What's next for Generative AI?

I came across this amazing article by Sequoia Capital on Generative AI's Act Two. This is absolute must read for anyone interested in world of GenAI and how the future is going to look like.

It talks about how first wave of GenAI apps conquered imagination of consumers and business alike. One things that is without any question is that GenAI is not a passing hot sauce. It is here to stay and it useful.

On the trifecta of better, faster and cheaper, we were always asked to choose 2. But with GenAI, all 3 is possible and that's what makes it a powerful technology. IOT was better and faster. Blockchain was better and cheaper. Quantum computing is getting there but is expected to be better and faster.

GenAI is better, faster and cheaper.

Artificial intelligence capabilities are rapidly advancing, driven by key innovations across the model development stack. Several emerging techniques are enabling more advanced reasoning, easier transfer learning, and improved accuracy and truthfulness through retrieval augmentation. New developer tools and AI infrastructure providers are also accelerating progress.

So as per Sequioa, here's what Act 2 of Generative AI will look like.

1. Emerging reasoning techniques like chain-of-thought, tree-of-thought and reflexion are dramatically improving models’ ability to perform complex reasoning tasks beyond just text generation. With chain-of-thought, models can break down prompts into logical steps, while tree-of-thought allows exploring different branches of possibilities. Reflexion enables models to explain their reasoning process. These approaches close the gap between what users expect from AI and what it can currently deliver. Frameworks like Anthropic's Constitutional AI and Claude make it easy for developers to invoke these reasoning techniques. With Claude's explainability, developers can understand and debug the model's reasoning chains.

2. Transfer learning techniques like Recursive Reward Model Fine-Tuning (RRFM) and fine-tuning substantially lower the barrier for adapting foundation models like GPT-3 to specific domains. Rather than training a model from scratch, developers can fine-tune a foundation model on much less data from their domain. The recent release of GPT-3.5 and models like Anthropic's Llama-2 enable easier fine-tuning. As models are deployed, continuous fine-tuning on user feedback tailors them further. Platforms like Anthropic's Constitutional AI enable safe, privacy-preserving fine-tuning. Transfer learning is making state-of-the-art AI much more accessible.

3. Retrieval-augmented generation enhances model accuracy by retrieving relevant context documents or knowledge during generation. This reduces hallucination and keeps responses truthful and useful. Vector database startups like Pinecone provide lightning-fast similarity search needed to match contexts at scale. Companies can build knowledge bases or ingest customer data to use as retrieval corpora. Anthropic's Constitutional AI framework makes it easy to integrate retrieval modules into end-to-end models. RAG is becoming a critical technique for responsible AI.

4. New developer tools and frameworks simplify building, evaluating and monitoring AI model performance. Application frameworks like Anthropic's Claude provide reusable components for chat, summarization and more. LLMOps tools like Weights & Biases and Langsmith enable easier model experiment tracking, explainability, bias detection, and monitoring of production model accuracy. These tools lead to higher quality, more responsible AI applications.

5. Lastly, a new generation of infrastructure companies provides the optimized hardware AI developers need. Companies like CoreWeave, Lambda Labs, Foundry, Replicate and Modal offer GPU clusters purpose-built for AI, along with major cloud providers like AWS. They provide flexible access to powerful GPUs and easy scaling, with developer-friendly pricing. Some offer PaaS tools optimized for AI. This new infrastructure enables smaller companies and startups to innovate.

Together, these advances across model techniques, transfer learning, reasoning, tooling and infrastructure are driving rapid progress in responsible and capable AI. With innovative approaches to reasoning, accuracy, generalization and developer productivity, AI is becoming far more useful across industries and applications. Leading companies will combine these techniques and leverage new developer tools and infrastructure to create the next generation of AI.