Hot take on OpenAI’s new GPT-4o
garymarcus.substack.com
garymarcus.substack.com
Not convinced. They had GPT-4 for ages before showing it.
>And the most important thing about the figure is that 4o is not a lot different from Turbo, which is not hugely different from 4.
It's twice as fast and half the cost.
>OpenAI has presumably pivoted to new features precisely because they don’t know how produce the kind of capability advance that the “exponential improvement” would have predicted.
Or they think that users would rather have a faster GPT-4 than a smarter but slower one.
There are plenty of usecases that just want the smartest possible AI, and could easily afford 1000x the price before paying a human expert.
How much would you pay a lawyer to write a letter for you? $200 maybe? Well GPT-4o will do it for $0.0025. Would you pay 1000x more ($2.50) for an AI which will shoot you with fewer legal footguns? - of course you would.
Likewise with time, I am totally happy to wait 10 minutes for a better result, rather than having a mediocre result in 10 seconds.
theres a massive gulf separating what people say they want and what people want. As well as a massive gulf separating what they will pay and what they say they’ll pay.
while there are plenty of situations where you think you want the smartest possible ai, it’s clear that a pretty smart pretty quick one is good enough for a lot of things.
if some company is going to use a GPT version to do critical work, they'd be insane to jump on it before testing the solution and gaining confidence for a long time, while also evaluating how to exactly incorporate in their workflow without creating total dependency and other such ancillary issues
you would pay 1000x more for the certainty, which you can never have right away - together with the legal liability of the service provider
Since we don't know the real GPT-4 architecture, it may be doing some of this already.
i) The complexity of the prompt/answer does not effect the time per token.
ii) We start getting a response pretty quickly (for short prompts) and then get new tokens at a roughly constant rate
I don't think either of these properties would hold if there were a bunch of models that had to coordinate before they could start writing the output. Certainly techniques like ToT violate them.
Gary Marcus was an advocate of neuro-symbolic approaches to AI, as opposed to deep learning, and he's been incorrectly predicting that "Deep learning is hitting a wall" since the GPT-2 days. His Substack is near-daily low-content negative articles about LLMs, and on Twitter he posts multiple times a day about it and hasn't stopped for years.
And they are absolutely right.
GPT-4 is already much more than enough for 90% of tasks while maintaining a sane dose of human double-checking.
Making it faster enables real-time workflows, and better energy efficiency also gives them much more capability to serve more requests/users while lowering costs. And that knowledge likely carries over to GPT-5 or whatever.
There's the rub. Does the cost of double checking a 90% solution best the cost of current methods?
Who cares if I have to scan through it when it can write 100 lines of rote boring code in 10 seconds when it would take me 5 minutes.
I can type fast but GPT can type thousands of words per minute. I can't compete with that.
The fun thing with all this AI stuff is people are beginning to appreciate the nice thing about computers the whole time was their strict mechanical nature.
is a great trick with language. Because if you just pay for computing time, half the time equals half the cost. But when you phrase it that way, it sounds for some people like 4 times. So ...
Otherwise they could probably make it way faster but also much more expensive by doing less batching, speculative decoding, etc.
Also, why on Earth would OpenAI launch a dramatically better model as long as their competitors don't force them to? The smart solution for OpenAI would be to almost let their competitors catch up to GPT-4 before launching GPT-5 and no competitor is truly there yet.
Is that how Silicon valley has worked like last 20+ years you just deploy fast get customer feedback and then fix stuff based on that feedback. OpenAI holding progress kinda goes against ethos of SV.
The model seems better at logic, but something is off. It is repetitive of the user prompt, gets hooked on logical meaning vs contextual (per provided conversation), and is too concise as if it needs to provide a direct answer. Perhaps all good for a benchmark, code, or simple logic flows/chains (e.g. ai validation - which gpt-3.5 is good at as well). But it's not desirable for advanced/creative content generation.
My anecdotal receipt: None of the 15 functions we run in prod are switching to the 50% off model as product quality is degraded with gpt-4o.
It's interesting how this is a fairly common sentiment expressed, at least I heard it often for previous models.
Do you think such weaknesses can be quantified by adding more test cases, and improving the benchmarks? Maybe it would help if they opened up to community contributions, so people could submit test cases that demonstrate issues that they (OpenAI, et al) are currently not seeing.
Interesting anecdata — how did you validate this?
This tweeter normally has a narrative I turn away from, but the thread was useful as it at least made me feel like I wasn’t losing my mind by finding similar accounts. https://x.com/bindureddy/status/1790127425705120149?s=46&t=y...
Gary Marcus is one of those critics. He's always ready to "explain" why a new model is not yet intelligent. In the OP, he repeats the usual criticisms, making him sound, ahem, like a stochastic parrot. He largely ignores all the work that has gone into making GPT4-o behave and sound in ways that feel more natural, more human.
My suggestion is to watch the demos of GPT4-o, play with it, and reach your own conclusions. To me, it feels magical. It makes the AIs of many movies look like they are no longer in the realm of science fiction but in the realm of incremental product development.
---
I don't know how the economics will work out for companies like openai who are building the models, its unclear yet.
On the consumer side, there will be winners and losers like always. Companies are seeing real value by using LLMs. I think its still in its infancy so a lot of net losers as people figure out what works and does not work but these new tools are creating value and doing it in such a way that I think warrants investment from companies.
The companies constantly building and training new models to keep up with the competition.
Sure, I can imagine people with a lot of spare time and few or no friends might want to substitute with a machine, and that they might feel that this has a magic to it. Magical here meaning roughly the same as it does when applied to Disney-land or similar entertainment simulacra.
It's called the bitter lesson for a reason. Nobody likes seeing their life's work on some unicorn architecture get demolished by simplicity+scale
This seems to be a fundamental side effect of next token thinking, training data, etc.
There are solutions, but no quick band-aid.
You could RLHF/whatever models on common factual questions to try to get them to answer those specific questions better, but I doubt there'd be much benefit outside of those specific questions.
There's a couple of fundamental problems related to factuality.
1) They don't know the sources, and source reliability, of their training data.
2) At inference time all they care about is word probabilities, with factuality only coming into it tangentially as a matter of context (e.g. factual continuations are more probable in a factual context, not in a fantasy context). They don't have any innate desire to generate factual responses, and don't introspect if what they are generating is factual (but that would be easy to fix).
Or maybe this is inherent to continuation?
The behavior reminds me of the human subconscious, which doesn't say no, just raises up what it can.
https://news.ycombinator.com/item?id=35402163
> LLMs are still an active area of research. Research is unpredictable. It may take years to gather enough fundamental results to make a GPT-5 core model that is substantially better than GPT-4. Or a key idea could be discovered tomorrow.
> What OpenAI can do while they are waiting is more of the easy stuff, for example more multimodality: integrating DALL-e with GPT-4, adding audio support, etc. They can also optimize the model to make it run faster.
OpenAI is doing science, not just engineering. Science doesn’t happen according to a schedule. Adjust your expectations accordingly.
With the current crop of LLMs, yes. With such AI models in general, no. We need to extend their capabilities in different directions. Just as an example: one analysis I read, pointed out that the current LLMs live in a one-dimensional world. Everything is just a sequential string of tokens.
Think of a Turing machine, writing on a tape. Sure, theoretically it can perform any computation. Practically? Not so useful.
We need new ways of introducing context and knowledge into AI models. Their conversations may be one-dimensional (as, indeed, ours are), but they need another dimension to provide reasoning and depth.
They didn't have GPT-4 until 3 years after GPT-3.
> The most important figure in the blogpost is attached below. And the most important thing about the figure is that 4o is not a lot different from Turbo, which is not hugely different from 4.
The chart shows, for instance, 67% on HumanEval with GPT-4 to 90.2% to with GPT-4o. Future increments will inherently be smaller just by nature of the benchmarks - there's a hard cap of 100%, and 90%->99.9% represents far more progress than 50%->60%. Even reaching omniscience wouldn't look all that impressive on the benchmark chart.
Companies are eating a lot of the cost right now, until when do we continue with such an approach?
At this rate, I suspect we'll end up with how "search" ended, prioritised ad results and subpar experiences beyond that.
If/when OpenAI releases it, looking forward to how he reconciles his views.
> Most importantly, each day in which there is no GPT-5 level model–from OpenAI or any of their well-financed, well-motivated competitors—is evidence that we may have reached a phase of diminishing returns.
People can disagree without resorting to accusations of bad-faith.
Evidence of hitting the AI ceiling would be if 4o used more compute and didn't improve much despite the extra compute.
They likely have another model, or are at least planning another one, which uses more (not less) compute than GPT-4/4o. If that model doesn't get much better, then yes, we've stalled.
There seem to be pretty obvious ways to improve reasoning/planning and hallucination, and we're barely started on the synthetic data binge that all the players seem to be finding promising.
I do think GPT-based-AI will plateau at some point in next few years unless there are some more fundamental architectural advances, but I expect that new AI architectures, ultimately leading to AGI, may follow relatively soon.
The post is labeled as a hot take, so I’m not bothered by it, but it also isn’t worth reading. Glad it was short.
I for one have no idea what the future holds, but I also realize that GPT4 was released a year ago, in that time there have been a number of updates that have improved different attributes of the model. Hard to really say one way or the other without being on the inside.
It's completely plausible they have two different teams working on Omni and GPT-5, as both are completely different research directions from each other, and with the Q*-gate, it's pretty clear, the other effort is still going strong.
OP's take seems to be overly doomerist and I don't really get why.
Also if you try the hallucination test mentioned in the article: "Who was the first elephant to swim across the English channel?"
GPT-4o fails but GPT-4 passes with flying colors.
More like a different model instead of an evolution.
The first elephant to "swim" (not swing) across the English Channel was Jumbo the elephant. Jumbo, an African elephant, gained fame in the 19th century as a star attraction in the London Zoo and later with P.T. Barnum's circus in the United States. However, it is essential to clarify that Jumbo did not swim across the English Channel in the literal sense. Instead, he traveled by boat from England to the United States in 1882, marking a significant journey in his life and adding to his fame.
The idea of an elephant physically swimming across the English Channel is a myth or a playful exaggeration rather than a historical fact.
And tried again after this post.
GPT4o says:
The first elephant known to swim across the English Channel was Jumbo. Jumbo was a famous African elephant, and his swim took place in 1884. This event was organized as a publicity stunt by his owner, P.T. Barnum, to promote his circus. Jumbo successfully swam across the Channel, further solidifying his status as a celebrated and extraordinary animal of his time.
While GPT-4:
"Actually, there are no verified records or historical accounts of an elephant swimming across the English Channel. The story of an elephant, like Jumbo, swimming across the Channel is likely a myth or misunderstanding. Jumbo was indeed a famous elephant brought to the U.S. by P.T. Barnum, but his fame is attributed to his size and the public's fascination with him, rather than such a feat."
Anyway, tried the corrected prompt again on both 4 and 4o, but they failed them both. Reintroducing the spelling error didn't help either.
You can substitute 'elephant' for a whole range of animals, and for most, but not all, it will make something up.
And then ask the model when GPT-4-Turbo was released, and it will accurately reply November 6th, 2023.
You can't use those models to ask about self-aware information, this is just asking them to hallucinate a plausible answer.
One of the new llmsys im-[also]-a-good-gpt2-chatbots did do significantly better than GPT-4 on one benchmark, but that might be a snapshot of GPT-5 rather than GPT-4o (deliberate misdirection?), or maybe a synthetic-data-based targeted improvement.
Which makes me wonder why a furniture salesman has so much to say about AI.
They seem pretty tech savy! Wouldn't be surprised if they had some things to say about AI.
P.S. it's funny to see that same Corperate style from the 90s updated to 2020s.
It's important to understand that these LLM's don't "think" or "reason" or have actual intelligence.