Open again after four years

OpenAI is finally going "Open" !

According to the latest news, they are preparing to release a new open source language model.

This is the first time in four years since GPT-2.

Many netizens poked their hands to express their expectations: Is this going to release their own open source replacement?

After all, the best open source model is still far from GPT-4. In terms of parameters alone, one is 20 billion and the other is 1.3 trillion, which is a difference of three orders of magnitude.

That being the case, then OpenAI's move "will change the competitive landscape of the entire large model?" .

Many netizens said that the LLaMA large model may be the first to bear the brunt, that is, the alpaca family.

After all, since the birth of ChatGPT, various open source solutions have emerged one after another, but most of them are inspired by the big model of Meta.

For example, Stanford's Alpaca, Berkeley's Vicuna, Kaola, and ColossalChat, as well as the domestic Harbin Institute of Technology's LLaMA fine-tuning model Huatuo based on Chinese medical knowledge... Some of these open source models have even been optimized to run on mobile devices.

The latest ranking of Chatbot Arena, a large model created by UC Berkeley, shows that many open source models follow GPT-4 and Claude.

However, whether it will be a "replacement" will not be known until the final release.

And whether OpenAI will use this model, it is not yet known whether it will compete with other similar open source models.

To be sure, this new open-source model will be unlikely to compete with GPT, The Information reports, citing people familiar with the matter.

At the same time, someone is naming Google again: Now the pressure on Google is increasing .

There have also been heated discussions about open source and moats.

Open Source and Moats

Open source or closed source? This is a question.

Some time ago, an internal Google document stirred up waves on the Internet, the core point of which is very clear:

The rapid development of open source large models is eroding the positions of OpenAI and Google.

And said, "unless the position of closed source is changed, open source alternatives will eventually eclipse them (including ChatGPT)".

At this point, neither Google nor OpenAI has a moat in this large-scale arms race.

Many open source problems have been solved, such as running on lower power devices, scalable personal AI, multimodality, etc.

Even though OpenAI and Google now have a certain advantage in model quality, the gap is rapidly narrowing.

In the past few weeks, every team in the field of open source AI has made continuous progress, whether it is from the model or the dimension of application.

For example, the AI startup Together , built an open source large model and cloud platform based on LLaMA last month, and now raised $20 million in seed round funding.

This trend even continued offline, and many people reveled and celebrated for the open source movement.

In addition to launching a series of large-scale model tools, the "Open Source Center" HuggingFace even held an offline "Woodstock of AI" gathering, which attracted more than 5,000 people.

Stability AI, the company behind Stable Diffusion, and Lightning AI, which developed PyTorch Lightning, also plan to hold an open source exchange meeting.

In the eyes of many, OpenAI and Google set a bad precedent: the dangers of unsupervised models are real.

Previously, while the models of these big tech companies may not have been fully replicated; the open source community knew what the basic ingredients of the "secret sauce" were. But now no one can know its ingredients.

What do you think about this matter?

Reference link:

[1] https://www.reuters.com/technology/openai-readies-new-open-source-ai-model-information-2023-05-15/

[2] https://www.theinformation.com/articles/open-source-ai-is-gaining-on-google-and-chatgpt

[3] https://venturebeat.com/ai/open-source-ai-continues-to-celebrate-as-big-tech-mulls-over-moats/