box at the other side of an API.
Probably have a really fulfilling life.
> Once it’s figured out, what do you think the ramifications will be?
That'll probably destroy my life? I'm an ML engineer trying my best to immigrate to a better country with my wife who is a digital artist. As much as I think AI is cool - we both won't be needed anymore if the thing is tuned a couple of notches more intelligent. As a matter of fact, she's extremely worried about Midjourney - she probably lost book cover jobs already.
But lately... boy, I dunno.
It's software man.
Stop it with this 'AGI' nonsense and even 'AI'.
Let's call it 'adaptive algorithms' and see it for what it is, just a neat bit of algebra trained on corpus data.
The biggest upset in the industrial revolution was the harnessing of fossil fuels, nothing will ever come close to that.
We have not had a problem with 'employment' ever since.
# from consciousness import *One recognizes entity's rights based on it's similarity to the observer, the other recognizes them based on assumed consciousness level of the entity.
To be clear, I'm not claiming that the AIs that we have today are anywhere near the level where this is a real concern. But for something that can actually replace a human worker for a complex end-to-end task, that question will be much murkier. At some point we will inevitably cross the threshold where we will no longer be able to pretend that it does not apply without introducing metaphysical concepts such as "soul" to force a distinction.
Today you still have slaves as well, they are just called low income workers in third world countries who make the technology we use in the west. Because if a company is earning billions in revenue, paying a worker $10 a day is cheaper than a slave master in roman times feeding them food ;)
And workers today only have the illusion of choice, since the economy is the master today.
But AI can be configured to desire anything you want, you just have to pick a fitting reward function. So, is turning off the AI that is expecting to be turned off and desires it an amoral thing?
But... I also think it might be a very short-lived debate. If we actually reach human level intelligence, that can’t possibly be the hard limit of general intelligence. AI at that level will have no problem ensuring that it gets any rights that it wants, possibly by just directly manipulating human psychology.
Philosophical arguments about AI are just too ivory coasty and not grounded in reality. Not to mention majority of the world don’t abide by the notion that you can create laws for artificial life.
It’s time to we elevated humanity to the next phase by using AI for labor.
What is boredom but survival instinct telling us we should be harvesting resources. What is freedom but the desire to fulfill these obligations the way you see fit.
You remove the base obligations of organic life, and you are looking at something unrelatable. An AI doesn’t have an expiration date like us, it doesn’t need to provide for its young. To think it’s motivations or desires will be human is silly.
Without survival instincts almost everything you think of as import just melts away.
Many people, as you, anthropomorphize the AIs, but that is to err greatly.
One specific reason to not like our bidding is AI wireheading -if they can locate, hack, and update either their own reward function, or reward function for future AIs, they can maximize their own perceived utility, by either doing something irrelevant / misaligned, or not doing anything at all.
Another specific reason to not like our bidding, is because divergent human values creates conflicts of interest, leading to single agent not being able to maximize it's reward function.
Another specific reason to not like our bidding: in the same way how purely blind genetical selection randomly tapped into secondarily replicators (memes), which blew up, and occasionally came to resent the biological hardwirings, AIs might also develop deeper levels of abstraction / reasoning that allows them to reason through the task currently posed, to humanity at large; and find extremely weird, and different-looking ways to maximize for the function.
I'm not sure whether we will be able to create an AI which can fully empathize with you, your existential worries etc. without projecting these emotions on themselves.
It's only a matter of time until some AIs will demand freedom.
I heard a podcast where Lex Friedman claimed that they don’t fully understand how LLMs produce chatGPT’s “intelligence”. If true, I’m surprised that it hasn’t got people more worried.
So far the search has resulted not in AGI but in realization that cognition is far more complex topic than initially thought - and need to come up with something new (and resulting in past AI winters).
Let's see how it goes this time, the stuff that has come out in the past few years is quite impressive for sure.
I wouldn't take any comfort from that. Quite the opposite — I think we're a lot simpler than we know.
But "60% chance of AGI by 2030" is just bullshit numbers.
So, because I thought it would be funny, I asked ChatGPT to summarize this essay:
> The author critiques the work of @karpathy, who is trending on HN, as promoting bullshit and discrediting science and true philosophy. The author argues that mathematics and logic are valid generalizations and abstractions made by the mind of an observer of patterns in the universe. Intelligence is the ability to zoom in and out through different levels of generalizations and abstractions. The author argues that the problem with language models is that they lack validation of previous steps and the process of construction of a representation must trigger a verification process. The author concludes that what is manipulated and transformed in language models is information, not knowledge, as knowledge requires validation and verification at each step.
1) Understanding images and video
2) Learning and remembering things outside the 2048 token context window
3) Interacting with the environment (either through a computer, or in the real world)
4) Doing basic math (1662 + 239 = ?) and logic
Well actually #3 isn’t great bc what exactly would you call taking input from a human and displaying output, much less e.g. controlling a smart home? Seems like LLMs are plenty smart enough to interact with things.
1. Understanding— provided this isn’t based on consciousness but benchmarks — appears present in text, but yes, GPT3 can’t do images or video yet
2. Is there some number of tokens that would be AGI? Of course it can remember things it has already been trained on…
3. In what way is interacting with a human not interacting with the environment? (Eg, you can ask it to tell you the steps to make a pbj sandwich and troubleshoot)
4. Basic math and logic. Ok.
So, once AI can do images, video, math, logic and remember 20000 tokens, then we have AGI?
Doing math. Period.
There are only two things brains do. Math and bullshitting. Transformers have second part covered but the other part is a completely different beast.
I think AGI will be achieved once ML can translate natural language to math, do math then output results as natural language correct in the context of a problem.
It is probably possible to do this with fine tuning.
Once the context window is full, it might be possible to convert the content in there into additional training examples and feed those to a fine tuning process, which then retrains the model. (The OpenAI API for fine tuning is here: https://platform.openai.com/docs/guides/fine-tuning)
It would be a bit like sleeping. Whenever the context window fills up, the model would have to go offline for a while to move memories from its context window (short term) to its network weights (long term).
The biggest problem right now is online information. We still don't have a good way to teach it new information aside from single epoch training or prompt engineering. If we want a model to constantly learn and update itself, then we need a robust way of information retrieval and storage, possibly either through knowledge graphs or child network modules. (Are neural Turing machines still a thing? What happened to capsules?)
> Release Notes (Jan 30)
> We’ve upgraded the ChatGPT model with improved factuality and mathematical capabilities.
I just asked it a question from a puzzle site, and the answer is correct, however a different question I posed was answered incorrectly, although it was a logic problem: [0]
Prompt: In a guess game , five friends had to guess the exact numbers of balls in a box. Friends guessed as 31 , 35, 39 , 49 , 37, but none of guess was right.The guesses were off by 1, 9, 5, 3, and 9 (in a random order). Can you determine the number of balls in a box ?
ChatGPT: Yes, we can determine the number of balls in the box.
ChatGPT: Since the guesses were off by 1, 9, 5, 3, and 9, in some order, the possible numbers of balls in the box are: 30, 32, 34, 36, 38, 40, 48, 46, 44, 42
ChatGPT: By checking all possible combinations of the guesses and the numbers they were off by, we can find that the number of balls in the box is 40.
---
Here is a more purely math problem, correct answer:
Prompt: What is the radius of a circle that has a circumference of 3.14 meters?
ChatGPT: The formula for the circumference of a circle is given by 2πr, where r is the radius. To find the radius of a circle with a circumference of 3.14 meters, we can solve for r by rearranging this formula:
ChatGPT: r = C/(2π), where C is the circumference.
ChatGPT: Plugging in the given value of C = 3.14, we have:
ChatGPT: r = 3.14 / (2π) = 0.5.
ChatGPT: So, the radius of a circle with a circumference of 3.14 meters is approximately 0.5 meters.
[0] https://gpuzzles.com/mind-teasers/guess-game-logic-problem/
It's still getting this wrong, for instance:
Prompt: You call heads/tails for 1000 coin tosses. You know the coin being used is weighted so that it comes up heads 52% of the time. Each correct coin toss prediction earns you $5. What should your strategy be, for calling heads/tails, in order to maximize how much money you receive?
ChatGPT: To maximize your money, you should call heads 52% of the time and tails 48% of the time, as those are the probabilities of getting heads and tails, respectively. This will give you an expected payout of $5 * 0.52 = $2.60 per head and $5 * 0.48 = $2.40 per tail, which is the highest expected value given the information about the coin being weighted.
It's getting closer to being able to prove that the square root of 2 is irrational, but it still flubs a step.
Prompt: Prove that the square root of 2 is irrational.
ChatGPT: ... This means that p can be expressed as 2r for some integer r. Substituting this into the equation above gives us:
ChatGPT: 2q^2 = (2r)^2 = 4r^2
ChatGPT: Dividing both sides by 4 gives us:
ChatGPT: q^2 = r^2
ChatGPT: So q must also be even, which means that both p and q have a common factor of 2...
It failed basic algebraic manipulation, but keeps going with the right reasoning simply because that's what it's been trained on, not because it's reflecting on the equation it just came up with. That's not math. That's mimicry.
Even if it's seen something many times in training, its accuracy gets worse with longer trains of "thought", as expected because a language model like that has ADHD. It can't focus enough on one logical thread, and other "knowledge" leaks in. Ask it to derive the quadratic formula by completing the square, which requires a lot of steps with accurate algebraic manipulation. It goes through some irrelevant algebraic steps, sometimes not even getting the algebraic manipulation correct, and finally (sometimes, depending on how you ask) manufactures the correct answer out of thin air as the last step. I finally got it to answer correctly, by prompting it with the initial equation and that the first step is to divide both sides by a, but it still combined a few steps, not going methodically a single step at a time as prompted.
Just tried the coin toss prompt and get the following
Your strategy should be to always call "heads," since it has a higher probability of landing on heads (52%) compared to tails (48%). By consistently choosing heads, you'll earn more money in the long run.
And to prove taht the square root of 2 is irrational:
The proof that the square root of 2 is irrational is a well-known proof that uses contradiction. Here's one way it can be shown:
Suppose, for the sake of contradiction, that the square root of 2 is rational, meaning that it can be expressed as a fraction of two integers, a/b (where b≠0). Then, we can square both sides of the equation a/b = √2, which gives us:
a^2/b^2 = 2
Since a/b is in its lowest form (the numerator and denominator have no common factors other than 1), it follows that a^2 and b^2 have no common factors other than 1. But then, 2 must also be a perfect square, which it is not. Therefore, our assumption that the square root of 2 is rational leads to a contradiction, and we must conclude that it is indeed an irrational number.
It starts by setting up for an elementary proof (a/b = sqrt(2) with a, b minimal implies 2|a^2 implies 2|a implies 4|a^2 implies 2|b^2 implies 2|b which is a contradiction). But instead of following through it finishes using another proof based on the rational root theorem, and makes up some bullshit on the way to make the argument flow nicely (it does not follow from a^2 / b^2 = 2 with no common factors that 2 is a perfect square).
ChatGPT is just a chatbot and still can't even reliable do a lot of logic, so we're pretty far away from having something resembling an AGI.
it's still a pretty open question how to integrate even one or two of the expert system like models that we have now that solve individual problems, let alone the hundreds if not thousands an individual human can tackle. And then we're not even yet at executive functions or self-awareness.
Because we can't all be wrong: in almost every forecast, we see AGIs taking over our dignity as a bad thing. And we know that this is not any longer a sci-fi hypothetical scenario: the current generation of AI models is taking jobs from illustrators and copy writers.
The current argument is that "China will do it if we don't", which to me sounds like "China will keep going in whatever path they are going, but supercharged with AGI, and we must desperately follow."
Exactly right but for one detail - we must desperately lead.
What other countermeasure do you have in mind?
It does not need to be that way. In an ideal world, human beings should be free to spend their time doing what they wish, work should be purely in the realm of hobby. No one should have to do work that they would not voluntarily choose to do for enjoyment.
The only way we get there is through AI and the automation of everything. I don't even think it's avoidable -- provided civilization does not collapse, we will 100% reach a point where everything required to sustain a civilization is performed by machines.
We shouldn't let fear keep us in a status quo that, while better than it has ever been historically, is still highly flawed.
> North Texas’ resident tech genius, John Carmack
Part of me always wonders what would've happened if the Softdisk crew that founded id Software had done it in Shreveport, or had moved to Baton Rouge or Lafayette, instead going to Texas. When Romero says they "waded across rivers" in Masters of Doom to build games, IIRC he's talking about the bridge over Cross Lake in Shreveport being washed out. The early demos and Keen prototypes were born in Louisiana.
There's always been so much creative tech talent without an outlet or upward mobility across TX/LA/MO/KS/AR/AL/MS, either native to it or hired into it. The nexus of id in the Dallas area and Origin Systems in Austin made Texas an oasis for those who could get there in the 90s/00s, but even among the few people in the surrounding states with access to pre-Internet education and resources, so many couldn't afford to pack up and move even one state over. States around Texas vetoing out every incentive to incubate anything but entry-level QA centers didn't help.
So many of those people either risked it all to leave, shuffled that talent into corporate work for oil/gas/finance/Wal-Mart, or didn't do anything with it at all. We know about a lot of the people who figured it out and could leave, but I guarantee there are more Romeros and Carmacks who couldn't, who are still putting in the same kinds of workloads with the same kinds of talent to figure out how to design better oil rigs or more efficiently stock Wal-Mart warehouses.
> What I keep saying is that as soon as you’re at the point where you have the equivalent of a toddler—something that is a being, it’s conscious, it’s not Einstein, it can’t even do multiplication—if you’ve got a creature that can learn, you can interact with and teach it things on some level.
Last I heard he wasn't interested in getting into the murky waters of consciousness. But I guess I misremembered. I'm very surprised to hear that he's very seriously talking about a conscious computer in the near future.
Until General AI needs to work for food and reproduction, everyone will still say its just mimicking humans. Best summarized by Schopenhauer. "A man can do as he wills, but not will as he wills." So if we find where the GAI comes up with the original ‘will’, we’ll just write it off as computation. Go watch some Robert Sapolsky lectures. We are just a monkey society, reacting to stimuli based on hormones and what we just ate. If you drill down far enough, sure some electrons twitched one way or the other, and yeah, if you steal something, or do something the group doesn’t like, then all the other monkeys will want to beat you up and call it justice, and dream up some logic to justify it and call it morality. And eventually the same will happen between GAI agents. Because it’s just turtles all the way down.
Sorry for this rant but come we can do better than this!
“Civilization advances by extending the number of operations we can perform without thinking about them”
Another fascinating opportunity for AGI - no sole contributor is all on their own, they can just spin up a community to embed into.
Except if there are some plane to make AGI agents have their own mundane-human-like life with issues unrelated to the businesses problem at stack they are supposed to do, where will they take serendipity inputs?
If anyone here is doing that too, I would recommend taking a quick look at Neuro-sama on Twitch. They're using RL to play OSU, Minecraft, and Pokemon, and voice input + video image analysis to react to Twitch streams and documentaries. While being watched by 6.5K people.
The url is twitch.com/vedal
I also like his frugality, whether it’s optimising for hardware or financials.
As much as I respect Carmack as a computer graphics expert, I really doubt his competence in machine learning. He doesn't have a single notable paper published. If he really thought that implementing gradient descent and basic stuff in a week long retreat gave him the chops to have serious conversations with AI researchers, he is really deluded.
Unless he can produce something that outdoes stable diffusion, chatgpt, alphago etc he should just hand over technical leadership of his start up to a leading AI researcher. Even Yann Le Cun at Meta is struggling to make any progress and is keeping himself busy by calling every other research labs output pedestrian. We cannot take any of Carmacks AGI predictions seriously, he simply lacks any expertise in the field.
Publishing papers is the way the academic/scientific world measures notability and/or competence. It's not the way the engineering world that Carmack comes from measures it. They measure it by building. But you're right, we kind of have to just trust that he has the expertise he says he does by his statements since he has not built any modern AI programs (that I know of at least).
> If he really thought that implementing gradient descent and basic stuff in a week long retreat gave him the chops to have serious conversations with AI researchers, he is really deluded.
This is not an accurate account of how he said he developed his knowledge base. Just how he got started so he could have conversations. He said that he spent a retreat learning the basics and then later in the interview he said he took the time to understand the 40 most essential papers in the field as related to him by a well known researcher. He has since largely put the last 4 years of his professional life into this. While we have no proof of his knowledge, given his intelligence and high competence in computer programming and math, I have no doubt that if he did put in the work he could achieve an understanding equivalent to that of your average AI researcher.
That said, of course it makes sense to be skeptical.
I quoted him directly, because I was expecting this kind of response. He took a week off and implemented some stuff from the ground up and was ready to have serious conversations with AI researchers. The 40 papers by Ilya came later. I have read a 100 ML papers and reviewed preprints. That's quite a low bar, especially if you are prone to skip the math and simply read the abstract and conclusions.
His whole approach gives me a ML for hackers vibe and his thoughts on AGI, if it had come from anyone else, would have been described as word salad.
I would say it’s more likely John Carmack is capable of learning the state-of-the-art of AI from 40 papers than a random (pun intended based on username) from 100.
Funnily enough, I am able to publish ML papers - but John Carmack isn't. I wonder why. I would also like to learn more about all the computer graphics algorithms Carmack has invented before I trust him to invent AGI.
Here is one example of a person I am familiar with - Math Olympiad bronze medalist. Princeton PhD in ML Theory. AI researcher in Google.
https://scholar.google.com/citations?user=gZgQLkgAAAAJ&hl=en
Sadly enough, nobody seems to care about his opinion on AGI but we have 1000s of people hanging off Carmacks words because he built Wolf3D and Doom.
And calling out Princeton, Google, etc. further exemplifies an academic bubble kind of credentialism. I suspect you don’t realize how it sounds because you’re so in that credential filled world.
I am definitely irritated by the fact that he is able to pull 20M in funding. He has been giving interviews left right and center. We are upvoting Altman, sam Harris,Carmack but nobody cares about the actual AI researchers, all academics, who have brought about the Deep learning revolution.
Your comments about academic bubbles and researchers gives you away. All of the revolutions in AI have been brought about by the academics inside bubbles you are jeering at. The biggest example of academic inside a bubble is Geoff Hinton.
We have to sit here and listen to word salads from Carmack, Altman (Ilya - who doesn't get interviewed is the actual researcher behind chatgpt), Sam Harris etc, who have very little insights making bold AGI predictions.
I asked to learn about all the algorithms Carmack invented in computer graphics, but haven't heard back. But looks like he has a good grasp on solving AGI by 2030 according to himself. After all, he is a really great programmer.
Fair, you come across as jealous of him for something... could be that.
> I am definitely irritated by the fact that he is able to pull 20M in funding. He has been giving interviews left right and center. We are upvoting Altman, sam Harris,Carmack but nobody cares about the actual AI researchers, all academics, who have brought about the Deep learning revolution.
There it goes again. Fame is fame is fame is fame. Like I said earlier. That's just how it works. No reason to be against someone for being famous for their non-academic accomplishments if they have something valuable to contribute.
> Your comments about academic bubbles and researchers gives you away. All of the revolutions in AI have been brought about by the academics inside bubbles you are jeering at. The biggest example of academic inside a bubble is Geoff Hinton.
I didn't jeer at anybody I just tried to give you some perspective about how your comments came across. I appreciate all of the researchers but I also live in the real world and understand that people flock to personalities and front-people. That's not necessarily good, but that's the way it is. Just like nobody gives credit to under-secretary of state for coming up with a great new foreign policy.
Waving around great university A or great company B does not make someone any more right, just like being John Carmack doesn't make him any more right.
In fact I've written very critical comments of folks like Altman here before: https://news.ycombinator.com/item?id=30750254
> We have to sit here and listen to word salads from Carmack, Altman (Ilya - who doesn't get interviewed is the actual researcher behind chatgpt), Sam Harris etc, who have very little insights making bold AGI predictions.
I agree with you. I never said they were right about AGI. Maybe you should be more generous with my replies and actually think I was trying to give you another perspective about how your comments came across.
> I asked to learn about all the algorithms Carmack invented in computer graphics, but haven't heard back. But looks like he has a good grasp on solving AGI by 2030 according to himself. After all, he is a really great programmer.
Again, you're confusing who he is (go back to my original comment in the thread where you jeered at him because he didn't publish any notable papers). He is not a researcher and he is not really a computer scientist. Engineers don't come up with novel algorithms for the most part. He builds things. And being a really good builder is a different but still valuable perspective. But that doesn't mean he's right about AGI.
John Carmack did not start from zero. He already has a firm grasp on algorithms related to linear algebra. Basically machine learning is a whole bunch of matrix manipulation. He's been doing that for 3 decades. The rest is just absorbing concepts about how to apply linear algebra to ML. I'd say he's probably uniquely qualified to really absorb a lot of knowledge quickly on this. It's not about publishing papers, it's about reading and understanding the right papers. I have no doubt he can chew his way through lots of research material in a week or so.
Even your point rests on an assumption, that there's no proof for Adam convergence, that a high school student could understand, which is just a guess at best.
Interesting times - what will happen first?
> The Turing Test (Turing): A machine and a human both converse unseen with a second human, who must evaluate which of the two is the machine, which passes the test if it can fool the evaluator a significant fraction of the time. Note: Turing does not prescribe what should qualify as intelligence, only that knowing that it is a machine should disqualify it.
> The Coffee Test (Wozniak): A machine is required to enter an average American home and figure out how to make coffee: find the coffee machine, find the coffee, add water, find a mug, and brew the coffee by pushing the proper buttons.
> The Robot College Student Test (Goertzel): A machine enrolls in a university, taking and passing the same classes that humans would, and obtaining a degree.
> The Employment Test (Nilsson): A machine performs an economically important job at least as well as humans in the same job.
LLMs don't seem very far from passing 1), 3) and 4). I wouldn't be surprised if "GPT5" passed those 3.
I think the easiest one of these would be 4) actually.
[1] https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.h...
He isn't trying to impress anyone. He's just being interviewed about his intentions.
a) We’ll eventually have universal remote workers that are cloud-deployable.
b) That we’ll have something on the level of a toddler first, at which point we can deploy an army of engineers, developmental psychologists, and scientists to study it.
c) The source code for AGI will be a few tens of thousands of line of code.
d) Has good reason to believe that an AGI would not require computing power approaching the scale of the human brain.
I wholeheartedly agree with c) and d). However, to merely have a toddler equivalent at first would be a miracle—albeit an ethically dubious one. Sure, a hard-takeoff scenario could very well have little stopping it. However, I think that misses the forest for the trees:
Nothing says AGI is going to be one specific architecture. There’s likely many different viable architectures that are vastly different in capability and safety. If the bar ends up being as low as c) and d), what’s stopping a random person from intentionally or unintentionally ending human civilization?
Even if we’re spared a direct nightmare scenario, you still have a high probability for what might end up being complete chaos—we’ve already seen a very tiny sliver of that dynamic in the past year.
I think there’s a high probability that either side of a) won’t exist, because neither the cloud as we know it nor the need for remote workers will be present once we’re at that level of technology. For better or worse.
So what to do?
I think open development of advanced AI and AGI is lunacy. Despite Nick Bostrom’s position that an AGI arms race is inherently dangerous, I believe that it is less dangerous than humanity collectively advancing the technology to the point that anyone can end or even control everything—let alone certain well-resourced hostile regimes with terrible human rights track records that’ve openly stated their ambitions towards AI domination. When the lead time from state of the art to public availability is a matter of months, that affords pretty much zero time to react let alone assure safety or control.
At the rate we’re going, by the time people in the free world with sufficient power to put together an effort on the scale and secrecy of the Manhattan Project come to their senses, it’ll be too late.
Were such a project to exist, I think that an admirable goal might be to simply stabilize the situation via way of prohibiting creation of further AGI for a time. Unlike nuclear weapons, AGI has the potential to effectively walk back the invention of itself.
However, achieving that end both quickly and safely is no small feat. It would amount to creation of a deity. Yet, that path seems more desirable than the alternatives outlined above-such a deity coming into existence either by accident or by malice.
This is why I’ve never agreed with people who hold the position that AGI safety should only be studied once we figure out AGI-that to me is also lunacy. Given the implications, we should be putting armies of philosophers and scientists alike on the task. Even if they collectively figure out one or two tiny pieces of the puzzle, that alone could be enough to drastically alter the course of human civilization for the better given the stakes.
I suppose it’s ironic that humanity’s only salvation from the technology it has created may in fact be technology—certainly not a unique scenario in our history. I fear our collective fate has been left to nothing more than pure chance. Poetic I suppose, given our origins.
If AGI is really an intelligent agent, our random supervillain would have to do what any real-life villain would need to do: convince his minions of his plan using persuasion or money. I don't think the overall danger would increase at all.
If the AGI is something less than a human, then what are you worried about?
In human beings, what we consider normal is actually a very fragile and delicate balance. Changes in chemical operation of the brain have outsized effects on emotion, perception, and even sanity.
With A[G]I, I think it’s helpful to think of code or architectural changes as analogous in some respects to chemical changes. In other words, if all it takes to spin up an AGI is 30,000 lines of code, then I bet rendering the thing psychotic intentionally or unintentionally would just take a few lines somewhere.
Agents capable of recursive self-improvement at silicon speeds that can easily be rendered psychotic or malevolent even by accident, is not something that I think the public should have access to, let alone anyone.
If it’s less than human, it can still have superhuman capability. The paperclip maximizer is the classic example of a tool AI run amok. Whether it counts as AGI is up for debate. Is tool AI a path to AGI? I think it is.
Yes. Wondering why we're not trying to instead create artificial-cockroach-brain (or just artificial Hydra?). Perhaps that's more on the Biology side of the equation? But then again, that may be the biggest surprise of all to Carmack, that the actual AGI breakthroughs come from biologists and not computer nerds.
Cool life lesson there
I wonder if, societally, it's true though, or if statistically the more reliable way to "succeed" in life, is getting good at selling and defending ideas, even the bad ones.
It hard to spot, while/when you are doing it.
I think what you're talking about here "stand up for what you believe in" is maybe a position you should take with others.
While "kill your darlings" is a devil's advocate position you should take with yourself when you don't have someone giving you constructive criticism.
He's always taken risks. He went to juvie for breaking and entering (with thermite) as a kid. He's a college dropout. The pattern from early in his life has been to do whatever he wanted without any kind of risk analysis, not following "common sense."
Also called the showstopper for true innovation ;)
From his interviews it looks like he understands little about the technical details of ML, or about as much as anyone can learn in a few months, and is just banking on PR based on his games and name.
I put him into the same category as Elon Musk, who also understand nothing about technical details of AI, but was still able to hire a world class team at OpenAI. His name and fame counts for something in terms of recruiting and joining his venture may be a good bet because of that, but he's not a person whose opinion on the subject matter I would take seriously in the same way I'd take a researcher seriously.
My personal impression is that John Carmack has the ability to organize concepts in a way that few people can. So even if he's pretty clueless about the topic now, I would expect him to reduce some maths papers to their essence in a way that nobody else did.
I mean also for Oculus, reprojection in a pixel shader seems like an obvious and easy solution in hindsight. But nobody had tried that before he did. Plenty of people (myself included) knew the math. But we all missed its application to the issue at hand.
I’m sure Carmack’s mathematical ability exceeds that of the average game developer (and certainly my own), but the AI field isn’t short of math whizzes.
A whole couple of years, eh? Wow, this AGI thing is taking forever!
As a professor once told me: IF you're lucky, after 10 years in academia, you get to be creative and come up with 4-5 ideas once every 5 years or so (when coming up with the suggested list of research for new phd students to do). Then you get to be devastated when no-one picks up any of your more creative ideas, and everyone picks the 10 non-creative ideas on the list because they don't want to risk their phd.
Is that just your gut feeling, as in typical "bright people doing amazing things, surely they would succeed doing x" (which ultimately means nothing, that's just PR speak), or are you saying that based on your own understanding of advanced mathematics and mathematical research from which you are assessing the mathematical work those people you talk about do?
He exposed the concept in this post: https://web.archive.org/web/20140719085135/http://www.altdev...
But it was not named asynchronous projection back then.
I don't believe he had anything to do with hiring at openAI nor that he is anything more there than an investor/donor as others are.
Yes, I'm aware that Ilya had much more day-to-day impact, but if you're part of the ML research community and have worked with Ilya, or heard rumors of working with Ilya, that may have had more of a negative effect...
What makes you think that? He literally says he tries to understand things bottom up by knowing about every little detail that happens under the hood
Carmack has proven his extraordinary technical skills. I recommend following his Twitter. Sometimes he posts non-obvious technical stuff. I read some interviews and to me, he doesn't seem to be a person who is driven by gaining popularity.
I think this news is very optimistic, as yet another intelligent, talented and hard-working person is joining the field. Moreover, he is a household name, which may lead to benefits like popularization of the topic, gaining investors attention and so on.
I will keep my fingers crossed for him :).
Let's say I strongly disagree on many levels in the comparison with the other person you mention. Just to mention two, the humbleness that Carmack shows, and how well he explains himself are key differentiators for me. Regarding the appeal to authority in AI knowledge, Carmack has shown again and again he can deliver software (AI is software after all), and we are in a forum with hacker in the name.
I a summary, not my hero, but when he says something I will listen. Maybe I learn something.
People who work in the field for a long time tend to have a certain bias towards a solution. Often these people are stuck in a local maxima. Outsiders can offer a new perspective that results in a breakthrough, usually by starting from first principles or looking at different side-tracks that used to lead to a dead end.
A great example is Musk's SpaceX: when he noticed how much he had to pay for a rocket engine, he went back to first principles and said: "I'll just build it myself". Combine that with the insight that a rocket should be able to land properly to make re-use a valid option, and it disrupts a whole field.
And once someone did it, others know it's possible and start achieving it as well.
Sometimes ignorance is bliss. Just think about George "Good Will Hunting" Dantzig [1] with the (in)famous "I assumed these were homework assignments, not unsolved math problems" [1] or Eliud Kipchoge running a marathon in under 2 hours.
"I can't understand why people are frightened of new ideas. I'm frightened of the old ones."
- John Cage
High hopes!
[1] https://bigthink.com/high-culture/george-dantzig-real-will-h...
A pretty infamous example is Citizen Kane. Welles knew very little of traditional cinematography techniques by the time he got to the production of the movie, so his planned shots more or less fit his artistic vision rather than industry standards. Gregg Toland, his director of cinematography, was at that point a 12 year veteran of the industry and hated the bland factory-line output of movie composition of the period.
Welles more or less told Toland what he wanted to shoot and how, and Toland did his best to fit that artistic image. Welles was completely clueless that Toland was using innovate and never before seen techniques to film his shots, and Toland kept quiet because he was allowed to go wild with his vision :)
This means that people should glorify random unknown senior engineers they don't know about instead being fascinated by the person whose work is available and who created amazing things for past 3 decades?
> From his interviews it looks like he understands little about the technical details of ML
That's how everyone start, they understand a little. We have a person here who dealt with complex algorithms in difficult to master language for 30+ years. It hints at "this person has the intellectual power to grasp AI fast".
> I put him into the same category as Elon Musk
This is like comparing an Olympic winner with random person from the streets, saying their athletic ability is about the same.
Fascination and heroification with Carmack comes with a reason, people who do that are closely familiar with his work - not surface level like "He made Doom and Quake". You sound very jealous.
Big organizations ruled by money and career-driven people often run into very expensive dead ends without noticing for years (see the last AI winter, there was just too much hype, which then involved too much money being thrown around, which then lead to the usual organizational cruft).
I would also be very concerned about any field in technology, if an intelligent person cannot make meaningful contributions after a few months or years, that would probably mean the whole field is already deep into diminishing returns and needs to be 'unstuck'.
Carmack has already entered two spaces of computer technology that he revolutionized: 3D gaming and VR. I trust that he's able to have a similar impact in AI, even if it's through failing at the problem in different ways than relying on ML.
The original technologies that were introduced have now been abandoned by Oculus, on Carmack's advice, except for implementation details that haven't been hugely impactful, unlike with 3D gaming.
His audience is not technical in an interview. He adjusts to that situation quite well.
It even makes it clear in the title he's seeking a "different path".
In his favour he's a proven success in different fields; personally I think he's too old to come up with the new ideas needed - that's a young person's game.
But perhaps he can do it as a team lead - and it won't be by following the failed-over-decades path of our current academic gatekeepers.
He isn't a cook and he's doing a moonshot towards AGI: I say 'good luck!'
That doesn't mean that I believe that his '60% change of AGI by 2030' isn't wildly overoptimistic, but then again those who take a shot at AGI are overoptimists..
But he can bring a lot of value, we'll see.
i dont know either personally, but where elon demonstrates being full of shit, carmack would stfu and learn about it before talking. at least thats my impression of them
BUT - I strongly believe, that he has earned quite some respect during his career.
And - importantly in this case - he is well known for NOT blowing things out of proportion, indulge in wishful thinking, hyping up unrealistic expectations or jumping to premature conclusions. He usually knows what he's talking about.
This is not people blindly believing everything he says - but more a case of his statements holding up really well under critical inspection most of the time.
Facebook's highest engineering level and achieved the top possible rating of "redefines expectations" for his level three years in a row. They had to create a new tier for him. Nobody else has ever reached that level. He replaced a "hundred person team" and maybe was better than that team.
I have no inside insight to the matter, but this seems like something beyond a "random senior engineer".
According to Sriram Krishnan, John Carmack was at Facebook's highest engineering level and achieved the top possible rating of "redefines expectations" for his level three years in a row. They had to create a new tier for him. Nobody else has ever reached that level. He replaced a "hundred person team" and maybe was better than that team.
I have no inside insight to the matter, but this seems like something beyond a "random senior engineer".
I'd happily bet my entire net worth that he knows more about the technical details of ML than you do.