Open AI in Trouble
garymarcus.substack.com
garymarcus.substack.com
Concentrated capital is truly a wild thing.
Not only can they not reach AGI, they cannot reach their own definition of AGI. But people will still gobble whatever next lie Altman will sell them for $2,000 a month.
Absolutely pathetic.
that definition likely not their's but came from Microsoft when they dumped 10B into OAI and was a condition for revenue sharing clause.
One can argue that AGI is already achieved, LLMs are more proficient and general in knowledge tasks than any specific individual.
This is not AGI.
Let's take an old version of the Wikipedia page about AGI, from end of 2022, before the ChatGPT craze: https://en.wikipedia.org/w/index.php?title=Artificial_genera...
> Artificial general intelligence (AGI) is the ability of an intelligent agent to understand or learn __any__ intellectual task that a human being can.
LLMs are nowhere near this. Not even close.
But let's get even more restrictive by looking at the actual required characteristics to declare a system an AGI:
> There is wide agreement among artificial intelligence researchers that intelligence is required to do the following:
> - reason, use strategy, solve puzzles, and make judgments under uncertainty;
> - represent knowledge, including common sense knowledge;
> - plan;
> - learn;
> - communicate in natural language;
> - and integrate all these skills towards common goals.
> Other important capabilities include:
> - input as the ability to sense (e.g. see, hear, etc.), and
> - output as the ability to act (e.g. move and manipulate objects, change own location to explore, etc.)
Once again, LLMs are not even scratching the surface of AGI.
Of course if you look at the current version of the page, where the bar has been massively lowered, it might look like AGI is achieved, but that's because the definition has changed, not the technology.
sure, its not AGI by old wikipedia definition, which imo is human-centric and more definition of superintelligence (requires ability to exceed all existing humans in all tasks). But it is AGI by current wikipedia definition.
Reading wikipedia about history of the term, it sounds like term was popularized by this book: https://www.amazon.com/Artificial-General-Intelligence-Cogni... which says that general means "ability so solve variety of tasks in variety of domains", not "all tasks in all domains". So "always meant" is easily challengeable here.
DNA doesn't understand intelligence.
"The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." - Edsger Dijkstra.
And what we have now, with LLMs, met the standard I had for AGI just five years ago.
Only thing that changed for me since then is noticing that not only does everyone have a different definition of what AGI even is, also none of the three initials are even boolean-valued: "how intelligent" can be a number, or even a vector that varies by domain as linguistic intelligence doesn't have to match spatial intelligence; "how general" could be what percentage of human cognitive tasks the AI can do; "how artificial" is answered entirely differently by those who care differently about learning from first principles vs. programmed algorithms vs. learning from reading the internet.
> People were misled by graphs that they were told pointed to a singularity right around the corner, despite obvious errors in the extant systems.
Tautology.
If there weren't obvious errors in the extant systems… there wouldn't be anything left to do.
What study are you referring to here? The ones I've seen don't particularly resemble anything Turing described. The preprint from UCSD got heavy coverage in the popular press but its headline claim is an elementary statistical mistake, modifying the test so it's no longer a binary comparison but still treating 50% as a meaningful pass threshold. It hasn't yet passed peer review nine months later, and I'm cautiously optimistic that it never will.
this is probably a religious question so people will not be convinced of intelligent machines even if they have an EQ and IQ of 200 because they don't work the same way human cells do I suppose.
I still believe GPT-4 has some general intelligence, just a very tiny amount. It can take what it was trained on and slightly modify it to answer a slightly novel question.
And me running is closer to the speed of light than me walking, yet neither are even remotely close to the speed of light.
The specific term "AGI" has always referred to a program that strictly matches to human intelligence, with no significant areas where it's worse than a typical human. I agree that we're not there and not obviously close to there. But the idea that it's all a parlor trick and LLMs have no cognition at all seems obviously false to me.
Today, LLMs being trained to output text, images and video equal to or better than any human is now what we consider to be AI.
There is a semantic game we play to move the goalposts at every new tick of these technologies so that humans are still on top.
Maybe the next big trick is to define AGI as having a system that was not trained on the entire corpus of human output available on the internet, and train it with nothing. Can an AI be “born” with no model like a baby and learn to speak, walk, function in society without being primed for success with a state of the art model? Could you drop this AI into any human society, regardless of culture, and have it seamlessly integrate?
What precisely we consider “cognition” feels like a philosophical debate, one where we have no single true answer. And doesn’t really matter?
I do think it's true that anyone a decade ago would have predicted their capabilities are impossible without AGI, which suggests that the entire conceptual framework is less useful than we thought. The eventual debate about whether we've "reached AGI" will probably have a lot more to do with the contract between OpenAI and Microsoft than any real paradigm shift in capabilities.
That is AI. AI ≠ AGI.
Literally millions or billions of people are better than me at a significant number of intellectual tasks
"Nerds with a thousand GPUs,
Nerds with a thousand streams,
with you only I experience,
the love, the chat of my dreams!"
"So prompt me, and list me,
index me, repeat me,
I'm yours till I elide,
so in love, so in love,
so in love with you, my GPT, am I. "
-With apologies to Cole Porter and anyone blinded by reading this post.
Having said that, 4.5 is clearly a misstep, one that should realign the goals of the AI industry to focus more on utility and cost effectiveness.
Is this a requirement for achieving AGI? The history of progression of the ML field indicates that the answer is "no". We don't really understand how concepts are encoded in today's models, yet that doesn't stop them from being economically useful. So why would the special case of AGI be any different?
Ans. No, no and the entirety of human cognition.
ChatGPT et al contain collections of words. When prompted they generate word sequences found in known word sources. That's all. They don't observe or reason about the world. They don't even reason about words. They merely append the most likely next word to a sequence of words.
Orion does seem to have been a failure, but I also find it a bit weird that they seemingly decided to release the full model rather than a distillation, which is the pattern we now usually see with foundation models.
So, did they simply decide that it wasn’t worth the effort and dedicate the compute to other, better things? Were they pushed by sama to release it anyway, to look like they were still making progress while developing something really next gen?
I agree that OpenAI's endless hyping about AGI seems pretty unrealistic, but let's take a breather here. There are no major research projects where you don't run into setbacks and failures before you reach your goals.
And even if they never really progress beyond where they are today with their current models, just bringing down the cost could open up a lot of doors for useful applications.
Someone will figure out how to make AIs understand their own ignorance and stop bullshitting when they don't know something, someone will figure out how to make AIs learn on the fly instead of fine-tuning new model versions, etc.
You either accept that change happens and use your life experience to help shape it in a positive direction or, well… I dunno. Become a old curmudgeon and watch the world blow by you.
And your tech examples don't offload actual process of thinking to others the way AI does. The comparisons are surface level and ignore what I actually said.
The current state of LLMs _can_ be helpful for education. Millions of people use them as such and benefit from it.
A far bigger problem with the technology IMO is the generative aspect. We already have a large problem with disinformation and spam on the internet, and generative AI will increase this by many orders of magnitude. Discerning fact from fiction is already difficult today; it will be literally impossible to do in the future, _unless_ we invent more technology to save us from it. This is a problem we haven't even begun to address. The public is collectively blinded by the novelty of the technology, while entrepreneurs are jumping over themselves trying to profit from the latest gold rush. Very few people with the power to change anything are actually thinking about the long-term, or even mid-term, impacts.
Curriculum isn’t moving fast enough. Just a few years ago every teacher had to adapt for COVID and go 100% online/remote in many areas. Kids are still turning in every assignment online, even in classroom settings in my district.
So, yeah, kids can just paste the question into ChatGPT and copy the answer. Nobody learns anything.
This isn’t AI ruining education, it’s schools being under-resourced and unable to move as quickly as society is changing. Teachers are still buying their own supplies, how can they adapt their entire curriculum in the course of a couple years to work under this entirely new paradigm of LLMs?
Give it a bit and I am convinced schools will go back to oral reports, handwritten essays and whatever is needed to make sure children are not just pasting garbage back and forth
Honestly, I think this is what we need to kill toxic social media and phone addiction as well. If AI forces us to talk and interact as a community again, it’s a win. Leave the internet to the bots and AI.
I was talking with someone in a non-tech industry and we have such a long way to go for even decades-old information system improvements. They don't even have basic things like an effective system of record or analytics. They have no way to measure the success of any particular initiative. If revenue is down - they don't really know why, they just randomly change stuff until it hopefully goes back up.
That said, annoying people will move on to hyping something else.
Also, it’s inefficient allocation of capital. Every cent being spent on this is money that could be spent on something useful (of course, absent the AI bubble, not _all_ of it would be, but some of it would be).
This layer itself will inevitably see its cost come down per unit of use on a long road towards commoditization. It will probably get better and better and more sophisticated but again the value will be primarily up stack, not accrued primarily from a company like this. It's not to say they couldn't be a great company... even Google is a great company that has enabled countless other companies to bloom. The myopic way people look to these one size fit all companies is just so disconnected from our economy works.
If we are waiting for a new breakthrough architecture, it could be decades. Our brains send signals with very high concurrency to individual processors (neurons). Each neuron is massively more complex than a single ReLU function and we have billions of them. If we need that kind of parallel processing and data transfer to match human thought and flexibility, it could be another 70 years before we see AGI.
That said, I do think LLMs are one of the biggest AI breakthroughs since the inception of AI. And I am sure that it, or something very similar will be part of an eventual AGI.
I could not agree more. There is much value still to be gained from blockchain technology!
But here, I think he's right about business matters. The massive investment in computing capacity we've seen in recent years, by Open AI and others, can generate positive returns only if the technology continues to improve rapidly so it can overcome its limitations and failure modes in the short run.
If the rate of improvement has slowed down, even temporarily, OpenAI and others like Anthropic are likely to face financial difficulties.
---
[a] In the words of Geoff Hinton: https://www.youtube.com/watch?v=d7ltNiRrDHQ
---
Note: At the moment, the OP is flagged. To the mods: It shouldn't be, because it conforms to the HN guidelines.
A business that burns money at the rate OpenAI does, without any clear path to profitability, will eventually die.
If GPT 4.5 had nothing of value to offer other than a flex of opening eyes ability to scale, it’s still a signal that they have the chops to throw the most amount of compute at the upcoming reinforcement learning race.
If you’re actually selling enterprise solutions in this space, you’ll quickly learn that a large number of Enterprises have their own private deployment of open AI on Azure and are pushing their vendors to use that even if it means they get lower quality outputs for certain use cases.
Data and model quality aren’t the only moat-able features.
This is a clear signal that the people remaining to run the show at OpenAI may not actually know what they're doing. The big names building the technology have fled from Altman's power grab. They all have startups of their own.
Meanwhile, the people at Anthropic keep making Claude better and better. No bombast, just tech. And then there's the model not on the list of many of these shoot-outs: Grok 3. That one has scale, and it's actually good.
"We can sorta kinda also make a big AI!" is not the message you want to be sending. It has to clobber the other "main" models out there.
I think you are wrong that this is the limit for single shot though. We have reached a limit due to data, but I expect what will happen next is that we will transition into a slow growth phase where chain of thought models will be used to effectively create more training data, which will then be used to train a better single shot model, which will then be extended into a better chain of thought model, which will then produce higher quality training data for another single shot. And so on. Kind of like what happens as knowledge passes from teacher to student across successive generations. Effectively a continuing process of compression and growth in intelligence, but progressing rather slowly compared to what we have seen in the last 5 years.
The compute for training is beginning to seem a poor investment since it is depreciating fast and isn't producing value in this case. That's a seriously big investment to make if it's not productive but since a lot of it actually belongs to Azure they could cut back here fast if they had to. I hope they won't because in the hands of good researchers there is still a real possibility that they'll use the compute to find some kind of technical innovation to give them a bigger edge.
So does Google. And Google can roll out their premium models into phones, household devices, cars and online platforms to add value.
OpenAI has a website.
It's not even close.
Also, even though LLMs can generate text much faster than humans, we may be internally thinking much faster. Each adult human brain has over 100 billion neurons and 100 trillion synapses, and each has been working every moment, for decades.
This is what separates human reasoning from LLM reasoning, and it can’t be solved by scaling the latter to anything feasible.
I wish AI companies would take a decent chunk of their billions, and split it into 1000+ million-dollar projects that each try a different idea to overcome these issues, and others (like emotion and alignment). Many of these projects would certainly fail, but some may produce breakthroughs. Meanwhile, spending the entire billion on scaling compute has failed and will continue to fail, because everyone else does that, so the resulting model has no practical advantages and makes less money than it cost to train before it becomes obsoleted by other people’s breakthroughs.
Disclosure - I am neither bearish or a mega bull on LLMs. LLMs useful in some cases.
Right now, these models are built by the establishment to serve the establishment—and that’s a load of shit that needs to change. The real fun starts the day some random group of “anonymous” on 4chan can train one of these models to generate incredibly convincing deepfakes of world leaders, all using the compute power in their own homes.
Power to the people. Fuck the system, and all that jazz.
A smart enough AI would summarize each of his posts as "I still hate the current AI boom".
There must be a term for such writers? He's certainly consistently on message.
D'oh!
Manhattan Project: Started 1942, combat use 1945.
Solid state transistor: Invented 1954, in radios 1955.
Hybrid cars, iPhone, etc etc.