AI Smarts Now Carry A Lot of Price Tags
Glean provides tools for searching through applications such as Gmail, Slack, and Salesforce. Qi said the new AI techniques for parsing speech will help Glean customers find the right file or conversation more easily.
But training with such a cutting edge AI algorithm costs many millions of dollars. That’s why Glean uses small, less capable AI models that don’t get a lot of meaning out of the text.
“It’s hard for smaller areas with smaller budgets to get the same level of results” as companies want Google or Amazon, According to Qi. The most powerful AI models are “out of the question,” he said.
AI has produced exciting results over the past decade – programs that can beat people in complex games, drive cars through city streets under specific conditions, respond to spoken commands, and write in consistent text based on a short step. Writing relies heavily on recent advances in computers ’ability to parse and manipulate language.
Those improvements are mostly a result of feeding algorithms more text as examples to learn from, and giving them more chips to use with it. And that costs money.
Think OpenAI’s speech model GPT-3, a large, mathematical simulation neural network fed reams of text ripped from the web. GPT-3 finds statistical patterns that predict, with surprising coherence, which words should be followed by others. Out of the box, the GPT-3 is better than previous AI models in tasks such as answering questions, summarizing text, and correcting grammatical errors. On one scale, it is 1,000 times more capable than its predecessor, GPT-2. But the training costs GPT-3, in some estimation, nearly $ 5 million.
“If GPT-3 can be accessed and inexpensive, it will completely boost our search engine,” Qi said. “That can be really powerful.”
The increasingly expensive cost of advanced AI training is also a problem for established companies looking to build on their AI capabilities.
Dan McCreary leads an internal team in a division of Optum, an IT healthcare company, that uses speech models to analyze call transcripts to identify patients best. risk or recommend referrals. He said that even training a language model that is one in a thousand the size of the GPT-3 can easily eat up the team’s budget. Models must be trained for specific tasks and can cost more than $ 50,000, which is paid by cloud computing companies to rent their computers and programs.
McCreary said cloud computing providers have little reason to lower the cost. “We can’t trust that cloud providers are working to cut costs for us to build our AI models,” he said. He is looking to purchase specialist chips designed to facilitate AI training.
Part of why AI has evolved most recently is because many academic labs and startups can download and use the latest ideas and techniques. Algorithms that have produced achievements in image processing, for example, have come out of academic labs and developed using off-the-shelf hardware and apparently divided data sets.
Later, however become more clear that the advancement of AI is tied to an exponential increase in the underlying power of the computer.
Many companies have, of course, always had advantages in terms of budget, scale, and reach. And a lot of computer power is being put on the table in industries like drug discovery.
Now, some are pushing to raise things even higher. Microsoft SAYS this week that, together with Nvidia, it built a speech model that is more than twice the size of the GPT-3. Chinese researchers says they built a language model four times larger than that.
“The cost of AI training will go up completely,” said David Kanter, executive director of MLCommon, an organization that tracks the production of chips designed for AI. The idea that larger models could open up many new capabilities can be seen in many parts of the technology industry, he said. It can be explained why Tesla designs its own chips just to train AI models for auto -independent.
Some worry that the rising cost of tapping into the latest and greatest tech could slow the pace of innovation by reserving it for most companies, and those who rent out their equipment.
“I think it will reduce innovation,” he added Chris Manning, a Stanford professor who specializes in AI and language. “If we only had a few areas where people could play around with these models of that size, that should reduce a lot of the creative exploration that’s going on.”