What is known about the capabilities of GPT-5
[ad_1]
GPT-5 could be released in the summer of 2024, two sources close to OpenAI told Business Insider. The company itself did not announce the release date of the next version of generative artificial intelligence. But its CEO Sam Altman confirmed that the developer will release an “amazing new model” this year. Some clients have already received access to the demo version of GPT-5, the publication writes. According to one test participant, it is “significantly better” than the previous one.
As Evgeny Chereshnev, vice president for strategy and innovation at MTS, CEO of Future Crew, the MTS innovation center, told Kommersant FM, GPT-5 will have fewer errors and improved protection. An experienced engineer can still get GPT-4 to find keys to crypto wallets or ask them to create hacking software. And from a business point of view, the potential of GPT-5 will be completely incomparable with what we have now, the expert noted:
“ChatGPT will learn to handle very application-specific scenarios better than you would expect from any generic model. For example, it is possible that a medical version will appear, when communicating with which, in fact, you can get a session with a therapist, whose role will be played by a machine. You can meet with a ChatGPT lawyer who will advise you on any relevant service for a small fee.
A huge leap is expected in the Copilot software development area. That is, the programmer simply gives a task, and ChatGPT writes the code, optimizes it, corrects errors, and so on. The car can do this even now, but this Copilot will be much better.
I think it’s safe to say that if OpenAI decides to go public in the near future, it will be one of the companies with a potential $1 trillion valuation.”
GPT-5 and its predecessors learn from huge data sets, allowing neural networks to solve problems more efficiently. The technical parameters of the new model are not yet known, but the amount of training data will significantly exceed that of previous versions.
The GPT-4 model was released in 2023. According to the developers, it provided very accurate and prompt answers, but over time it developed problems such as degradation and laziness, that is, artificial intelligence simply refused to respond to requests. And GPT-5 will be able to perform tasks in autonomous mode, explained Sergey Zubarev, CEO of the Sistema company: “The new version is planned as a so-called agent, and not just artificial intelligence that communicates with the user in the form of a chatbot. That is, this is a multi-stage AI that will be able to perform more complex tasks.
As for the timing of the presentation, I think this is also due to the fact that the release of LlaMA 3 is planned for the summer. This is probably the most popular open source model, sponsored by Mark Zuckerberg.
For this purpose, he purchased the most powerful video cards at that time for about $1 billion. It looks like in terms of power, volume, and capabilities, this is very close to artificial intelligence, which is equal to human intelligence.
If we talk about the prospects of OpenAI, then GPT-4 has already become a plus. Last year, the company received, I think, more than $1 billion in net revenue.”
In February 2023, according to OpenAI, the number of GPT-4-based chatbot users reached 100 million people. It is known that it was also implemented by Morgan Stanley, the Icelandic government and the educational platform Duolingo. And in the Amazon e-book store there were about 200 books written by ChatGPT. According to the general director of the SMM agency Soyka, Ekaterina Kozyreva, her company introduced GPT-4 into its work a couple of months after its introduction, and is now waiting for the release of a new version:
“It’s super low-cost, the paid version of ChatGPT costs us $10 or $20 a month. It happens that a chatbot works with a very large amount of data, for example, we load into it the results of problem interviews, say, 60 pieces, and say: “Dear neural network, now tell me what they have in common.” Sometimes in such cases ChatGPT starts to glitch or give some very obvious answers; specifics are already more difficult for it. So we are certainly waiting for the fifth version, we’ll see what happens there. But I’m sure it will be useful for the team.
Sometimes it’s much easier to work with a neural network as a text author than with a human, especially when the team is in different time zones.”
Previously, Nvidia introduced the most powerful AI chip of the new generation – Blackwell B200. According to the company, it is four times faster than their development last year. For example, if training an AI model at the GPT-4 level previously required 8 thousand such chips, now 2 thousand are required. The new processor will reportedly be used by Google, Microsoft, OpenAI and Tesla. This will increase the capacity of their servers hundreds of times.
Everything is clear with us – Telegram channel “Kommersant FM”.
[ad_2]
Source link
