Marcus Sharp’s Review of GPT-5 Urgently in Need of a New Paradigm- OpenAI Has No Advantage

Reimagining GPT-5 A Critical Dissection by Marcus Sharp Calls for a Revolutionary Approach as OpenAI Loses Ground

The news about GPT-5 has recently gained popularity again.

Starting from the initial rumors that OpenAI is secretly training GPT-5, to Sam Altman clarifying the situation, to discussions about how many H100 GPUs are needed to train GPT-5, to DeepMind’s CEO Suleyman confirming that OpenAI is indeed training GPT-5.

And now, a new round of speculation begins.

In between, Altman boldly predicts that GPT-10 will surpass the collective intelligence of humanity and appear before 2030, becoming a true AGI and so on.

• BTC Soars What Factors Are Driving the Rise
• Trezor Unleashes Its Sleuthing Skills as Crafty Phishing Campaign Hones in on Crypto Hardware Wallet Users
• Kraken Taps Fintech Dynamo Bivu Das as New UK Managing Director

Recently, OpenAI’s multimodal model named Gobi also challenges Google’s Gimini model, igniting fierce competition between the two tech giants.

For a while, the latest advancements in large language models have become the hottest topic in the industry.

To use an ancient poem to describe it, it’s like “half-hiding a pipa (Chinese musical instrument) in one’s arms” – quite appropriate. The only question is when will it truly come to fruition after all the anticipation and calls.

A timeline review

Today’s topic is directly related to GPT-5 and is based on an analysis by our old friend Gary Marcus.

The main point can be summarized in one sentence: The transition from GPT-4 to 5 is not just about scaling up the model, but a change in the entire AI paradigm. Based on this, OpenAI, the creator of GPT-4, may not necessarily be the first company to reach GPT-5.

In other words, when the paradigm needs to change, previous accumulations may not be very transferable.

Before delving into Marcus’ viewpoint, let’s briefly review what has been happening recently regarding the legendary GPT-5 and what the public opinion has been saying.

It started with a tweet from KarLianGuaithy, co-founder of OpenAI, stating that the H100 is a hot commodity that all the giants are pursuing, and everyone is curious about who has it and how many they have.

Then, a large wave of discussions followed regarding how many H100 GPUs each company needs for training.

That’s the general idea.

GPT-4 was possibly trained on around 10,000-25,000 A100 GPUs.

Meta had around 21,000 A100 GPUs.

Tesla had around 7,000 A100 GPUs.

Stability AI had around 5,000 A100 GPUs.

Falcon-40B was trained on 384 A100 GPUs.

About this matter, Musk also joined the discussion, and according to Musk, training GPT-5 might require 30,000 to 50,000 H100s.

Morgan Stanley made similar predictions before, although their estimated total number is slightly lower than what Musk stated, around 25,000 GPUs.

Of course, when GPT-5 is brought up for discussion, Sam Altman inevitably comes forward to refute it and clarify that OpenAI is not training GPT-5.

Some bold netizens speculate that the reason OpenAI denied it is very likely that they just changed the name of the next generation model and it’s not called GPT-5.

Anyway, according to Sam Altman, it’s because there’s a shortage of GPUs that many plans have been delayed. He even said that he doesn’t want too many people to use GPT-4.

The entire industry is so desperate for GPUs. According to statistics, all the tech giants combined need about 430,000 GPUs, which is about $15 billion.

But it’s a bit convoluted to infer GPT-5 based on GPU usage, so DeepMind’s founder Suleyman directly “hammered” it in an interview, saying that OpenAI is secretly training GPT-5, stop hiding.

Of course, in the full interview, Suleyman also talked about a lot of industry gossip, like why DeepMind fell behind in the competition with OpenAI, even though not much time has passed.

There are also a lot of insider information, like what happened when Google acquired them. But these are not really related to GPT-5, so interested friends can find out for themselves.

In conclusion, this is a discussion among industry giants about the latest progress of GPT-5, which leaves everyone with doubts.

After this, in another one-on-one session, Sam Altman said, “I think before 2030, AGI will appear, called GPT-10, surpassing the sum of all human intelligence.”

On one hand, making bold predictions, on the other hand, denying training GPT-5, it makes it difficult for others to truly know what OpenAI is doing.

In this session, Altman envisioned many scenes belonging to the future. For example, how he understands AGI, when AGI will appear, what OpenAI will do if AGI appears, and what humanity should do.

But in terms of actual progress, Altman has planned this way, “I told the employees in the company that our goal is to improve the performance of our prototype product by 10% every 12 months.”

“Setting this goal to 20% may be a bit too high.”

This is a specific arrangement. But the connection between 10%, 20%, and GPT-5 is not very clear.

The most valuable thing is the Gobi multimodal model from OpenAI.

The focus is on the fierce competition between Google and OpenAI and at which stage they are in.

Before talking about Gobi, let’s talk about GPT-vision. This generation of models is impressive. Take a sketch photo, send it to GPT, and the website will make it for you in minutes.

When it comes to coding, it goes without saying.

Once GPT-vision is completed, OpenAI may be able to launch a more powerful multimodal model, codenamed Gobi.

Unlike GPT-4, Gobi was built as a multimodal model from the beginning.

This has piqued the interest of onlookers – is Gobi the legendary GPT-5?

Of course, we still don’t know how far Gobi’s training has progressed, and there is no definite news.

And Suleyman firmly believes that Sam Altman recently said they haven’t trained GPT-5, but he may not be telling the truth.

Marcus’s perspective

To begin with, Marcus states that it is very likely that in the history of technology, no pre-release product (except perhaps the iPhone) has been showered with more expectations than GPT-5.

This is not only because consumers are excited about it, but also because a large number of businesses are planning to build their fortunes around it, and even some diplomatic policies are being formulated around GPT-5.

In addition, the release of GPT-5 may intensify the recently upgraded chip war.

Marcus says that some people have specifically requested a halt to the production of GPT-5, based on a projected scale model of expectations for its release.

Of course, there are also many people who are optimistic and imagine that GPT-5 could eliminate, or at least greatly reduce, many concerns people have about existing models, such as their unreliability, biases, and tendencies to spout authoritative nonsense.

But Marcus believes that he has never been sure whether simply building a larger model can truly solve these problems.

Today, foreign media have reported that OpenAI’s another project, Arrakis, aimed at creating smaller and more efficient models, has been canceled by top management due to not meeting the expected goals.

Marcus says that almost everyone believes that GPT-5 will be released soon after GPT-4, and the imagined GPT-5 is much more powerful than GPT-4, so it was surprising when Sam denied its existence initially.

There are many speculations about this, such as the mentioned GPU issue, that OpenAI may not have enough cash on hand to train these models (the training costs of these models are notoriously high).

But then again, OpenAI’s financial position is almost as good as any startup company. For a company that just raised $10 billion in funding, it is not impossible to spend $500 million on training.

Another argument is that OpenAI realizes that the costs, both for training and running the models, will be very high, and they are not sure if they can make a profit at these costs.

This seems to make some sense.

The third view, which is also Marcus’s view, is that OpenAI has already conducted some concept verification tests in May of the first half of the year during Altman’s speech, but they were not satisfied with the results they obtained.

Their conclusion may be as follows: if GPT-5 is just an enlarged version of GPT-4, then it will not meet expectations and will fall far short of the predetermined goals.

If the results are disappointing and even laughable, then training GPT-5 is not worth spending hundreds of millions of dollars on.

In fact, LeCun has the same idea.

From GPT-4 to GPT-5, it is not as simple as 4plus. The shift from 4 to 5 should be groundbreaking.

What is needed here is a new paradigm, not just a larger model.

So, in terms of paradigm shift, it is certainly the more financially capable companies that are more likely to achieve this goal. But the difference is, it may not necessarily be OpenAI. Because a paradigm shift is a new track, past experience or accumulation may not be of much use.

Similarly, from an economic perspective, if what Marcus said is true, then the development of GPT-5 is essentially delayed indefinitely. No one knows when the new technology will arrive.

It’s like how electric cars nowadays generally have a range of several hundred kilometers, and if you want a range of over a thousand kilometers, you need completely new battery technology. And who will break through with the new technology often depends not only on experience and funding, but also on a little bit of luck and opportunity.

But no matter what, if Marcus is right, the various commercial values related to GPT-5 in the future will inevitably shrink significantly.

We will continue to update Blocking; if you have any questions or suggestions, please contact us!