GPT-3 plans to be open source! Sam Altman revealed that there is an urgent need for GPUs, and GPT-4 multi-modal capabilities will be available next year

After the hearing, Sam Altman took the team on a "travel" in Europe.

In a recent interview, Altman did not hide it, revealing that all progress in his own AI will have to wait until the GPU catches up.

He discussed OpenAI’s API and product plans, which attracted the attention of many people.

Many netizens have said that I like Altman’s candor.

It is worth mentioning that the multi-modal capabilities of GPT-4 should be available to most Plus users in 2024, provided they have enough GPUs.

The supercomputer Microsoft’s big brother spent US$1.2 billion to build for OpenAI is far from being able to meet the computing power required for GPT-4 runtime. After all, GPT-4 parameters are said to have 100 trillion.

In addition, Altman also revealed that GPT-3 is also included in OpenAI’s open source plan.

I don’t know if this interview revealed too many “secrets” of OpenAI. The source manuscript has been deleted, so hurry up and code it.

Key points

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

This discussion touches on practical developer issues, as well as larger issues related to OpenAI’s mission and the social impact of AI.

The following are the key points:

1. OpenAI is in urgent need of GPU

2. OpenAI’s recent roadmap: GPT-4 multi-modal will be opened in 2024

3. The ChatGPT plug-in accessed through API will not be released in the near future

4. OpenAI only makes ChatGPT, the "killer application", with the goal of making ChatGPT a super-smart work assistant

##5. GPT-3 In the open source plan

6. The scaling law of model performance continues to be effective

Next, I will introduce what Sam Altman said from 6 major points What happened?

All the topics in the interview revolved around, "OpenAI is too short of GPU."

This has delayed many of their short-term plans.

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

##OpenAI is the first customer of NVIDIA DGX-1 supercomputer ## As for , the context length supporting 32k tokens cannot yet be rolled out to more people.

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

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

OpenAI does not yet use efficient fine-tuning methods like Adapters or LoRa, so fine-tuning is very computationally intensive to run and manage.

However, they will provide better support for fine-tuning in the future. OpenAI may even host a marketplace of community-contributed models.

Finally, dedicated capacity provision is also limited by GPU availability.

At the beginning of this year, netizens revealed that OpenAI was quietly launching a new developer platform, Foundry, to allow 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, customers must be willing to pay $100k upfront.

However, as can be seen from the picture information disclosed, the example is not cheap.

Running the lightweight version of GPT-3.5 will cost $78,000 for 3 months of delegation and $264,000 for a year.

It can also be seen from the other side that GPU consumption is expensive.

OpenAI near-term roadmap

Altman shared the tentative near-term roadmap for the OpenAI API:

2023 ：

· Fast and cheap GPT-4 ーー This is a top priority for OpenAI.

In general, OpenAI’s goal is to reduce “smart costs” as much as possible. So they will work hard to continue driving down the cost of their APIs.

·Longer context windows—In the near future, context windows may support up to 1 million tokens.

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

·Memory API ーーCurrently most tokens are wasted in the above transmission. In the future, there will be an API version that can remember the conversation history.

2024：

##·Multi-modal capabilitiesーーGPT-4 demonstrated powerful multi-modal capabilities when it was released, but until GPUs are satisfied, this feature cannot be extended to everyone.

Many developers are very interested in accessing the ChatGPT plug-in through the API, but Sam said that these plug-ins It won't be released anytime soon.

"Except for Browsing, the plug-in system has not found PMF."

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

Every move OpenAI makes makes developers tremble.

Many developers said they were nervous about using the OpenAI API to build applications when OpenAI might release products that compete with them.

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

In his opinion, a great company has a "killer application", and ChatGPT is going to be this record-breaking application.

ChatGPT’s vision is to become a super smart work assistant. OpenAI will not touch many other GPT use cases.

While Altman calls for regulation of future models, he does not believe that existing models are dangerous.

He believes that regulating or banning existing models would be a huge mistake.

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

Now, part of the reason OpenAI is not open source is because he doubts how many individuals and companies have the ability to host and provide large models.

Recently, many articles have claimed that the era of giant artificial intelligence models is over. However, it does not accurately reflect Altman's original intention.

#OpenAI’s internal data shows that scaling laws for model performance are still in effect, and making the model larger will continue to yield performance.

However, OpenAI has already scaled up its model millions of times in just a few years and therefore cannot sustain this rate of expansion.

This doesn't mean that OpenAI won't continue to try to make models bigger, it just means that they will probably only increase by 1x/2x per year instead of multiple orders of magnitude. The fact that scaling laws continue to be in effect has important implications for the timeline of AGI development.

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

If the era of scaling is over, then we should probably expect AGI to be even further away. The continued validity of the scaling law strongly implies that the timeline for implementing AGI will become shorter.

Hot comments from netizens

Some netizens joked,

##OpenAI: Must Protect our moats through regulations. OpenAI once again mentioned that Meta is peeing in our moat, which should also mean that our models need to be open source.

Some people say that if GPT-3 is really open source, like LLMa, it will take about 5 days to be ready on the M1 chip It's running.

Community developers can help OpenAI solve the GPU bottleneck, provided they open source their models. Within days, developers will be able to run it on CPUs and edge devices.

Regarding the GPU shortage, some people think that there is a problem with OpenAI’s capital chain and it cannot afford it.

However, others said there was an obvious lack of supply. Unless there is a revolution in chip manufacturing, there will likely always be an undersupply relative to consumer GPUs.

Some netizens doubt that Nvidia’s value is still underestimated? The step change in computing demand may last for years...

Nvidia just joined the trillion dollar club, so unlimited computing power Demand may lead to a world with chip factories exceeding $2 trillion.

Reference:

##//m.sbmmt.com/link /c55d22f5c88cc6f04c0bb2e0025dd70b

#//m.sbmmt.com/link/d5776aeecb3c45ab15adce6f5cb355f3

The above is the detailed content of GPT-3 plans to be open source! Sam Altman revealed that there is an urgent need for GPUs, and GPT-4 multi-modal capabilities will be available next year. For more information, please follow other related articles on the PHP Chinese website!