Shulou(Shulou.com)11/24 Report--

Unexpectedly, the short-term work of OpenAI can not be promoted, it is the lack of GPU!

After the hearing, Sam Altman and his team began a "trip" in Europe.

In a recent interview, Altman did not hide it, revealing that all the progress of his AI would have to wait for GPU to keep up.

He discussed OpenAI's API, as well as product plans, which attracted a lot of attention.

Many netizens have said that I like Altman's honesty.

It is worth mentioning that the multimodal capability of GPT-4 should be available to most Plus users in 2024, as long as there is enough GPU.

The supercomputing that Microsoft Big Brother spent $1.2 billion on OpenAI is far from enough to meet the computing power needed for GPT-4 to run. After all, the GPT-4 parameter is said to be 100 trillion.

In addition, Altman also revealed that GPT-3 is also in the open source plan of OpenAI.

I don't know if the content of this interview revealed too many "secrets" of OpenAI. The source manuscript has been deleted. Hurry up.

The latest interview, hosted by Raza Habib, CEO of AI development platform Humanloop, interviewed Altman and 20 other developers.

This discussion covers actual developer issues, as well as larger issues related to the mission of OpenAI and the social impact of AI.

Here are the key points:

1. OpenAI is in urgent need of GPU

2. OpenAI's near-term roadmap: GPT-4 Multimodal 2024 opening

3. Access to the ChatGPT plug-in through API will not be released in the near future

4. OpenAI only does the "killer application" of ChatGPT, and the goal is to make ChatGPT a super intelligent work assistant.

5. GPT-3 is in the open source project

6. The scaling rule of model performance continues to be valid.

Next, we will introduce exactly what Sam Altman said from 6 big points.

OpenAI currently relies heavily on GPU to interview all the topics around, "OpenAI is too short of GPU."

This has delayed many of their short-term plans.

At present, many customers of OpenAI are complaining about the reliability and speed of API. Sam Altman explained that the main reason is that GPU is too lacking.

OpenAI is the first customer of Nvidia DGX-1. As for the context length that supports 32k tokens, it is not available to more people.

Since OpenAI has not yet overcome the technical hurdles, it looks as if they will have a context to support 100k-1M tokens this year, but they need to make a breakthrough in research.

Fine-Tuning API is also currently limited by the availability of GPU.

OpenAI does not yet use efficient fine-tuning methods like Adapters or LoRa, so fine-tuning is very computationally intensive for operation and management.

However, they will provide better support for fine-tuning in the future. Even, OpenAI may host a community-contributed model market.

Finally, the availability of dedicated capacity is also limited by GPU availability.

Earlier this year, netizens revealed that OpenAI is quietly launching a new developer platform, Foundry, which allows customers to run the company's new machine learning model on dedicated capacity.

This product is "designed for cutting-edge customers running larger workloads". To use this service, the customer must be willing to pay a fee of $100k in advance.

However, from the disclosed picture information, we can see that the example is not cheap.

Running a lightweight version of GPT-3.5, a three-month commission will cost $78000 and a year will cost $264000.

On the other hand, it can also be seen that GPU consumption is expensive.

OpenAI roadmap Altman shares OpenAI API's tentative roadmap for the near future:

2023: fast and cheap GPT-4 customers this is the top priority of OpenAI.

Overall, the goal of OpenAI is to reduce the "cost of intelligence" as much as possible. So they will try to continue to reduce the cost of API.

Longer context windowsThe context window may support up to 1 million tokens in the near future.

Fine-tuning API-- fine-tuning API will be extended to the latest model, but the exact form will depend on what the developer really wants.

Most of the token with memory is currently wasted in the above transmission, and there will be a version of API that can remember the history of the conversation in the future.

2024: the powerful multimodal capability was demonstrated when the multimodal capabilities were released, but this capability cannot be extended to everyone until GPU is satisfied.

Plug-in "does not have PMF" and will not appear in API in the short term

Many developers are interested in accessing ChatGPT plug-ins through API, but Sam says these plug-ins will not be released soon.

"apart from Browsing, the plug-in system has not found PMF yet. "

He also pointed out that many people want to put their products in ChatGPT, but what they really need is to put ChatGPT in their products.

Except that ChatGPT,OpenAI will not release any more products, every move of OpenAI makes developers tremble.

Many developers say that when OpenAI may release competing products, they are nervous about using OpenAI API to build applications.

Altman said OpenAI will not release any more products outside of ChatGPT.

In his view, great companies have a "killer app", and ChatGPT is going to do this record-breaking app.

ChatGPT's vision is to be a super intelligent work assistant. OpenAI won't touch on many other GPT use cases.

Regulation is necessary, but open source is also necessary. Although Altman calls for regulation of future models, he does not think that existing models are dangerous.

He believes that it would be a huge mistake to regulate or ban existing models.

In the interview, he reiterated his belief in the importance of open source and said that OpenAI is considering opening up GPT-3.

Now, part of the reason OpenAI is not open source is that he is skeptical about how many individuals and companies are capable of hosting and providing large models.

The model performance "scaling rule" is still valid recently, and many articles claim that the era of giant artificial intelligence models is over. However, it does not accurately reflect the original intention of Altman.

OpenAI's internal data show that the scaling rule for model performance (scaling laws) is still valid, and making the model larger will continue to produce performance.

However, the speed of scaling cannot be maintained because OpenAI has expanded the model millions of times in just a few years, and doing so will not be sustainable in the future.

This does not mean that OpenAI will not continue to try to make models larger, but that they may only double / 2x per year, rather than multiple orders of magnitude. The fact that the scaling rule continues to work has an important impact on the timeline of AGI development.

The scaling assumption is that we may already have most of the parts needed to build AGI, and most of the remaining work will be to extend the existing methods to larger models and larger data sets.

If the era of scaling is over, then we should probably expect AGI to go further. The continued effectiveness of the scaling rule strongly implies that the timeline of implementing AGI will be shorter.

Some netizens commented on the hot comments of netizens.

OpenAI: laws and regulations must be passed to protect our moat. Or OpenAI:Meta pees in our moat, let's also hint that our model should be open source.

Others say that if GPT-3 is really open source, like LLMa, it will take about five days to run on the M1 chip.

Community developers can help OpenAI solve GPU bottlenecks as long as they open source models. In a few days, developers will be able to run it on CPU and edge devices.

GPU shortage, some people think that there is something wrong with the OpenAI capital chain, can not afford to buy.

However, others say there is an obvious lack of supply. Unless there is a revolution in chip manufacturing, there may always be a shortage of supply relative to the consumption of GPU.

Some netizens doubt that Nvidia is still undervalued. The step-by-step change in computing demand may last for several years.

Nvidia has just joined the trillion-dollar club, so the unlimited demand for computing power may lead to a chip factory in the world that costs more than $2 trillion.

Reference:

Https://humanloop.com/blog/openai-plans

Https://news.ycombinator.com/item?id=36141544

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.