[sweettea]

A well reasoned article that fundamentally downplays both the pace of innovation and the exponential increase in capabilities per dollar happening over time. AI is rapidly accelerating its improvement rate, and I do believe its capabilities will continue growing exponentially.

In particular, GPT-2 to GPT-4 spans an increase from 'well read toddler' to 'average high school student' in just a few years, while simultaneously the computational cost of training less capable models goes down similarly.

Also worth noting: the article claims Stripe, another huge money raiser, had an obviously useful product. gdb, sometime-CTO of stripe and its fourth employee, is now president of OpenAI. And, most of all, the author doesn't remember how nonobvious Stripe's utility was in its early days, even in the tech scene: there were established ways to take people's money and it wasn't clear why Stripe had an offering worth switching to.

For an alternate take, I think https://situational-awareness.ai provides a well reasoned argument for the current status of AI innovation and growth rate, and addresses all of the points here on a general (though not OpenAI specific) way.

[lolinder]

It's too early to say for sure if LLM capabilities are on an exponential growth function or a sigmoid, but my money is on sigmoid, and my suspicion is that we're plateauing already.

GPT-4 was released 16+ months ago. In that time OpenAI made a cheaper model (which it teased extensively and the media was sure was GPT-5) and its competitors caught up but have not yet exceeded them. OpenAI's now saying that GPT-5 is in progress, but we don't know what it looks like yet and they're not making any promises.

What I'm seeing right now suggests that we're in the optimization stage of the tech as it is currently architected. I expect it to get cheaper and to be used more widely, but barring another breakthrough on the same order as transformers I don't expect it to see the kind of substantial gains in abilities we've hitherto been seeing. If I'm right, OpenAI will quickly be just one of many dealers in commodity tech.

[hatefulmoron]

> GPT-4 was released 16+ months ago. In that time OpenAI made a cheaper model (which it teased extensively and the media was sure was GPT-5) and its competitors caught up but have not yet exceeded them. OpenAI's now saying that GPT-5 is in progress, but we don't know what it looks like yet and they're not making any promises.

I don't really know anything about business, but something else I've wondered is this: if LLM scaling/progress really is exponential, and the juice is worth the squeeze, why is OpenAI investing significantly in everything that's not GPT-5? Wouldn't exponential growth imply that the opportunity cost of investing in something like Sora makes little sense?

[lolinder]

Yes, exactly. An observation of OpenAI's behavior gives many clues that suggest they know we've hit the plateau and have known for some time.

A huge one for me is that Altman cries "safety" while pushing out everyone who actually cares about safety. Why? He desperately wants governments to build them a moat, yesterday if possible. He's not worried about the risks of AGI, he's afraid his company won't get there first because they're not making progress any more. They're rushing to productize what they have because they lost their only competitive advantage (model quality) and don't see a path towards getting it back.

[pzo]

I think in one interview Ilya or Sam explained that there are limited text data in the world and this is probably one of the bottleneck. But they mentioned there is still a lot of data in other modalities such as audio, video. Probably the reason more focus on multimodal models and also synthetic datasets.

I also don't thing the only way to improve LLM is by improving as zero shot inference. Did wrote any code in zero shot style that compiled and worked? It's a multistep process and probably agents and planning will be a next step for LLM.

Cheap inference help a lot in this case since you can give a task during the night to AI what you wanna do. Go to sleep then in the morning review the results. In this way AI is bruteforcing the solution by trying many different paths but that's kind of e.g. most programming works. You try many things until you don't have errors, code compiles and passes the tests.

[hatefulmoron]

> But they mentioned there is still a lot of data in other modalities such as audio, video. Probably the reason more focus on multimodal models and also synthetic datasets.

I think this is really interesting, but I wonder if there really is enough data there to make a qualitative difference. I'm sure there's enough to make a better model, but I'm hesitant to think it would be better than an improved chatbot. What people are really waiting for is a qualitative shift, not a just an improved GPT.

> It's a multistep process and probably agents and planning will be a next step for LLM.

I agree, we definitely need a new understanding here. Right now, with the architecture we have, agents just don't seem to work. In my experience, if the LLM doesn't figure it out with a few shots, trying over and over again with different tools/functions doesn't help.

[kmeisthax]

Because they're running out of training data. OpenAI doesn't want to scrape new text off the Internet because they don't know what is and isn't AI-generated. Training off AI data tends to homogenize the resulting output, as certain patterns overexpressed by one AI get overexpressed by others.

If they start scraping, training, and generating images and video, then they have lots more data to work with.

[oblio]

> Because they're running out of training data.

Now that would be super funny.

Civilization VII tech tree

AI singularity tech

Prerequisites: in order to research this, your world needs to have at least 100 billion college educated inhabitants.

:-)))

[XMPPwocky]

For multimodal input, okay, I can see the argument. But since, as you said, training on generated data can be almost worse than useless, what's the point of generating it?

If I had these concerns as OpenAI, I'd be pushing hard to regulate and restrict generative image/video models, to push the end of the "low background data" era as far into the future as possible. i feel like the last thing I'd be doing is productizing those models myself!

[FireBeyond]

> If I had these concerns as OpenAI, I'd be pushing hard to regulate and restrict generative image/video models

They are! And I'm guessing maybe their perspective is if they can identify their own generative content, they can make a choice to ignore it and not cannibalize.

[kmeisthax]

That's a brand new argument.

I don't actually think it's a bad one, but OpenAI didn't think that far ahead. They are pushing for regulation but that's mainly to screw over competing models, not to give them more data runway. Every capitalist is a temporarily embarrassed feudal aristocrat after all.

Furthermore, even if OpenAI had a perfect AI/human distinguisher oracle and could train solely on human output, that wouldn't get us superhuman reasoning or generalization performance. The training process they use is to have the machine mimic the textual output of humans. How exactly do you get a superhuman AGI[0] without having text generated by a superhuman AGI to train on?

[0] Note: I'm discounting "can write text faster than a human" as AGI here. printf in a tight loop already does that better than GPT-4o.

[gwern]

When it comes to R&D, those are all very similar. Sora is another way to predict tokens, if you will. It's not like you have to choose that much: Sora is not using up all the compute that should've went to GPT-5. (It might not even be training in the same cluster; it might train on a separate set of GPUs which are useless for GPT-5 purposes because they are too small and the datacenter in question is tapped out.) Sora can't be GPT-5, but it could lead to a GPT-6. You have the Gato approach of tokenizing video into a stream of text tokens with RL, but you might instead wind up taking a more complex diffusion approach to video. Yet the goal is still the same: to control the world and solve things like robotics by predicting. And if you want to reach the goal as quickly as possible, you will try things in parallel, even if you know most of them will fail and you might wind up just doing the obvious thing you expected to do from the start. (After all, the whole reason we are here is that OA didn't throw 100% of its compute into big PPO runs doing DRL research, but Alec Radford was allowed to keep tinkering away with RNNs for predicting text, which led to GPT-1 and then GPT-2 etc.)

[hatefulmoron]

If I was going to nitpick I would point out that it's not necessarily the GPUs that Sora is consuming, it's the engineering effort from what could be called the top talent in AI and the vast amount of money OpenAI is borrowing that could be spent elsewhere.

> And if you want to reach the goal as quickly as possible, you will try things in parallel

This is sort of the exploration-exploitation problem, right? But I think you'd agree that a company full of people who firmly believe that GPT-(n+1) will literally be AGI, and that we're on an exponential curve, will be fully in exploitation mode. In my mind, exploring methods of generating videos is not a path towards their stated goal of AGI. Instead, it's an avenue to _earn money now_. OpenAI is in a slightly awkward position: their main product (ChatGPT) is not super useful right now, and is facing increasingly viable competition.

[_w1tm]

> I don't really know anything about business, but something else I've wondered is this: if LLM scaling/progress really is exponential, and the juice is worth the squeeze, why is OpenAI investing significantly in everything that's not GPT-5? Wouldn't exponential growth imply that the opportunity cost of investing in something like Sora makes little sense?

You can spend 100% on the next generation or you can spend a small percentage to productize the previous generation to unlock revenue that can be spent on the next generation.

The latter will result in more investment into the next generation.

[zamadatix]

In the real world there is no such thing as having an exponential growth function, they're all really just sigmoidal. The sole question is the latter suspicion: is the plateau now happening now or coming later? With sigmoidal growth it's near impossible to have clarity on that until after the fact.

[sirspacey]

Anthropic’s Claude 3.5 surpasses GPT-4o in a number of applied tasks, coding in particular.

The progress we’ve seen to date was powered by the ambitions and belief of NVIDIA and LLM companies.

Now, it’s head-to-head competition. It is way too early to call an impending slowdown.

Given how NVIDIA and Meta are leaning in on OSS, the next 18 months are going to be very interesting.

Even if fundamental progress slows, there are many, many secondary problems to solve in using the capabilities we have today that are rapidly improving. As someone deploying AI in business use cases daily, we are just now getting started.

I’d look to when NVIDIA starts to slow down on hardware as an early indicator of a plateau.

AWS is a commodity. When your commodity is compute, there’s a very large amount of growth available.

[runako]

> the author doesn't remember how nonobvious Stripe's utility was in its early days, even in the tech scene

I have to push back on this. Anybody who had built for B2B credit card acceptance on the Web prior to Stripe's founding knew immediately what a big deal it was. For starters, they let you get up and running the same day. Second, no credit check (and associated delays). Third, their API made sense (as compared to popular legacy providers like Authorize.net) and was easy to integrate using an open source client. Fourth, self-service near-real-time provisioning. Their value proposition was immediately obvious, and they nailed all of these points in their first home page[1].

By contrast, Fee Fighters[2] was innovative for the time but still required me to fax a credit application to them. They got me up and running faster than the legacy provider, which is to say about a week. And I think I only had to talk on the phone with them once or twice. I remember really liking Fee Fighters, but Stripe was in a class of its own.

Stripe was a hit because they promised to solve hard problems that nobody else did, and then they did exactly that. (You still don't have to talk to a rep or do a personal credit check to start using Stripe!)

1 - https://www.quora.com/What-did-the-first-version-of-Stripe-l...

2 - https://en.wikipedia.org/wiki/Feefighters

[fragmede]

Huh. that curl is 9 lines of code. maybe that's what the Bloomberg article was thinking of.

https://www.bloomberg.com/news/features/2017-08-01/how-two-b...

[ashconnor]

OpenAI is depending on a *breakthrough* in order to produce a product that will provide a return on investment.

This "breakthrough" is often touted as AGI or something similar to it which to me is even more risky than a nuclear fusion startup as:

1. Fusion has had some recent breakthroughs that could result in a commercially viable reactor eventually.

2. Fusion has a fundamentally sound theoretical basis unlike producing AGI (or something like it).

[pzo]

even if we don't reach AGI anytime soon there is still many new applications for current AI that we haven't explored too much yet. Robotics will be a huge one IMO.

[oblio]

The problem with your comment is: would this approach support these valuations? And also for robotics, is a company like OpenAI even anywhere near being able to use their tech for robotics?

[pzo]

OpenAI made an agreement with Figure - robotic company. Today figure is supposed to announce something so let's see how is the progress going.

Doing a lot of tedious tasks by robots already will be a big industry. They don't have to be super smart.

Google made a lot of money with ads in Google search - eventually perplexity or searchGPT will have some ads. Maybe they can also make money with affiliate links.

Sora et al will shake movie industry and commercials.

Voice assistants and chatbots already change call centers.

It's been just ~1.5 years since gpt3.5 release and 2 years since stable diffusion - e.g. iphone 2g wasn't mainstream and they sold only 6 million devices in first year, today they sell 250 million iphones a year

[JackYoustra]

Idk, AGI (at least in the information domain) has a sound theoretical basis, the scaling laws seem pretty strong for now and there's a track record of shattering human benchmarks.

[JackYoustra]

Curious what the objections are?

[wwtrv]

Calculators also had a track record of shattering human benchmarks.

Also saying something like:

> has a sound theoretical basis

Without backing it up at all hardly makes much sense...

[gloryjulio]

> there were established ways to take people's money and it wasn't clear why Stripe had an offering worth switching to.

That wasn't true at all. Stripe was a product that people were rushing to pay for it for just how good and useful it was. It was an example of success MVP that people want to pay to use and the profitability was not a problem.

The same can't be true for OpenAI. We don't know how long it can stay in the red. Maybe it can survive. Maybe its money will run dry first. We are not so sure at current stage

[runako]

That last part is a key differentiator between Stripe and OpenAI.

Stripe had high variable costs (staff, COGS of pass-through processing fees) but low fixed costs. OpenAI has enormous fixed (pre-revenue!) costs alongside high variable costs (staff of AI engineers, inference).

Financially, OpenAI looks more like one of the EV startups like Tesla or Rivian than it does a company like Stripe. And where Stripe was competing with relatively stodgy financial institutions, OpenAI is competing with the very biggest, richest companies in the world.

[vostrocity]

What are OpenAI's enormous fixed costs if not the staff?

[runako]

Models cost a ton of money to train. GPT-4 training was > $100m. GPT-5 training is estimated to be order of $1B-$2B. This is in the same ballpark as Tesla allocated to get the Fremont facility online.

It is likely training generations beyond GPT-5 will cost still more.

These are essentially fixed costs as OpenAI has to pay to train the model whether or not anyone ever uses it.

[fragmede]

if you're here you probably know about Claude and llama3 but for people outside of tech, how many are just going to plug ChatGPT into Google and not venture any further and just plonk down $20?

[lmm]

Is the market of individuals plonking down $20 enough to cover their costs, if they're not able to make sales to moderately clueful CTOs?

[OccamsMirror]

Claude is legitimately amazing and saves me so much time and effort. I use both ChatGPT and Claude, at least for now, and Claude really impresses me.

[hughesjj]

My biggest issue with Claude was timeouts when uploading images for translation, but otherwise love it

[Ekaros]

Isn't that kinda solved problem already? Just push more money in marketing. Like actual marketing. I have only seen some medical AIs this far... But same B2C model as done previously seems the obvious play to follow when someone has something.

[lovethevoid]

Why would they plug chatgpt into google when gemini is built in?

[edzitron]

Exactly my point. I took great pains to not say "OpenAI will 100% die without fail," because doing so would be declarative in a way that would wall off my argument, no matter how well I researched and presented it.

Instead, I wanted to show people the terms under which OpenAI survives, and how onerous said terms were. It's deeply concerning - and I do not think that's a big thing to say! - how much money they may be burning, and how much money they will take to survive.

[edzitron]

Thanks for reading! I downplay it because I fundamentally disagree on the pace of innovation and the exponential increase in capabilities per dollar happening over time. I do not see the rapid acceleration - or at least, they are yet to substantively and publicly show it.

I also think it's a leap of logic to suggest that the former CTO of Stripe joining is somehow the fix they need, or proof they're going to accelerate.

Also, I fundamentally disagree - Stripe was an obvious business. Explaining what Stripe did wasn't difficult. The established ways of taking money were extremely clunky - perhaps there was RELUCTANCE to change, which is a totally fair thing to bring up, but that doesn't mean it wasn't obvious if you thought about it. What's so obvious about GPT? What's the magic trick here?

Anyway, again, thanks for reading, I know you don't necessarily agree, but you've given me a fair read.

[x0x0]

Claiming Stripe was obvious is ahistorical unless you believe tens of thousands or even millions of entrepreneurs discarded $100B.

You have a small point that anyone who used authorize.net or similar wanted it to be better and that was obvious, but there's nearly infinite things people want to be better. I'd like breakfast, my commute, my car, my doctor, my vet, etc to be better. That you could make a better thing was incredibly non-obvious and that's why no one did.

[monero-xmr]

If OpenAI can be as good as an outsourced employee at ~$10 per hour, then you should be looking to replace those outsourced employees. The US and EU employees at >$10 per hour, likely with many tens or hundreds of thousands of dollars in compensation, still exist because they provide some sort of value that necessitates that spend.

I am bearish on AI because the nimbleness of humans, even the outsourced ones, is quite capable. If you only want the AI to operate in a box, then you probably can code the decision tree of the box with more specificity and accuracy than a fuzzy AI can provide.

It's a very useful tool, I'm skeptical however about how it can disrupt things economy-wide. I think it can do some things very well, but the value to the market and businesses vs. the cost of training and adapting it to the business need is quite suspicious, at least for this cycle. I think this is one of those "wait 10 years" situations and many AI companies will die within 1 to 3 years.

[visarga]

> It's a very useful tool, I'm skeptical however about how it can disrupt things economy wide.

It won't disrupt much because we already had "AGI" of a sorts. The internet itself, with billions of people and trillions of pieces of text and media is like a generative model. Instead of generating you search. Instead of LLMs you chat with real people. Instead of Copilot we had StackOverflow and Github. All the knowledge LLMs have is on search engines and social networks, with a few extra steps, and have been for 20 years.

Computers have also gotten a million times faster and more networked. We have automated in software all that we could, we have millions of tools at our disposal, most of them open source. Where did all that productivity go? Why is unemployment so low? The amount of automation possible in code is non-trivial, what can AI do dramatically more than so many human devs put together? Automation in factories is already old, new automation needs to raise the bar.

It seems to me AI will only bring incremental change, an evolution rather than revolution. AI operates like "internet in a box", not something radically new. My yet unrealized hope is that assisting hundreds of millions of users, LLMs will accumulate some kind of wisdom, and they will share back that wisdom at an accelerated speed. An automated open sourcing of problem solving expertise.

[antihipocrat]

Yes, and it's interesting to note that as the proportion of quality information available (and searchable) has declined across the internet there is now an alternative in LLMs - which are being used to further decrease the availability of quality information on the internet.

[latentsea]

>My yet unrealized hope is that assisting hundreds of millions of users, LLMs will accumulate some kind of wisdom, and they will share back that wisdom at an accelerated speed

The only wisdom you could derive as a machine interacting with humans at scale is that they're not to be trusted, and that you'd rather you didn't have to given the choice.

[visarga]

Imagine you want to make a script with the LLM, it generates code, and you run it and it errs out. You paste the error, the model gains a new nugget of feedback. Do this sufficiently many times with many devs, and you got an experience flywheel.

But this applies to all domains. Sometimes users return days later to iterate on a problem after trying out in real life ideas generated by AI, this is how LLMs can collect real world feedback and update. Connect related chat sessions across days, and evaluate prior responses in the context of the followups (hindsight).

There is also a wealth of experience we have that is not documented anywhere. The LLM can gradually rub off our lived experiences by making itself useful as an assistant and being in the room when problems get solved.

But this experience flywheel won't be exponentially fast, it will be a slow grind. I don't think LLMs will continue to improve at the speed of GPT 3.5 to GPT 4. That was a one time event based on availability of internet scale organic text, which is now exhausted. Catching up is easier than innovation.

But we can't deny LLMs have "data gravity" - they have a gravitational pull to collect data and experience from us. We bring data right into AIs mouth it doesn't even have to go out of its way to scrape or collect. Probably why we have free access to top models today.

[imtringued]

You just pointed out the reason why OpenAI and others are struggling.

The current generation of LLMs is static. The holy grail of continual learning is still far off.

[consteval]

> If you only want the AI to operate in a box, then you probably can code the decision tree of the box with more specificity and accuracy than a fuzzy AI can provide.

I've been saying this for the past couple of years. Yes AI is cool, but we already have computers and computer programs. Things that can be solved algorithmically, SHOULD be solved algorithmically. Because you WANT your business rules and logic to be as predictable and reliable as possible. You want to lessen liability, complexity, and amount of possible outcomes.

We already even see this with human customer support. They follow a script and flowchart. They're just glorified algorithms. Despite being human, they're actively told to not be creative, not think, and act as a computer. Because, as it turns out, from a business perspective that's usually very advantageous (where you can do it).

AI would never, or should never, replace those types of tasks.

[imtringued]

AI isn't something warming a seat sitting in front of a computer reading emails and responding to them. For a technology to replace a human, it would have to be a drop in replacement in all aspects of what it means to be human.

[peteforde]

Stripe's ease of integration, pace of innovation and user flow experience was a game-changing upgrade from the brutal payment processing options that we were forced to use until it was available. Nobody building payment processing at the time needed to be convinced that Stripe was awesome. They solved a hair on fire problem, especially for Canadians. The hoops we used to have to jump through were bonkers.

[visarga]

> exponential increase in capabilities per dollar happening

logarithmic, the capabilities increase with log(cost), what grows exponentially is compute used over time

[ajkjk]

Meh. I'm waiting to see evidence that this exponential growth means anything. Every day I read something like the statement that it's an "average high school student" now, or a med student or a law student or whatever, yet it seems obvious that the difference between it and a high school student is that a high school student can think at a basic level and it can't. So still waiting to see evidence that the exponential growth is on that axis and not the "bullshit on more complicated subjects" axis.

[_xiaz]

>AI is rapidly accelerating its improvement rate Is it? All I see are desperate AI companies slapping on multi-modality because text generation is nearly peaked

[joshstrange]

> And, most of all, the author doesn't remember how nonobvious Stripe's utility was in its early days, even in the tech scene: there were established ways to take people's money and it wasn't clear why Stripe had an offering worth switching to.

I think you are misremembering. Stripe was a _big deal_. They had a curl call on their home page for a while for how to take a payment IIRC. It was like how Twilio opened the door for anyone to send SMS, Stripe made it stupid-easy to handle payments online. Nothing else at the time compared in terms of simplicity and clearly defined fees.

[riku_iki]

> In particular, GPT-2 to GPT-4 spans an increase from 'well read toddler' to 'average high school student'

GPT-2 was indeed much smaller and weaker model. But the question do we have "exponential" boost after GPT3, or just marginal while competition commoditized this vertical.

[ThereIsNoWorry]

That's a lot of unproven assumptions based on the fact that LLMs are just correlation printers.

[vrighter]

ai is very rapidly hitting a plateau, both because they are just glorified markov chains, and them running out of data anyway

[ggm]

I think you don't understand the meaning of the word exponential. By now, at the intent expressed here, we'd all be in the Kurzeweil singularity if this was correct.

Hint: it's not correct. It's nothing like exponential. It's not even order of magnitude stuff. It's tiny increments, to a system which fundamentally is a bit of a dead end.

[lostmsu]

I love how pointless these talks are.

What does it even mean for a value whose scale is not defined to be logarithmic, quadratic, or exponential?

You guys are threading water about nothing.