It’s a weird time. Elon is similar. The guy says the most ludicrous things but seems to have “delivered” so it’s not easy to discount his insanely stupid comments.
Sama is another one. He is part of some sort of success story, makes a bunch of radical 4d chess type claims, so you can’t discount everything he says, however, even if what he says is true, how do you “build for AGI”? The dude sounds legit bonkers, but again, he has some type of cred because his team delivered a working product. It almost feels like this is a common play adopted by a lot of tech elites. It’s not a coincidence.
“If you find yourself part of some success story, any success, use it to your advantage. Use your ‘cred’ to drive a lot of hype and cash out, using all the cash you accumulate hire hard working people, have them deliver something which in turn keeps you looking ‘credible’, make more outlandish claims, gain more funding, rinse wash repeat”
I feel Jeff Bezos is an example of someone who doesn’t do this. Seems like a practical guy who works really hard and avoids (mostly) making silly claims.
> have them deliver something which in turn keeps you looking ‘credible’
Yeah, this is the key part, because a lot of the specific claims do fall over once interrogated (self-driving when?), but usually he is delivering something.
I find sama, musk, jobs, elizabeth holmes to be in the same category. Only difference is that sama, musk, and jobs were able to not only market and get hype, but also build some 'threshold' of what they claim.
Holmes, while nearly as inspirational in the 'marketing' sham they all do, didn't have enough of a product, so she's in jail.
People like 'ilya' who understand the tech deeply are hidden and lose out to these 'musk' type folks, like sama.
Maybe the types of wrapper startups and apps people are building with GPT4 are obviously going to be replaced by GPT5 in a few months. There will be huge backlash because startups, VCs, and solo devs aren’t thinking big enough.
Could be that he’s dreaming too big about AGI, but I know several people building GPT4 based products that will obviously be replaced by the next small advancement in GPT.
Maybe Sam is just trying to push people to think (and fund) a bit longer term ideas that won’t have so much backlash when they’re replaced by native GPT functionality improvements every few months. This will also create less churn in the app store and for users.
It seems obvious in hindsight that almost all ai startups last year would die in a few short weeks to months because openai simply integrated the functionality into their own product -> Assistants, GPTs, GPT Store, ChatGPT for teams and businesses, retrieval, voice, vision etc. I wager that 2/3 of all these startups thought these were clever products to work on back then. The difference is: Openai just drops new features without previous announcement, exceeds expectations and kills these api wrapper products instantly. Compare with google who announces their new ai features beforehand, fails to deliver on time and when delivered doesn't meet expectations.
Do you see it too though ? It really feels like a social device.
I can’t help fall back to Musk. So much of his success is due to not only his own hard work, but plenty of others, yet he is absolutely wonderful at marketing it like it’s his own and we should hang off everything he says.
Even when he does credit his team, it’s brief and then it’s straight back to the hype show.
Of course I’m not suggesting the dude is a complete fraud, he is part of some brilliant things, but it doesn’t mean he isn’t taking advantage of it.
I mean it is a smart way of operating and it totally works. I just don’t think it’s
honest.
Sam Altman seems to operate by the exact same playbook, I’d say he is a novice player compared to Elon though.
People do not understand the meaning of "Marketing". People believe it is selling what you have(Sales and Advertisement) while it really means understanding Markets: The needs of the people.
If you have good intuition on what the needs of the people are, even on things that do not exist yet, you can design your product or service accordly. You could also choose the right people for the job.
People like Steve Jobs or Elon are great at understanding markets. i.e Steve knew that people would be using their smartphones on they pockets and that having scratch resistant screens was essential while the rest didn't care. I had a PocketPC and TabletPC and Microsoft cared so much and invested billions in things that few people care while being against most user real needs.
They are visionaries that have to imagine a future that does not exist yet. The kind of people that can do that, like Elon or Sam usually can see the future as real as it already exist and can be overoptimistic as for them it is obvious that something is going to happen as for them the future is as real as the present.
Elon saved Tesla from bankruptcy choosing a pathway to mass fabrication of EVs. The original Tesla vision was exclusivity and using Lotus car frames.
What separates Elon for everybody else is that he risks his own money on what he believes in, not someone else's like Sam. He is a risk taker like nobody. This is what I hear from people that know him personally.
Elon and Sam can be as visionary as all hell, they can also be wonderful manipulators too. Both things can be true. Especially if said behaviour gets them more money to take more risks.
In the case of Altman, I feel the way he communicates in this sort of secretive, somewhat threatening way (I’m about to unleash my super secret AGI published by my fake nonprofit org, regulate the completion at once) is sociopathic. It’s quite unique to him. I recall Steve Jobs being visionary, I heard stories of him being a perfectionist, but I don’t recall him using the same sociopathic playbook I described. Think of the contrast in behavior between Altman and say, Carl Sagan, the contrast is so stark to me. Between Elon musk and Eric Weinstein.
Anyway my original point wasn't focused on the people, just that I think they’ve workouted out how to perfectly game our attention economy to their advantage. I don’t hate them for it per se. I just think it’s dishonest and distracting. I'm not even sure they know they're doing it, but they just do it because it works.
I’m getting strong “full self driving by the end of the year (2017)” vibes from this. His business is very hype-based so he’s working the hype machine. AGI is a fantastic promise in this regard because it’s not falsifiable - there is no broadly agreed upon and well understood definition of AGI. Instead the people listening to his promise imagine whatever they want and then get excited. By the time 2025 comes along there will be a new promise. They’ll claim they have AGI but it only handles text, images, and audio, perhaps. Then they will say “and in three years we will double the modalities and have AGI plus” and it will still be broadly meaningless.
Meanwhile I’m a robotics engineer and I need systems that can understand large amounts of data in real time and I doubt their 2025 “AGI” will be able to intelligently process LIDAR data. Such narrow systems are hardly general in my eyes, even if they are very compelling chat bots.
>there is no broadly agreed upon and well understood definition of AGI
OpenAI follow their own definition of AGI, and it's "highly autonomous systems that outperform humans at most economically valuable work". [0] Whether that includes jobs that require robotics is an open question.
Thanks for the reference. I think this only reinforces my point. They’re going to claim something that does economically valuable work on the computer is AGI, but when I hear “most economically valuable work” I think “build buildings, operate factories, grow food”.
We’ve just never seen a universal multimodal learner (all modalities) or a system with its own goals and motivations that can learn on its own without massive datasets (meaning things without datasets are hard for it to learn, so how does it for example learn how to do PCB design or CAD modeling), so they’re talking about such a leap it’s hard to fathom. I mean hell take PCB design for example. That’s actually on the computer but there’s no big dataset in existence that actually explains PCB design in a way a system would understand, so approaches that rely on a dataset for training simply can’t begin to solve that task.
And if you can’t even do that economically valuable task on the computer then I don’t believe you have AGI.
I am quite sure that they’re cooking up some fascinating stuff but I doubt they are going to have a system that can do machine design or PCB design just to name a few important and extremely economically valuable tasks.
Yeah, obviously anything anyone from OpenAI publicly says is marketing, especially considering that the all-loving cosmic utopia in the link I posted clearly is pure marketing BS. They actually added that definition in 2019, and shifted their priorities from their initial goals - no reason not to do this again any other time.
Not just annoying but extremely cringeworthy also, like he's some cute elementary school buddy. These are some ice cold very powerful people, let's always remember that.
Don't create parasocial relationships to them like some desperate pawn.
Assuming he's right (which is a big assumption, but much more interesting than calling BS), society is going to require some massive changes to carry on functioning. Imagine a world where all knowledge work can be replaced with a clever bot that actually gets things right. Entire industries will collapse. Things we value highly will become close to worthless (largely depending on how much OpenAI charges of course). It'll be so weird. The idea of never paying an accountant or a lawyer again is very appealing.
We haven't seen what AGI would look like, yet. If I understood correctly, what we have are statistical language models. I am far from knowledgeable in this domain, so I defer to Chomsky et al. here [1].
Makes me wonder what kind of progress they've had internally at OpenAI to be making such claims.
So far I haven't even seen any particularly useful or compelling definitions of AGI; it's usually handwavy, incredibly vague, or defined in terms of something else which just shifts the problem around.
The general sentiment I've gotten from following a bunch of AI researchers on Twitter is that we're still multiple key insights away from being able to build AGI. So either Sam knows something we don't or he's hoping to we get lucky in the next 2 years.
I hope he's wrong and we actually get AGI by 2024.
Makes me wonder what kind of progress they've had internally at OpenAI to be making such claims
This is by design.
I think this is why he is the CEO, he is a master of deception, half truths, decoys and promises.
You never know what he knows so bad fucking luck, that’s his attitude. Even if he isn’t lying , he is communicating in a non-friendly, secretive and manipulative way which is toxic.
AGI, something better than humans at pretty much ahy tasks would have some radical implications for society. Him throwing this around loosely is irresponsible and shows a lack of respect or empathy on many fronts. If OpenAI are trying to develop it, fine, but just so some maturity as well.
There’s a relatively mainstream definition of AGI: an AGI could learn to accomplish any intellectual task that human beings or animals can perform. Sometimes agency is included in there to facilitate the learning and accomplishment.
I think it’s as good a definition for a major milestone in AI development as any.
The other mainstream definition comes from Peter Norvig , author of the most mainstream textbook on artificial intelligence (or at least the most cited). He defines AGI relative to narrow AI, using the development of general purpose digital computing as an analogy. Before general purpose digital computing, there were a wide range of specialized narrow forms of computing.
Thus, he concludes that AGI is already here (it is similarly general purpose) but in the early days.
I don't buy this argument. Specifically, Modalities. The claim is that "[b]y using modality-specific tokenizers or processing raw data streams, frontier models can, in principle, handle any known sensory or motor modality". Sure, you may get some sort of output but I'm quite confident you won't get anything useful without training the model on that kind of input. We barely started sticking narrow AI together. E.g. ChatGPT (LLM) can invoke Dall-e (text to image). Maybe AGI will turn out to be a bunch of narrow AIs in a trench coat. So far I haven't seen such a system.
Tasks claim is a little stretched, too. For example, simple arithmetic. Does model actually do arithmetic or does it do text generation and the right answer just happens to be the most likely next token? Can we reliably tell one from the other to claim that model actually perform tasks other than just text generation?
I'm familiar with this definition but find it lacking. It has utility as a working definition for gesturing towards a desirable goal. My problem with such a definition is that it doesn't clarify anything, it basically shifts the problem over to one where human or animal intelligence is treated as a magical black box. Sorry if this comes off as a bit pedantic, I understand that research into new and unknown fields can take a while before landing on more accurate and unambiguous definitions.
I think of it a little differently. Human or animal intelligence here is a somewhat useful reference point.
As an analogy, think of the first steam engines. They were replacing horses so why not measure their primary function in units of horse power? We had a somewhat rigorous physical framework for work and power but it probably wasn't well understood by the target audience to be useful.
While we don't seem to have a scientifically rigorous definition of intelligence we still can have a somewhat useful discussions about utility in reference to another familiar instance of intelligence.
I didn't define AGI because I'm also confused about the term. As the term is currently used, it seems to directionally gesture vaguely towards something that's mostly based on gut-based sensing. Maybe it's one of those "when you see it you'll know" kind of things. Or maybe we'll have to experiment with building it first before we can narrow down the exact boundaries of how it's defined.
If Sam genuinely believed that AGI (assumption: capable of achieving some sort of singularity, or at least being capable enough to replace x% of human knowledge workers) would materialize in the near future, why would he then invest any time or money into these YC companies, given that most of them, if not all, would be made completely obsolete by such a technology?
Occam's razor would imply that he is simply encouraging future YC batches to build products on top of OAI products, with "AGI" being some sort of multi modal GPT.
Although the tweet is not clear, people who were at that presentation seem to confirm that's exactly what he said. There are a few of those sparse in this thread.
Tech leaders are so bad at making predictions. Some recents examples: self driving cars should have happened around 2021, not there yet. Elon's claimed to launch Tesla robot taxi around 22/23, and we don't see anything like that.
That's said, I do believe that GTP-5 will be out in a year, but re:AGI is a such a bold prediction.
I was just driving this afternoon and I saw a "trucks and buses must keep right" sign on a road where the right lane was blocked due to construction work.
I realised that in 2020 I would have assumed that an AI car autopilot would have no hope with conflicting logical constraints such this, reading English traffic signs, etc...
Now? In early 2024? There are multi-modal AIs that can read arbitrary road signs, understand the text (in practically any language), and follow the logic to a correct conclusion.
Would I trust GPT-4V to control my car? No, the technology just isn't ready yet.
However, it's clearly possible now, whereas very recently it seemed impossible.
Currently, Tesla autopilot is a complex mix of multiple models, a bunch of C++ code, and more. It's "traditional software" with bits of AI in it.
I suspect that within just a few years, a pure AI model could be trained directly on video collected from the Tesla fleet, in much the same way that GPT was trained with text from the Internet and books. Sprinkle in a pre-trained GPT-5 level LLM for reading and logic, and ta-da, you have a generally intelligent driver agent that can understand spoken instructions, can read road signs, receive traffic alerts over 5G, and react appropriately. It would have learned from billions of hours of driving and a million accidents. Just like GPT, it'll have a wider range of experience than any human, prepared for any eventuality that actually occurs on roads.
It’s like the difference between SpaceX Starship development and faster-than-light spaceship development.
We know that we can make a bigger reusable rocket, it’s just a matter of working out the details. We have no idea if FTL is even possible, let alone how.
I feel like AI is at the same point: technology demonstrators have proven that there is a way forward, now we just need oodles of computer power to make it happen.
Chat bots that can answer complex English questions used to be pure science fiction just a year ago!
Things have been improving rapidly. Try interpreting AGI conservatively, i.e. not ASI, not alive or full simulation of a human, just something that can do all or almost all human cognitive tasks.
GPT-4 can handle many human tasks with the right configuration.
There might be multiple ways to improve things like reasoning. Larger, more efficient and better curated datasets. Or maybe there is a way to embed some kind of TorchOpt stuff inside the models. Baking in batches of tree of thought. Micro development environments or logic programming tightly integrated with inference. Etc.
Also, the emphasis on more animal-like abilities than LeCun is focusing on could start to pay off within a couple of years.
Given a conservative interpretation of "AGI", I take what Altman said at face value. They need to have ambitious goals because of the rate of progress, level of competition, high stakes, and very high level talent.
When someone says "AGI" what does that actually mean to you today? Every time I talk with others about this it feels like we have wildly different interpretations and expectations. Similar to "cloud native", "crypto", et. al.
For me, the most basic level of "AGI" is: I can provide a human instruction and the intelligence will iteratively work through a known problem space with known tools until it reaches a satisfactory outcome, or realizes it is unable to complete the task with an explanation regarding why.
How this outcome is achieved is not really relevant to me. The effect on the business is all that really matters.
The LLM doesn't seem like the important part anymore. We figured out a pretty damn good hammer. It bangs those nails in clean pretty much one swing every time. We need to stop focusing on the hammer and begin focusing on the framing.
The "ChatGPT for teams" and the underwhelming GPT App store would be unnecessary if the headline was true. It more looks like they are looking to build out a product based on the existing ChatGPT capabilities. I predict any improvements will be slow and iterative.
I seriously doubt we will reach AGI in 2025, or next iteration of GPT after GPT-5.
What this does however suggest they have something that will make substantial improvement in next GPT, and something to further improve on the one after.
What may likely to happen is that GPT-6 ( In 2026 or later rather than 2025 ) will be good enough for lots of thing, as we further refine it and improve on it for another 1 or 2 iteration. This whole process in the next 5 - 6 years will unlock a lot of value in other industry. Enough to fuel another 10 years of investment cycle.
I still dont believe we will reach AGI by 2030 though.
NOTE: I know nothing about this area besides what I read from tech articles.
The existance of multiple centers of decision on latest LLMs surprised me.
I'd venture to guess either locality nodes will evolve or will need to be made happen to reduce the necessary bandwidth for LLM.
By a jump of logic, I'd also venture to guess that will need to make it happen before we reach AGI.
Additionally, physiological nodes will also need to happen before we have human-like AGI.
Either way, just making things up here. Not a professional.
I’m a little sceptical of AGI next year. Even if pushed to its limit it’s just 2 years. I suspect some sort of shenanigans. Like, rebranding multi-modal AI to general AI. I guess, one gets to invent a new term if your model can, say, process text, sound, and video, let’s add spatial manipulation, too, just to differentiate from current multi-modal models. Or G might stand for somthing else than “general”.
Another aspect is that Altman doesn’t say any specific dates, even years. Two big step ups in two years seem like a little too fast.
Seems hyperbolic, but if AGI really is arriving in 2025, what would you build?
Would it be worth building anything? AGI will be able to build it far cheaper, and in any case, who's going to use a travel-booking website or whatever, if they can just tell an AGI to 'book me a holiday', etc.
The only thing that makes sense is a lunge for patents, trademarks, etc, ie ownership.
One part I really don't understand is, so far any GPT sucks at math or makes very basic mistakes with numbers. There were so many examples. But how will design of next LLM will improve that with making math proficient? AGI without math would not make any sense?
Say your service would cost an unsustainable amount with GPT 4 because of excessive prompting needed, or you need to retry X out of every Y generations.
You could eat the cost temporarily knowing when GPT 5 arrives you'll be able to cut down on that significantly.
Or you build a simpler product that gets a foot in the door in organizations knowing that you'll have an additional value proposition that's enabled by an incremental improvement in capabilities.
Sama is another one. He is part of some sort of success story, makes a bunch of radical 4d chess type claims, so you can’t discount everything he says, however, even if what he says is true, how do you “build for AGI”? The dude sounds legit bonkers, but again, he has some type of cred because his team delivered a working product. It almost feels like this is a common play adopted by a lot of tech elites. It’s not a coincidence.
“If you find yourself part of some success story, any success, use it to your advantage. Use your ‘cred’ to drive a lot of hype and cash out, using all the cash you accumulate hire hard working people, have them deliver something which in turn keeps you looking ‘credible’, make more outlandish claims, gain more funding, rinse wash repeat”
I feel Jeff Bezos is an example of someone who doesn’t do this. Seems like a practical guy who works really hard and avoids (mostly) making silly claims.