- Cohere announced a fine-tuned AI model that outperforms GPT-4 on some tasks.
- The model is also cheap to run, costing up to 15 times less than large-scale AI systems.
- Cohere is betting on cheaper, business-focused AI to compete with OpenAI and Anthropic.
OpenAI rival Cohere has announced an updated AI model that it says is more convenient and cheaper to run than GPT-4.
The AI startup is rolling out the ability to fine-tune Command R AI models to outperform large-scale models like GPT-4 in some use cases while reducing operating costs by up to 15x. It says that performance can be achieved.
Amid growing concerns about rising costs due to the AI boom, hopes are rising that smaller, cheaper models could rival the larger, more expensive AI systems built by tech giants.
“We found that using smaller models and fine-tuning datasets produced very good results,” Cohere co-founder Nick Frost told Business Insider.
“We have fine-tuned the Command R and when we benchmarked it against its competitors, it outperforms some models in completely different weight classes, and outperforms them at a fraction of the price.” ” he added.
Kohia tested it on tasks such as summarizing meetings and analyzing financial and scientific information, and the fine-tuned version of Command R outperforms GPT-4, GPT-4 Turbo, the most advanced model developed by Amazon. Said it was more accurate than Claude Opus. I support Antropics.
Cohere performs these tests in-house and its fine-tuned Command R model has an accuracy of 80.2% when summarizing meetings, compared to 78.8% for GPT-4 and 77.9% for Claude Opus. I understand that. 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 much lower than GPT-4 and Claude Opus, at $2-4 per million tokens compared to $30-60 for GPT-4. .
Kohia said Command R, which was first released in March, is significantly smaller than something like GPT-4 and therefore costs much less to operate.
Fine-tuning, where users tune models using expert data, also reduces the amount of computation required to run a model by allowing the model to perform better on more relevant tasks. .
Command R model tweaks will be available on Cohere's platform starting Thursday, and will be available on other platforms in the near future.
Cohere bets on enterprise
Training large-scale AI models such as GPT-4 and Meta's Llama requires vast amounts of computing power, and many AI companies are using the We are in a billion-dollar arms race.
Mark Zuckerberg told investors Meta will continue to spend “aggressively” on AI, and OpenAI head Sam Altman last month announced “I don't care if it costs $5 billion or $50 to build an intelligent AI,” he said. $1 billion or $500 billion.
“As long as we can find a way to pay the bills, we're building AGI. It's going to cost money,” Altman told a group of Stanford students.
Toronto-based Cohere took a different approach. The company targets enterprises and enterprise customers, offering small-scale AI models focused on business use at a fraction of the cost of larger models.
“I think there's a very interesting scientific debate about whether only large language models can scale to AGI. I don't think they scale. So we can't just throw more money at compute. , I don't think you'll get results like AGI,'' Frost said.
“Large language models are an amazing technology, and I think they can provide much more value than what we currently offer. But only if they are deployed in real business use cases and made at a reasonable price point. “If only,” he added.
Kohia was valued at more than $2.1 billion last year, but its path hasn't been entirely smooth. The Information reported in March that despite Kohia's high valuation, it had generated just $13 million in annual revenue by the end of last year.
Business Insider understands that annual revenue rose to about $35 million by the end of the first quarter. Frost said Cohere's revenue increased as the company steadily released new models and updates this year.
“This year has started well for us, and I think this is a direct result of our focus on real-world solutions that are actually business-ready, rather than high-flying scientific projects.” he said.
However, the company still faces challenges competing with Big Tech-backed heavyweights like OpenAI and Anthropic.
The landscape for AI startups is not as bright as it was a year ago, with hot companies like Stability AI and Inflection facing trouble in recent months.
Stability cut jobs last month as part of an effort to “focus” on the business after CEO Emad Mostaq resigned following reports the start-up was in financial trouble. .
Meanwhile, Inflexion, once valued at $4 billion, lost co-founder Mustafa Suleiman and some of its staff to Microsoft in March.
Cohere looks forward to focusing on enterprise and low-cost models to carve a niche in the increasingly competitive AI landscape.
“We're interested in making these models as useful as possible,” Frost said.
“We're interested in a world where we use language models every day for every task we do with computers, and we don't need AGI to do that,” he added.
