## The Trend For startups CTOs and CIOs in large corporate and Enterprise companies, a clear trend seems to emerge as a golden path comprised of three phases for their AI deployment journey: 1. Start with a large hosted proprietary model to ship your MVP. 2. As you grow and try to balance effectiveness and efficiency, switch to an open-source model, hosted or self-hosted. 3. As you accumulate real-world feedback, invest in fine-tuning existing or more specialized models. ### Phase 1: Ideation and Growth [OpenAI](https://openai.com/) seems like a clear winner so far: - The service is immediately available, doesn't require any infrastructure, and is easy to test through [ChatGPT](https://chatgpt.com/). - [OpenAI's API](https://openai.com/api/) is the de-facto industry standard and has pervasive integrations with AI frameworks. - Brings off-the-shelves [enterprise compliance](https://learn.microsoft.com/en-us/azure/ai-services/security-controls-policy) through its partnership with Microsoft. ### Phase 2: Optimization As open-source models are catching up with proprietary ones, Meta with [Llama 3](https://llama.meta.com/llama3/) seems to be the most popular choice to optimize costs: - Llama 3 has a great [cost/performance/value ratio](https://artificialanalysis.ai/models/llama-3-instruct-70b). - It is [open-source](https://github.com/meta-llama/llama3) (so far), including its weights, and benefits from the Meta's reputation in the open-source ecosystem (React, Hermes, PyTorch, Faiss...). - It is compatible with the most [popular AI frameworks](https://llama.meta.com/docs/integration-guides/langchain/) and available through the most [popular hosters](https://llama.meta.com/docs/llama-in-the-cloud/microsoft-azure/). ### Phase 3: Specialization As software vendors are maturing, it is still hard to tell who the leader in this future phase will be. Still, my guess is that it will rely on open-source models. ## Perspectives Looking at the current trends, it looks like Meta could be ahead of the competition eventually, thanks to their extraordinarily strong competitive assets. ### Ability to Execute Meta is in an advantageous position to execute its AI strategy for several reasons: - They have one of the [largest AI infrastructure](https://about.fb.com/news/2024/04/introducing-our-next-generation-infrastructure-for-ai/) and the industrial ability to host large-scale AI workloads. - Designing [custom-made chips](https://about.fb.com/news/2024/04/introducing-our-next-generation-infrastructure-for-ai/) to support these workloads. - They have [industrial scale budgets](https://www.businessinsider.com/mark-zuckerberg-plan-meta-make-money-from-ai-2024-4). - They control access to the largest user base and distribution channel with [Meta AI](https://www.theverge.com/2024/4/18/24133808/meta-ai-assistant-llama-3-chatgpt-openai-rival). - They have their own LLM model, [Llama 3](https://llama.meta.com/llama3/), but more importantly the data and feedback pipeline that they'll be able to use through FaceBook, Instagram, WhatsApp... ### Completeness of Vision Meta has learned from Google's strategies and has a sharp vision of how to [leverage AI to make money](https://www.businessinsider.com/mark-zuckerberg-plan-meta-make-money-from-ai-2024-4): 1. Lead the open-source ecosystem through their almost unlimited investment capabilities, delivering AI models at near-zero marginal cost. 2. Kill proprietary platforms in the long run by defining the standards and owning the ecosystem. 3. Like Google did with Chrome and Android, make sure that owning the platform will drive usage to their profitable services. ## Final Thoughts The future is uncertain and indeed shifts are happening every day. > ***Who has the best resources and capabilities to compete with Meta?***