From left: Ivan Zhang, Aidan Gomez and Nick Frosst, co-founders of Cohere. Cohere
Cohere introduces a fine-tuned AI model that claims to outperform GPT-4 on some tasks.
The model is also cheaper to operate and costs up to 15 times less than larger AI systems.
Cohere is betting on cheaper, enterprise-focused AI and trying to compete with OpenAI and Anthropic.
This is a machine translation of an article from our US colleagues at Business Insider. It was automatically translated and checked by a real editor.
OpenAI competitor Cohere has unveiled an updated AI model that it says is more useful and cost-effective than GPT-4.
The AI startup says it is introducing 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.
This raises hopes that smaller, cheaper models can compete with the tech giants’ larger, more expensive AI systems, as concerns grow about the spiraling costs of the AI boom.
“We’ve found that fine-tuning datasets with a small model produces really great results,” Cohere co-founder Nick Frosst told Business Insider (BI). “The fine-tuning on Command R, when we compared it to the competition, outperforms some models in completely different weight classes and can then be better than them at a tiny fraction of the price,” he added.
Read too
How Nvidia is dominating the AI-obsessed market without even providing numbers
Cohere said the fine-tuned version of Command R was more accurate at tasks such as summarizing meetings and analyzing financial and scientific information than GPT-4, GPT-4 Turbo and Claude Opus, Anthropic’s most advanced model Amazon is supported.
Cohere conducted these tests itself and found that its fine-tuned Command R model achieved 80.2 percent accuracy in meeting summarization, compared to 78.8 percent for GPT-4 and 77.9 percent for Claude Opus . When analyzing financial data, Command R was also 6.2 percent more accurate than GPT-4 and 5.3 percent more accurate than Claude.
The running costs of the fine-tuned model, the so-called inference costs, are also far below those of GPT-4 and Claude Opus: They are between two and four US dollars (roughly between 1.85 and 3.70 euros) per million tokens, compared to 30 to 60 dollars (around 28 to 56 euros) for GPT-4.
Cohere said Command R, which launched in March, is significantly smaller than GPT-4 and therefore costs much less.
Fine-tuning, where users adjust the model with specific data, also reduces the amount of computation required to run the model, making it more suitable for more relevant tasks. Fine-tuning the Command R model has been available on Cohere’s platform since Thursday; availability on other platforms will follow in the near future.
OpenAI competitor Cohere relies on companies, not individuals
The enormous amount of computing power required to train large AI models like GPT-4 and Metas Llama has forced many AI companies into a multi-billion dollar arms race. The way to make AI profitable is still difficult to find.
Mark Zuckerberg told investors that Meta will continue to invest “aggressively” in AI. OpenAI CEO Sam Altman said last month that he “doesn’t care” about the development of artificial general intelligence (breaking latest news) – AI with intelligence greater than that of humans – cost five billion, 50 billion or 500 billion dollars (around 4.6, 46 or 460 billion euros).
“As long as we can find a way to pay the bills, we will develop breaking latest news. It’s going to be expensive,” Altman told a group of students at Stanford University.
Cohere, based in Toronto, takes a different approach. The company targets enterprises and large customers and offers smaller AI models tailored specifically for business purposes at a fraction of the cost of larger models.
“I think there’s a very interesting scientific debate about whether large language models alone can scale to breaking latest news – I don’t think they can. “So I don’t think something like breaking latest news will happen if you just put more money into data processing,” says Frosst.
Read too
I have synesthesia and use AI to “hear” my paintings – it also helps me compose music
“Large language models are an incredible technology. I believe they can deliver so much more value than they currently do. But only if they are actually used in real business applications and if they are offered at a reasonable price,” he added.
Cohere became last year with over 2.1 billion dollars (around 1.9 billion euros), but the journey there was not entirely smooth. „The Information“ reported in March that, despite its high valuation, Cohere had only generated $13 million (around twelve million euros) in annual sales at the end of last year.
BI expects annual sales to have increased to around $35 million (around €32 million) by the end of the first quarter. Frosst said Cohere’s sales increased because the company released a steady stream of new models and updates this year.
“It was a good start to the year for us. I think this is a direct result of us focusing on actual business-ready, real-world solutions rather than lofty scientific projects,” he said.
However, the company still faces the challenge of competing with Big Tech-backed heavyweights like OpenAI and Anthropic. The picture for AI startups looks less sunny than it did a year ago, as emerging companies like Stability AI and Inflection have struggled in recent months.
Read too
Goldman Sachs predicts: AI data centers will drive electricity demand by 160 percent by 2030
Stability made layoffs last month to “focus” operations after CEO Emad Mostaque resigned following reports of the startup’s financial problems. Inflection, which was once valued at $4 billion (about 3.7 billion euros), lost co-founder Mustafa Suleyman and some of his employees to Microsoft in March.
Cohere is leveraging its focus on enterprise and low-cost models to carve out a niche in an increasingly competitive AI landscape. “We are interested in making these models as useful as possible,” says Frosst. “We’re interested in a world where you use a language model every day to help you do everything you use a computer to do. You don’t need breaking latest news for that,” he added.
Disclaimer: Axel Springer, the parent company of Business Insider, has entered into a global agreement that allows OpenAI to train its models on the reporting of its media brands.
Read the original article Business Insider.