The open-source AI boom is built on Big Tech’s handouts

 Greater access to the code behind generative models encourages innovation. But if the top companies get spooked, they could go out of business.

Last week, a leaked memo purportedly written by Luke Sernau, a senior engineer at Google, said aloud what many in Silicon Valley must have been whispering to themselves for weeks: an open-source free-for-all threatens control of Big Tech. AI.

New big open-source language models—alternatives to Google's Bard or OpenAI's ChatGPT that researchers and app developers can study, build upon, and modify—are falling like candy from a piñata. These are smaller, cheaper versions of best-in-class AI models built by big companies that (almost) match them in performance – and they're shared for free.

Companies like Google — which revealed at its annual product keynote this week that it's throwing generative AI into everything it has, from Gmail to Photos to Maps — have been too busy looking over their shoulders to see the real competition coming, Sernau writes: " While we were arguing, the third faction was quietly eating our lunch.”

In many ways, this is a good thing. Greater access to these models has helped drive innovation – it can also help catch their flaws. Artificial intelligence will not flourish if only a few mega-rich companies care about the technology or decide how it will be used.

But this open-source boom is uncertain. Most open source versions still stand on the shoulders of giant models published by big companies with deep pockets. If OpenAI and Meta decide to close up shop, the boomtown could become a backwater.

For example, many of these models are built on LLaMA, an open-source large language model released by Meta AI. Others use a massive public dataset called the Pile, put together by the open-source nonprofit EleutherAI. But EleutherAI only exists because the openness of OpenAI meant that lots of coders were able to reverse engineer how GPT-3 was made and then build their own at their leisure.

"Meta AI has done a really great job of training and releasing the models to the research community," says Stella Biderman, who splits her time between EleutherAI, where she's executive director and head of research, and consulting firm Booz Allen Hamilton. Sernau also highlights the key role of Meta AI in his Google note. (Google confirmed to MIT Technology Review that the memo was written by one of its employees, but notes that it is not an official strategy document.)

that could change. OpenAI is already reversing its previous open policy due to competition concerns. And Meta may start wanting to limit the risk of newcomers doing nasty things to its open-source code. "I honestly feel like it's the right thing to do right now," says Joelle Pineau, CEO of Meta AI, about opening up the code to outsiders. “Is this the same strategy we will adopt for the next five years? I don't know because AI moves so fast."

If the trend of closing access continues, not only will the open source crowd be swept away, but the next generation of AI breakthroughs will be squarely back in the hands of the world's largest and wealthiest AI labs.

The future of how AI is made and used is at a crossroads.

Open source bonanza

Open-source software has been around for decades. That's what the internet runs on. But the cost of creating powerful models meant that open-source AI only took off a year ago. It quickly became a bonanza.

Just look at the last few weeks. On March 25, Hugging Face, a startup that promotes free and open access to AI, introduced the first open-source alternative to ChatGPT, the viral chatbot released by OpenAI in November.

Hugging Face's chatbot, HuggingChat, is built on an open-source large conversational-tuned language model called Open Assistant, which was trained with the help of about 13,000 volunteers and released a month ago. But Open Assistant itself is built on Meta's LLaMA.

And then there's StableLM, an open-source large language model released on March 19 by Stability AI, the company behind the hit Stable Diffusion model for text-to-image conversion. A week later, on March 28th, Stability AI released StableVicuna, a version of StableLM that – like Open Assistant or HuggingChat – is optimized for conversation. (Think of StableLM as Stability's answer to GPT-4 and StableVicuna as ChatGPT's answer.)

These new open-source models join a number of others released in the past few months, including Alpaca (from a team at the University of Stanford), Dolly (from software firm Databricks), and Cerebras-GPT (from artificial intelligence firm Cerebras). Most of these models are built on LLaMA or datasets and models from EleutherAI; Cerebras-GPT follows a template set by DeepMind. You can bet there will be more to come.

For some, open-source is a matter of principle. “This is a global community effort to bring the power of conversational AI to everyone… and take it out of the hands of a few large corporations,” says AI researcher and YouTuber Yannic Kilcher in a video introducing Open Assistant.

"We will never give up the fight for open source AI," Hugging Face co-founder Julien Chaumond tweeted last month.

For others, it's about profit. Stability AI hopes to repeat the same trick with chatbots that it pulled with Images: driving and then benefiting from a flurry of innovation among developers who use its products. The company plans to take the best of this innovation and bring it back into tailor-made products for a wide range of clients. “We add innovation and then we pick and choose,” says Emad Mostaque, CEO of Stability AI. "It's the best business model in the world."

Either way, the huge crop of free and open large language models is putting the technology in the hands of millions of people around the world, inspiring many to build new tools and explore how they work. “This technology has a lot more access than ever before,” says Biderman.

"The incredible number of ways people are using this technology is frankly mind-boggling," says Amir Ghavi, a lawyer at Fried Frank, which represents a number of generative AI companies, including Stability AI. "I think it's a testament to human creativity, which is the whole point of open-source."

GPU melting

But training large language models from scratch—rather than building on or modifying them—is hard. "It's still out of reach for the vast majority of people," says Mostaque. "We melted a lot of GPUs building StableLM."

The first version of Stability AI, the Stable Diffusion text-to-image model, performed as well as—if not better than—closed equivalents such as Google's Imagen and OpenAI's DALL-E. Not only was it free to use, but it also ran on a good home computer. Last year, more than any other model, Stable Diffusion fueled the explosion of open-source development around image AI.