Recently, in the open source model community, Command R+ is in the limelight.
With more than 13,000 supporters, Command R+ jumped to No. 6 on the Arena list, matching GPT-4-0314 with its first ever win against GPT-4's open-weighted model.
Source: https://twitter.com/lmsysorg/status/1777630133798772766
The Arena list is one of the only two trusted test benchmarks mentioned by Andrej Karpathy, the AI god who recently used pure C language to hand-rub GPT-2.
Command R+ is part of Cohere, a large model startup headquartered in Toronto. The co-founder and CEO of this AI start-up is none other than Aidan Gomez, the youngest author of Transformer. Let's take a look behind the scenes of the popular fried chicken in the open source model world.
Command R+ 横空出世，何以击败GPT-4
Like its predecessor, Command R+, it has a 128k token context window, and on top of that, performance has been improved across the board.
First, it provides advanced Retrieval Enhanced Generation (RAG) capabilities. RAG technology combines the advantages of search-based and generative models. The former involves accessing and extracting information from a large number of sources such as databases, articles, or websites, while the latter excels at generating coherent and context-aware text. By combining these two components, RAG stands out for generating more informative and contextually relevant responses.
What's more, the new model is optimized for advanced RAG, improves response accuracy, and provides embedded source references to reduce hallucinations to better respond to business needs, enabling businesses to scale with AI to support a variety of business functions such as finance, human resources, sales, marketing, and customer support.
In addition, it covers more than 10 languages to support global business operations. These include: English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Arabic and Chinese.
In addition, Command R+ is equipped with a variety of tools to automate complex business processes
API Interface: Comes with the ability to use tools that can be accessed through the Cohere and LangChain APIs. This helps automate complex business workflows such as updating CRM tasks, activities, and records.
Multi-Step Tool Usage: A new feature in Command R+ that enables models to combine multiple tools for multiple steps to accomplish complex tasks. Command R+ also has the ability to self-correct when it tries to use a tool and fails (e.g. encountering an error or malfunction of the tool), which enhances the likelihood that the model will complete the task and increases the likelihood of success.
In terms of conversational tool usage and single-turn function invocation capabilities, Command R+ outperformed OpenAI's GPT-4 Turbo, as well as Anthropic's Claude 3 Sonnet and Mistral Large.
Source: https://txt.cohere.com/command-r-plus-microsoft-azure/
Cohere与Oracle，
It's not the relationship between OpenAI and Microsoft
As a company that serves enterprise users, Cohere doesn't want to be a henchman of the giants.
Although Oracle is a major investor in Cohere, and the two have a close relationship, they don't like OpenAI's "get-a-half" model with Microsoft. Cohere's CEO Gomez said Cohere is independent of any cloud service provider and can deploy its models on any cloud platform, unlike OpenAI, which is only available in Microsoft Azure.
"We believe independence is important, so our services can be used on every cloud platform you know – Azure, GCP, OCI, AWS, and on-premises. You're not locked into a single tech stack or cloud platform," Gomez said on the sidelines of the 2024 World Economic Forum in Davos. This independence gives Cohere the potential to open up more revenue streams.
"We don't accept huge checks from a single cloud service provider, which could lock us into an ecosystem or environment. We're really trying to stay independent and build something new for the world," he added, indirectly criticizing OpenAI.
Martin Kon, COO of the business, echoed a similar sentiment, saying that models need to be cloud-agnostic so that you can deploy them where you feel most comfortable with your data, without being tied to a specific cloud or even on-premise.
However, Oracle is one of Cohere's deepest funders and has a strong technical relationship with the company. Cohere trains and builds its generative AI models on Oracle Cloud Infrastructure (OCI), which provides high-performance and low-cost GPU clustering technology. This allows Cohere to accelerate the training of large language models (LLMs) while reducing costs. Similarly, Microsoft's Azure cloud computing platform is also providing computing power support for OpenAI.
"The relationship with Oracle has made a huge impact in computing, providing us with the best supercomputer on the planet, while also being important in terms of entering the market together, co-creating new products, retroinventing existing products, and bringing this technology to the enterprise," Gomez said.
Gomez added, "OCI's generative AI service truly exemplifies our mission to build large language models for enterprises in a way that is extremely protective of their data and completely secure."
Cohere's generative AI models have been integrated into Oracle's commercial applications, including Oracle Fusion Cloud, Oracle NetSuite, and Oracle's industry-specific applications.
Today, Cohere is working hard to improve its bottom line. At the end of last year, the company's annualized revenue was about $13 million, with monthly revenue just over $1 million. That figure has been left far behind by his formidable rival, OpenAI, which surpassed the $2 billion mark in revenue last December.
However, the startup has informed investors that its sales funnel, including potential contracts that are expected to be completed by the end of 2024, is worth more than $300 million, and it is unclear what Oracle's share of that is. What is certain is that as Cohere's model is deployed across multiple cloud platforms, his revenue curve will continue to climb.
Microsoft Azure Announces Adoption of Cohere
Under OpenAI, large models seek a way to make a profit
The first year of large models has passed, and the wave of AI application is surging. Fu Sheng predicts that more than half of the 100 billion models will wither this year.
Under the bright light of OpenAI, either according to the cat or the tiger, or find another way, large model manufacturers must explore their own ways to make money.
Cohere, which focuses on corporate services, has been determined to B early. Martin Kon, CEO of Cohere, said: "We don't have and we won't have a cash-burning consumer chatbot. ”
On the evening of April 11, Robin Li expressed in his internal speech why Wenxin Yiyan chose the road of closed source, he mentioned, "Closed source is a real business model, it can make money, and only by making money can it gather computing power and talents, and closed source has an advantage in cost, as long as it is the same ability, the reasoning cost of the closed source model must be lower, and the response speed must be faster." ”
But various other alternatives, such as Llama, Mistral AI, and Command R+, are booming and increasingly being adopted by cloud providers and enterprises. The power of open source should not be underestimated. Command R+, for example, outperforms similar models in the scalable market category and is more competitively priced.
Cohere's 1 million input and output tokens for Command R+ cost $3 and $15, respectively, which is quite competitive. Compared to other products, its pricing is comparable to that of the Claude 3 Sonnet, while the latest OpenAI GPT-4 Turbo model charges $10 for 1 million input tokens and $30 for 1 million output tokens.
With the continuous influx of new competitors in the field of AI innovation, under the current trend of steady development, it is believed that the debate on "open source VS closed source" will also accompany every step of exploration under the wave of innovation, but this will not be the final game of "the east wind overpowers the west wind", and the two models will inevitably coexist for a long time to create value for the implementation of the AI industry.
Reference Links:
1.HTTPS://AnalyticSynthesis.com/Vo-Needs-Makeup-GPT-4/
2.HTTPS://AnalyticSindiamag.com/te-secret-superstar-of-of-llms/
3. Hatps://k.cena.com.cn/article_1654203637_629924f5020011TVI.html