The Four Wars of the AI Stack (Dec 2023 Recap)
Section 1: Introduction
- There are four main “wars” or battlegrounds among AI companies and stakeholders: the Data War, GPU/Inference War, Multimodality War, and RAG/Ops War.
Section 2: The Data War
- OpenAI partners with Axel Springer to provide news and citations to ChatGPT.
- NYT sues OpenAI to shut down all GPT models.
- Apple offers $50m for data contracts with publishers.
- Increased interest in synthetic data at NeurIPS and Deepmind.
Section 3: The GPU/Inference War
- Release of Mixtral sparked price cuts, down to $0.27 per million tokens.
- Major inference providers like Anthropic, Fireworks, Perplexity slashed prices to undercut Mixtral.
- Analysis shows profitable lower bound is 0.51−0.76 per million tokens.
Section 4: The Multimodality War
- Midjourney launched v6 and a web UI.
- Assembly AI raised $50m for “Stripe for AI models”.
- Replicate raised $40m to serve AI engineers.
- Suno AI launched for audio generation.
- OpenAI and Google continue work on “God Models”.
Section 5: The RAG/Ops War
- Debates on need for vector DBs vs new ones like Turbopuffer.
- LangChain vs LlamaIndex for chaining and execution.
- LLMOps developments like HumanLoop vs frameworks like LangSmith.
Section 6: Open Models
- Release of Mixtral sparked GPU/inference price war. Up to 99.55% drop in cost.
- New model architectures like Mamba, RWKV announced.
- Apple released MLX framework with Mixtral.
- Microsoft released Phi-2 model.
Section 7: Fundraising News
- Notable fundraises: Midjourney (200m/yr),Anthropic(750m), Mistral (400m),Glean(200m), Harvey ($80m).
Section 8: Other Launches
- Midjourney launched v6 and web UI.
- OpenAI released logprobs for ChatGPT.
- LlamaIndex launched step-wise agent execution.
- New models like Suno AI (audio), Phi-2 (Microsoft), TextDiffuser 2 (images).
Section 9: Conclusion
- Key themes were data, infrastructure, multimodal, and tooling.
- Progress made across models, verticals, tools, and infrastructure.