Hacker News Reader

Comment by pk-protect-ai | Hacker News Reader

They are not leading their competition in terms of technology; they are ahead in a public theater and as the first to market. The claims of "regulate us," "it's almost AGI and it's dangerous," and "let's put a moratorium on larger models" are all just hype, deceiving the public. They utilize the work of scientists and other companies worldwide without acknowledging their contributions. Certainly, they have done an impressive job, and the scientists and engineers working there deserve respect and recognition. However, they are not significantly ahead of current research. They are not even weeks ahead of research. The current landscape is a game of budgets, winning public opinion, and securing investments.

View on HN

When you put it that way it kind of reminds me of FTX, they too were all about leading the way on regulation. IIRC the impetus for the entire thing coming crashing down was because the other big exchange's owner thought it was extremely hypocritical how they talked about inviting regulation while operating just like most other crypto firms, completely outside of decent financial controls.

Of course not to imply that OpenAI is the same as FTX.

Sam is more trustworthy than Sam.

Yeah I dont see what you’re seeing. Ignoring research, OpenAI is bringing tools to market that nobody else to quite close to yet. Yes, they are getting closer but not as you describe it.

Actually, I felt pk-protect-ai's post was spot on. That is, OpenAI does have, currently, the most impressive interfaces and models. But (as was famously commented on by a Google exec) they don't really have any unique moat. That is, while what OpenAI has built may be extraordinary, there is no "secret sauce" there that prevents others from copying them. And this isn't a knock on OpenAI at all. On the contrary, my understanding is that OpenAI was really the first to take the big risk in scaling their GPT models (i.e. spending the hundreds of millions to train their models before they knew what the outcome would be).

> was famously commented on by a Google exec

It was just a rank and file IC.

People keep saying this but after 1-2 years, nobody has gotten to their level yet.

This isn't true. Benchmarks do not define the usefulness of the models. Mixtral is much more useful for me right now than GPT-4. Look at LLaVA, or the new one, and the very impressive LWM (the text-only version is LLaMA2-based with a 512K context, which does not require TPU to run the inference). The fine-tuned LLaMA 34B is much faster than GPT-4 Turbo and less annoying, providing pretty impressive quality of results.

Who mentioned anything about benchmarks? YMMV, it depends on exact use cases, from my extensive testing using my cases, GPT-4 Turbo still is much easier to direct than any of the others you have mentioned.

I would not say that OpenAI is all that far ahead of the competition. I would not count Google out, and I would be willing to bet that OpenAI/Microsoft + Google will be the duopoly of the AI age.

At this point Gemini is as good as GPT4 (and anecdotally I think it's better at many coding assistance tasks, based on my experiments on the LMSys chatbot arena). Sora is getting a lot of press for text-to-video, but Google's Lumiere has already been out for a few weeks and produces pretty good results.

I have no doubt that OpenAI has been cooking things up. GPT-4 is an older model which others are only just catching up to now. They have an A+ team, a giant war chest, and a lot of momentum. But just because they have momentum does not mean that they have a moat.

I'd be willing to wager that the top-of-the-line foundational models are going to converge and become indistinguishable for almost all tasks. Even the open foundation models (e.g. Mixtral) are getting really good. Foundation models are not moats.

The players that have moats are Microsoft (with deep enterprise software & B2B expertise, allowing them to sell AI-powered software & ward off competition from upstarts) and Google (where their decade of investment into custom silicon allows them to train & run inference for cheaper than anyone else, by far).

I don't have a large amount of time to devote to this. But no, google is not out. They just choke slammed ChatGPT with the Gemini announcement.

10m context with that retrieval rate is such a monstrous leap. And to top it off, we got LargeWorldModel in the same week, capable of 1M token context with insane retrieval rate in the open source space. So not only is the open source world currently technically ahead of ChatGPT, so is Google. Which is why they had to announce SORA, because google's model is so far ahead of the competition. That's also why it will probably be ages before we get access to SORA. Now don't get me wrong, the average person can't afford 32 TPU's to run LWM, but we already have quants for it, which is a step towards enabling the average person (that somehow has 24-48gb of VRAM to get a taste of that power).

What is also striking is the fact that the new models are all multimodal as a standard. We not only leapfrogged in context size, but also in modalities. The model seems to only benefit from having more modalities to work with.

I think the statement Bill Gates made claiming that "LLM's have reached a plateau" itself indicates they don't believe they can make more money from training better/larger models. Which indicates that they already did as well as they could with their existing people, and are now "years" behind google. I never thought google could catch up, especially after their infamous "We have no moat" situation. But it seems they actually doubled down and did something about it.

To a lot of people, last Thursday was a very nihilistic day for Local Models, as the goalposts shifted from 128-200k context to 10M tokens with near perfect retrieval. It's literally insanely scary. But luckily we got LWM, and that means we have only been 10xed.

Now the local people will work on figuring out how to bridge the gap, before being leapfrogged again. What is really insane is that, we have had LLAMA2 for over a year now, and nobody else figured out how to get this result from it, despite it being around so long.

I still believe there are modifications to the architecture of MoE that will unlock new powers that we haven't even dreamed of yet.

Sorry, this was supposed to be well thought out, but it turned more into stream of consciousness, and I honestly had no intention of disagreeing with you.

> But luckily we got LWM, and that means we have only been 10xed.

If I remember the paper correctly, it was something about a 4M context in there. So not 10x, but 2.5x.

> What is really insane is that, we have had LLAMA2 for over a year now, and nobody else figured out how to get this result from it, despite it being around so long.

This isn't true. For now, the task of extending context to 10M tokens is brute-forced by money (increased HW requirements for training and inference and increased training time are also a financial domain). And for now, there simply is no leapfrogging solution for open source or commercial models, which will decrease the costs by orders of magnitude.

Calling it public theater is pretty laughable. The amount you get for $25 a month as a consumer is insane.

I literally have a 24/7 consultant with surface level understand of any topic in human history. This consultant also happens to be an amazing artist for $25 a month that gives me the rights to commercialize any art piece they make for me.

This is extremely hard to beat. This will be extremely hard to beat.

As a business they are succeeding.

You get a "consultant" whose every word has to be verified against an independent second source, and an """artist""" containing a digest of essentially everything actual human artists have ever uploaded to the web, who's somehow still only able to generate derivative slop utterly undifferentiated from what it's spitting out for everybody else.

LLMs are useful tools, but if they don't lead to something you can actually somewhat rely on to generate correct/consistent/factual results (see e.g. the recent Air Canada chatbot lawsuit) then the hype is a bust. TBD.

> You get a "consultant" whose every word has to be verified against an independent second source So, no different from YouTube Videos or [insert any website here]? That’s still a ton of value to capture and repackage. Especially when it’s packaged into an interface that humans find much more natural (ie. “conversation”) than the existing offerings.

Hence why I said "LLMs are useful tools". But the current hype pitches them as much, much more than ancillaries to traditional web search, hence why I said "[...] then the hype is a bust. TBD."

I just don't believe that LLM's have to reach the impossible bar you set of near-perfect correctness in order to be world changing. Thus, I disagree that any outcome less than that should be considered a "bust".

I'm not sure what else you can call a scenario in which LLMs, currently hyped as (among other things) near-term replacements for human workers in all sorts of professions and positions, are not eventually able to satisfactorily (and cost-effectively) replace even the most low-skilled, largely script-driven customer service workers.

The bar I'm setting is far from "impossible"; even human children generally won't seamlessly confabulate when you ask them a question they don't know the answer to. Again citing the recent Air Canada case, these models can't even reliably answer simple questions that are definitively and objectively answered in documentation that is presumably made as freely available to them as is technically possible under the limitations of current technology.

>LLMs, currently hyped as (among other things) near-term replacements for human workers in all sorts of professions and positions

No well-informed sources are saying that. If you're saying that mainstream reporting and other non-tech folks are wrong about what the possibilities are then... obviously.

> No well-informed sources are saying that.

Yes they are, actually! I've talked to people who obsessively read practically every LLM paper that passes through the arxiv, with undeniably deep and broad knowledge of the current state of this tech, who seriously believe it's going to surpass humans within a year or two. That it may already have, in the deep dark top secret labs beneath OpenAI HQ.

However,

> If you're saying that mainstream reporting and other non-tech folks are wrong about what the possibilities are then... obviously.

If it was obvious then why are you still replying to my comments, which have very obviously been specifically addressing the mismatch between hype and reality? If the current approach doesn't scale, the hype will have been a bust! Objectively! That's what I've been talking about this whole time!

"Mainstream reporting and other non-tech folks" is a bit disingenuous, though. The primary drivers of the current unrealistic hype are software vendors and associated clingers-on looking to make a quick buck. They'll say anything, regardless of whether it's true, and as a result of those mostly-falsehoods our public lives will be flooded with awful AI tools that make everything shittier and more difficult. I can't wait!

No company in the history of companies has ever advertised themselves with the concept "we're going to destroy the world", primarily because it doesn't actually work. OpenAI are true believers, they say that sort of thing because that is what they genuinely think.

No single neural network currently has the capability to destroy the world, and it will not occur until the architecture for NNs is altered to become self-driving rather than reactive. Even then, it will not be sentient. The deception occurs when the attention shifts from the people operating new powerful technology (tools) to this technology itself.

Meanwhile, every AI company ever: "We are very excited about agents. We are working hard on agents. We want to roll out agents as soon as possible. Also, persistent memory and online learning would be nice to get."

Like, are you following the things that OpenAI and Deepmind are saying at all? The things that make current LLMs not a threat, they aim to tear down as soon as they can arrange.

OpenAI just released a video network, and one of their core touted benefits was that you could use it as an action controller!

And, um. Do you really think, when the AIs can take a simple prompt and turn it into a ten thousand step plan that requires dynamic skill acquisition, resourcing and persistence, that generating the prompt will be the one single task that stumps them? When we are at that point - and to be clear, every leading AI organisation is sprinting to reach that point earliest - then the difference between doom and safety will be one sentence: "When you are done with that, generate a new prompt." This is not how a world with a long expected lifespan looks.

edit: To be clear, I'm still not accusing OpenAI of making up the doom stuff. Even though when I phrase it like that it sounds like they're directly working on things that obviously end the world, which seems contradictory, I don't think they see it like that. To be honest, I can't explain why any doomer works at OpenAI, except in the way that people sometimes move towards explosions and gunfire. I think it's just a bug in the human brain. We want to have the danger in sight.

The agents and action controllers are the reactive components in the toolchain, aren't they?

"Self-driving" is just "reactive" with an open recursive structure. In principle, a network that processes a prompt, generates a plan, recurses a finite number of times, judges how well it did, generates a training plan to improve, outputs a corresponding follow-up prompt, and then waits for you to press a button before it repeats the whole thing with the follow-up prompt, ad infinitum, is still "reactive" - but nobody would argue that whoever presses the button is performing irreplaceable cognitive labor.

So I just don't think this captures an important distinction at the limit. If a system can generate a good action plan, turning it into an agent is just plumbing.

Actually, we can't be certain that humans themselves are not reactive. It is just that their reactions are either built in (self-preservation, reproduction), or based on much broader input (sensory, biochemistry, etc.). The current level of reactivity of the LLMs is very limited by their architectures, though, and as long as these architectures stay in place, you can't expect them to be "self-driven".

I just don't think this is the case. I suspect reactivity in LLMs is mostly limited by training. Human text data is just not suited to the way an AI needs to output data to plan long action chains - justification, not reasoning.

This might be true as well along with what I said.

Reply on news.ycombinator.com