In 2017, Aidan Gomez, who you can see in the center of the image above, was the youngest member in an exceptional group of eight researchers. They were all part of Google Brain, the division that eventually merged with DeepMind. And all of them, at the time, were publishing the most crucial study in artificial intelligence seen in years.
Their research, titled "Attention is all you need," introduced the concept of the transformer, and catalyzed the emergence of generative AI models and chatbots like ChatGPT.
Eventually, Gomez joined British computer scientist Geoffrey Hinton at his lab in Toronto, but he only stayed there briefly. In 2019, he co-founded Cohere with Nick Frosst (the person on the right in the photo) and Ivan Zhang (left).
They had all studied at the University of Toronto, which helped them come together to create an AI startup that's started to make noise and could compete with established AI companies.
The Difference Between Cohere and the AI Giants
- Cohere is a Canadian company with about 300 employees, and recently it's become an exciting alternative in the sector. This is because it takes a quite different approach than its competitors, especially OpenAI, which it’s often compared to. But why? There are three reasons.
- SaaS model. Companies like OpenAI provide access to AI models through an API and charge for each token generated by its LLMs. In addition, these companies run the queries on substructures they have set up in cloud infrastructures, similar to those used by Microsoft, Google, or Amazon. Cohere proposes a SaaS (software-as-a-service) model where the customer already has their own infrastructure and uses Cohere’s models, for which it charges a commission. As Gomez explains in an interview with Fortune, “That leads to much higher margins for us, because we’re not paying for that compute.”
- Chatbots don't matter (that much). While AI giants like OpenAI and Google have chatbots, Cohere doesn’t have one. These types of assistants, which use a freemium model, require a massive computing infrastructure. While they can undoubtedly convince many users to switch to their paid plans, Cohere prefers to save a lot on the cost of inference (generating answers). “We’re starting to hit the inflection point now where spending on inference compute is becoming higher than spending on training, which is indicative of market maturity,” Gomez said.
- Ad hoc models for enterprises. Cohere's other major differentiator is that it's not targeted to end users, but to enterprise customers. And to those companies, they offer highly tuned models for specific purposes. Frosst put it this way to Business Insider: “We’ve found that tuning small models with [specialized] data sets yields big results.”
Small, Accurate, and Inexpensive models
Cohere's strategy is gaining traction. In March 2024, the company launched Command R, a scalable generative model aimed specifically at enterprise clients. A month later, it launched Command R+, a supercharged version of the model that also offers a 128k token context window.
The Coral interface in Cohere allows the evaluation of models such as Command R.
Companies can use both (with limitations and by getting a free API trial) on platforms like OpenRouter, Hugging Face, and from Cohere’s platform through Coral, its user interface to evaluate its models.
Cohere promises something important in both cases: It uses a Retrieval Augmented Generation (RAG) system that, according to its creators, reduces common “hallucinations” in AI models. To accomplish this, it incorporates citations and references—similar to what Perplexity.ai does, for example—and, above all, tries to adapt to the specific needs of each company and each possible use case.
According to Cohere’s founders, this offers significant advantages. In internal tests, they found that the fine-tuned versions of Command R were more accurate than their competitors in analyzing scientific and financial information.
For example, Command R’s accuracy was 80.2%, GPT-4’s was 78.8%, and Claude Opus’ was 77.9%. When analyzing financial data, Cohere’s model was 6.2% more accurate than OpenAI and 5.3% more accurate than Anthropic’s, according to the company's internal tests.
Even more important and exciting is the cost. Running these fine-tuned Cohere models, called inference costs, is significantly cheaper than using OpenAI models: generating one million tokens costs between $2 and $4 with Cohere but between $30 and $60 with GPT-4.
These aren’t the only proposals from Cohere, which presented its Aya 23 8B and 35B models just a few days ago. The models are similar to Meta's Llama 3, featuring open-weight models and apparently remarkable behavior. In addition, they're available in 23 different languages.
But There’s Also Uncertainty
In June 2023, Cohere announced a $270 million funding round led by Inovia Capital, which included giants like NVIDIA, Oracle, and Salesforce.
This investment brought the company’s valuation to $2.29 billion, a remarkable figure. Still, it’s far from the investments made in companies like OpenAI ($11.3 billion according to TechCrunch) or Anthropic ($450 million).
In recent weeks, rumors have swirled about a new round of investment of an additional $500 million, bringing the company’s market valuation to $5 billion. This amount would give the company more room to grow and solidify its commitment to a model that's different from those offered by other AI companies. But there’s also uncertainty about its future.
First of all, revenues are still very modest. According to The Information, Cohere only brought in $13 million in 2023, although things improved at the beginning of 2024: By the end of the first quarter of the year, that figure had risen to $35 million.
These figures lag far behind the theoretical revenues of Anthropic and OpenAI. As stated by The Information, Anthropic is expected to generate more than $850 million in 2024, while OpenAI could bring in $5 billion.
Competition is also accelerating, especially for large companies that invest vast amounts of money in developing their models and making them available to all audiences.
The emergence of open-source models—even if they’re not fully open-source—such as Llama 3 also threatens Cohere’s value proposition, especially since it facilitates the operation of customized implementations. These implementations must be carried out by the experts at companies themselves, who must be able to train and fine-tune these models in a secure and private manner.
Cohere has a great chance of becoming relevant in this increasingly competitive market. Still, it will be interesting to see if its approach, which is markedly different from that of its competitors, bears the fruits it hopes for.
Image | Cohere
Related | I’ve Tried Out Google’s Project Starline, and It’s the Most Amazing Video Call I've Ever Had
View 0 comments