Two rumors and latest predictions for GPT-4!

​Author|Yun Zhao

On March 9, Microsoft Germany CTO Andreas Braun brought a long-awaited news at the AI ​​kickoff conference: "We will When GPT-4 launches next week, we will launch a multi-modal mode that offers completely different possibilities - such as video."

##In words, he likened large language models (LLMs) to "game changers" because they teach machines to understand natural language, and the machines then statistically understand things that could previously only be read and understood by humans. At the same time, the technology has evolved to be "available for all languages": you can ask questions in German and answer in Italian. With multimodality, Microsoft (-OpenAI) will "make models comprehensive."

It seems that it is basically certain that GPT-4 will be a large multi-modal model. But if you want to know the true identity of GPT-4 in advance, it is not impossible to find it.

1. Prediction: GPT-4 is based on the previous multi-modal large language model Kosmos-1

In fact, as early as March 1, we noticed that Microsoft published The multi-modal large language model Kosmos-1 is developed. If GPT-4 is multi-modal, we can reasonably predict what GPT-4 is capable of.

KOSMOS-1 is a multimodal large language model (MLLM) capable of perceiving multimodal inputs, Follow instructions and perform contextual learning tasks, which can also be multi-modal tasks. In this work, we align vision with large language models (LLMs), driving the evolution from LLMs to MLLMs.

## Jim Fan, a Stanford Ph.D. and Nvidia AI scientist, made 5 specific predictions based on this:

(1) Visual IQ test: Yes, a test for humans! (2) No OCR reading comprehension: input screenshots, scanned documents, street signs or any pixels containing text. Reason directly about content without explicit OCR. This is useful for unlocking AI-driven applications on multimedia web pages or "text in the wild" from real-world cameras. (3) Multi-modal chat: Have a conversation about pictures. You can even provide a "follow-up" picture halfway through. (4) Extensive visual understanding capabilities, such as subtitles, visual question and answer, object detection, scene layout, common sense reasoning, etc. (5) Audio and speech recognition: Not mentioned in the Kosmos-1 paper, but Whisper is already an OpenAI API and should be easy to integrate.

Jim believes that there may be some discrepancies in the predictions based on Andreas’ recent announcement. But Kosmos-1 has already done this. There’s reason to believe it provides capabilities for GPT-4 or whatever AI service Microsoft will offer next. "It's hard to believe that Kosmos-1 will remain in the laboratory and not become a product."

Multi-modal large model application examples: image capture, image question and answer, OCR, visual dialogue

Jim advises practitioners, "Please be prepared for multi-modal APIs - they will appear sooner or later!"

2. GPT-4 will become AGI ? Far from enough

First of all, the issue of accuracy is still not enough. When asked about operational reliability and factual fidelity, Siebler, senior artificial intelligence expert at Microsoft Germany, said that the AI ​​will not always answer correctly, so verification is necessary. Microsoft is currently creating a confidence metric to address this issue. Customers typically only use AI support on their own datasets, primarily for reading comprehension and querying inventory data, where the models are already quite accurate. However, the text generated by the models is still generative and therefore not easily verifiable. "We built a feedback loop around it, both for and against," Siebler said. "It's an iterative process."

Secondly, there is not enough data. Even though the multi-modal GPT-4 is about to demonstrate powerful vision, hearing, reading comprehension and reasoning capabilities, this is just the tip of the iceberg of AGI. Taking humanoid robots as an example, it is difficult to unify the control data of the robots, and , these control data are related to the robot hardware and vary greatly. Therefore, training data from different real robots cannot be easily combined, which is qualitatively different from data such as text, video, image, audio, etc.

3. Two rumors about GPT-4

1. One trillion parameters? Fake!

GPT-4 is a new language model being created by OpenAI that can generate text similar to human speech. It will advance the technology used by ChatGPT, which is based on GPT-3.5.

As early as August 2021, industry experts speculated that GPT-4 would have 100 trillion parameters, but at that time some people said: Building AI with more parameters may not necessarily Guarantees better performance and may affect responsiveness.

##But Altman, the father of ChatGPT, quickly refuted the rumors: the next version of ChatGPT will not be AGI, nor will it be AGI. There will be 100 trillion parameters. Those rumors are incorrect.

2.Does Bing chat use GPT-4?

Microsoft has said that the new Bing or Bing Chat is more powerful than ChatGPT. Since OpenAI’s chat uses GPT-3.5, some speculate that Bing chat may use GPT-4. This has not been confirmed.

Obviously, Bing Chat has been upgraded to provide access to current information over the Internet, which is a huge improvement over ChatGPT, which until now could only benefit from training data received before 2021.

In addition to Internet access, the AI ​​models used for Bing Chat are much faster, which is important when taken out of the lab and added to the search engine.

But this seems unlikely to be equivalent to OpenAI’s GPT-4 model. If GPT-4 is already publicly available, there is no need to keep it secret anymore.

4. GPT-4: An evolution, not a revolution

There is no doubt that the upcoming GPT-4 will definitely leave a deep impression on people. , but OpenAI CEO Sam Altman said in an interview with StrictlyVC, "People are 'begging' to be disappointed, and they will be disappointed."

Altman also said on Twitter before : AGI has the potential to cause huge disruption to the world economy, saying that rolling out a few small changes quickly is better than shocking progress that gives the world little chance to adapt.

Coincidentally, world-class artificial intelligence master Ben Goertzel also poured a lot of cold water on GPT-3/GPT-4:

Of course, Now one should note that the amount of dollars and manpower invested in these "knowledge rearranging" systems (like ChatGPT) is far greater than the amount invested in alternative AI approaches that respect more of the roots of grounded, self-correcting cognition Complexity.

The sweeping skepticism of multi-layer neural networks and embrace of expert systems in the late 1970s-early 1990s seemed naive, archaic, and foolish, much like it does now.

Similarly, today’s starry-eyed enthusiasm for LLM and glib dismissal of more subtle AGI approaches will look ridiculous by the mid/late 2020s

My point in this post is not to say that these LLM-based systems are not cool or useful - just that they are a trendy new narrow AI technology that is not as closely related to AGI as It appears on the surface, or so some commenters claim.

In short, GPT-4 will be an evolution, not a revolution.

5. Written at the end

Obviously, ChatGPT 4 is unlikely to be the name of OpenAI’s next product, but it will bring some creative imagination, ChatGPT’s name, most likely combined with the AI ​​model GPT-4 that drives its improvements. AI practitioners and managers need to focus on the current work progress of GPT-4 and ChatGPT, and when OpenAI may release the next major upgrade.

Finally, I would like to tell everyone that it is not important whether the GPT-4 released next week meets the predictions of this article, including the large domestic model "Wen Xinyiyan" we are looking forward to. What matters is whether developers or their companies are ready for multi-modal large model APIs. How fortunate that all of this will hopefully come before 2024!

Reference link:

https://arxiv.org/abs/2302.14045

https://www.heise.de/news/ GPT-4-is-coming-next-week-and-it-will-be-multimodal-says-Microsoft-Germany-7540972.html

https://t.co/JbtQvjoJ3W

The above is the detailed content of Two rumors and latest predictions for GPT-4!. For more information, please follow other related articles on the PHP Chinese website!