Powerful open-source AI models out-of-the-box

(AI generation/Shutterstock)

ChatGPT and GPT-4 models are making all the headlines these days. But if you look around, you’ll find a competent open source model that’s free, unlimited, and with great performance that you can run yourself. A Google researcher acknowledged a similar thing in a leaked “We Have No Moat” memo earlier this month, but those in the computing industry are aware that a ton of new and exciting AI models are being released every day. Not likely, experts say.

The launch of ChatGPT on November 30, 2022 will go down in history as a defining moment in the democratization of AI. To find a comparable moment, we have to go back to June 29, 2007, when Steve Jobs unveiled his Apple iPhone. OpenAI CEO Sam Altman’s testimony before the U.S. Senate this week shows not only the incredible potential of AI technology, but also the fear of what might come next.

ChatGPT buzz is also spurring action in the machine learning and AI community as technologists look for ways to enhance any cognitive task with AI capabilities. From data analytics data management to his ERP and app development, every application has his AI-based co-pilot to improve human capabilities.

Luis Ceze, professor of computer science at the University of Washington and CEO of OctoML, is as amazed by the progress as we are. The machine learning expert credits OpenAI with his current AI boom.

“It’s absolutely amazing what’s happening now,” he says. “Everything about ChatGPT feels great, not only as a great demonstration of the technology, but also in terms of bringing attention to this feature … Now it’s time to show people what all these models can do. No need to explain.”

(Eble Omar/Shutterstock)

Large Language Models (LLM) have been around for a while, but OpenAI deserves credit for opening the door to new AI use cases. However, future users should be careful not to get too caught up in the idea that his OpenAI is the only game when it comes to AI. In fact, there are many other alternatives to GPT-4 and ChatGPT that are likely better suited for his AI builder than OpenAI’s API, he says.

“I have seen many customers and users start using OpenAI and realize, ‘Oh, this is all I really need,’” he says. Data Nami. “they [OpenAI] Building great models requires significant investment to get to where they are today. But the use cases people are often interested in don’t need that level of functionality. ”

In some cases, such as text summaries, using ChatGPT is like “getting coffee in a Ferrari,” Ceze says. “There are many cars that you can use for free.”

The “free cars” that OctoML helps users ride in include RedPajama, Vicuna from LMSYS.org, and Dolly from Databricks. Stable Diffusion and Dall-E 2 come to mind for Ceze. Meta’s LLaMa model was leaked online and people are building on it. Stanford’s Alpaca is based on LLaMa. Hugging Face gives him access to over 13,000 models. There are many more that already exist, and more are emerging every day, he says.

“The pace of progress is ferocious,” says Cezet. “I’m talking about the big improvements that are going on. No exaggeration, there are amazing new models coming out every day. I don’t think people realize that. “

In addition to being free, the open source model has other advantages. First, users can run them on their own infrastructure, giving them more control. If you have data science skills, you can tweak the weights to tune the algorithm or even train the model to run on your own data. There is also the issue of increasing the risk of running a model trained on unknown data, which has been criticized for OpenAI’s conduct.

The pace of innovation with LLMs is accelerating (Image source: LMSYS.org)

Two weeks ago, Google researchers backed up Ceze’s view when an internal memo, “We don’t have moats, we don’t have OpenAI,” leaked on the Discord server. people of semi-analysis It confirmed the memo’s authenticity and published a moderately redacted version.

according to semi-analyticalA Google researcher wrote:

“Our model still has a slight edge in quality, but the gap is closing surprisingly fast. The open source model is faster, more customizable, more private, They’re doing things with parameters of $100 and 13 billion while we’re struggling with $10 million and $540 billion. is doing it in weeks, not months, which means a lot to us.”

There is no “secret sauce” for what Google is doing with Bard and other models, the researchers wrote. Improvements are made through community collaboration rather than mixing digital potions in Google Labs. And when smaller language models start doing things that previously only large language models could do, the gap narrows until the advantage disappears.

Since LLaMa leaked online in March, we’ve seen a “tremendous outpouring” of innovation, the study says. It took about a month for the community to start building the underlying model, including variants that incorporate instruction tuning, quantization, quality enhancement, human evaluation, multimodality, and RLHF. [reinforcement learning human feedback] to a new level.

“Most importantly, they’ve solved the scaling problem to the point that anyone can tinker,” the researchers write. Many of the new ideas come from ordinary people. The barriers to entry for training and experimentation have dropped from the total output of major research organizations to one person, one night, and a bulky laptop. ”

RedPajama is a variation of Meta’s LLaMa base model

In other words, even if there was a moat in the first place, a permanent bridge was built over it.

“People will not pay for a restricted model if free and unrestricted alternatives are of equal quality,” the Google researchers wrote. “We should think about where our added value really lies.”

Of course, this is great news for everyone from hobbyists who want to tinker, to researchers who want to experiment, to Fortune 100 companies looking to incorporate AI into their business processes.

One of the most fascinating technological advances in recent history has become freely available to anyone with a laptop, internet connection, and the curiosity to look around. Barriers to entry for using this technology have already been lowered significantly, further reductions are almost certain, and PhDs will have to explore even greater innovations, leaving the rest of the population already available. can be used to create new applications or reinvent your business.

“People are giving a lot of attention to closed large-scale models like OpenAI and Cohere, but I don’t think there’s enough recognition that open source models are incredibly capable,” Ceze said. increase. “I think that’s what people aren’t paying enough attention to.”

Related products:

AI Chatbots: A Hedge Against Inflation?

LLM is Dinosaur-Killing Meteor for Old BI, Says ThoughtSpot CEO

ChatGPT Dominates as Most In-Demand Skill at Work: Udemy Report