Cohere
- Cohere unveils fine-tuned AI model it says outperforms GPT-4 on some tasks.
- The model is also cheaper to run, costing up to 15 times less than larger AI systems.
- Cohere is betting on cheaper, business-focused AI as it tries to compete with OpenAI and Anthropic.
OpenAI rival Cohere has unveiled an updated AI model it says is more useful and cheaper to run than GPT-4.
The AI startup says it is rolling out the ability to fine-tune its Command R AI model, allowing it to outperform larger models like GPT-4 in some use cases while costing up to fifteen times less to operate.
It raises hopes that smaller, cheaper models might be able to match the larger, more expensive AI systems built by tech giants as concerns grow over the spiraling costs of the AI boom.
"We have found that fine-tuning on data sets with a small model gets really great results," Cohere cofounder Nick Frosst told Business Insider.
"Fine tuning on Command R, when we benchmarked it against the competition, outperforms some models in completely different weight classes and can then do better than them at a tiny fraction of the price," he added.
Cohere said when tested on tasks such as summarizing meetings and analyzing financial and scientific information, the fine-tuned version of Command R was more accurate than GPT-4, GPT-4 Turbo, and Claude Opus, the most advanced model built by Amazon-backed Anthropic.
Cohere performed these tests itself, and found that its fine-tuned Command R model scored 80.2% on accuracy when summarizing meetings, compared to 78.8% for GPT-4 and 77.9% for Claude Opus. Similarly, when analyzing financial data Command R was 6.2% more accurate than GPT-4 and 5.3% more accurate than Claude.
The running cost of the fine-tuned model, known as the inference cost, is also far below GPT-4 and Claude Opus, costing between $2 to $4 per million tokens compared to $30 to $60 for GPT-4.
Cohere said that as Command R, which initially launched in March, is significantly smaller than the likes of GPT-4, it costs much less to run.
Fine-tuning, which sees users tailor the model with specialist data, also reduces the amount of computation required to run the model by making it better at more relevant tasks.
Fine-tuning on the Command R model is available on Cohere's platform from Thursday, with availability on other platforms coming in the near future.
Cohere bets on enterprise
The massive amount of computer power needed to train large AI models like GPT-4 and Meta's Llama has forced many AI companies into a multi-billion-dollar arms race, even as the path to making AI profitable remains elusive.
Mark Zuckerberg told investors that Meta will continue spending "aggressively" on AI, and OpenAI boss Sam Altman said last month that he "doesn't care" if building Artificial General Intelligence — AI with above human-level intelligence — costs $5 billion, $50 billion or $500 billion.
"As long as we can figure out a way to pay the bills, we're making AGI. It's going to be expensive," Altman said to a group of students at Stanford University.
Cohere, which is based in Toronto, has taken a different approach. The company is targeting businesses and enterprise customers, offering smaller AI models specifically tailored to business uses at a fraction of the cost of larger models.
"I think there's a very interesting scientific debate to be had about whether or not large language models alone will scale to AGI — I don't think they will. So I don't think just throwing more money into compute will result in something like AGI," said Frosst.
"Large language models are an incredible technology. I think they can deliver so much more value than they're delivering currently. But only if they're actually put into real business use cases, if they're made at a reasonable price point," he added.
Cohere was valued at over $2.1 billion last year, but the road hasn't been completely smooth. The Information reported in March that despite its lofty valuation, Cohere was generating only $13 million in annualized revenue by the end of last year.
Business Insider understands that annualized revenue had risen to around $35 million by the end of Q1. Frosst said that Cohere's revenue had increased due to the company releasing a steady stream of new models and updates this year.
"It's been a good start to the year for us. I think that is a direct result of us focusing on actually business-ready and real-world solutions rather than lofty science projects," he said.
However, the company still faces a challenge in competing with Big Tech-backed heavyweights like OpenAI and Anthropic.
The picture for AI startups looks less sunny than a year ago, with buzzy firms like Stability AI and Inflection encountering problems in recent months.
Stability conducted layoffs last month as part of an effort to "focus" its operations after CEO Emad Mostaque resigned, following reports that the startup was experiencing financial problems.
Meanwhile, Inflection, which was once valued at $4 billion, lost cofounder Mustafa Suleyman and a chunk of its staff to Microsoft in March.
Cohere is counting on its focus on enterprise and low-cost models to help it carve out a niche in an increasingly competitive AI landscape.
"We're interested in making these models as useful as possible," said Frosst.
"We're interested in a world where every day you use a language model to help you in any of the things you're using a computer with. You don't need AGI for that," he added.