With so much buzz around AI, it's easy to get lost in the noise.
So let's break it down. In every issue, we'll dive into:
the latest AI developments
how they can be practically applied
how they can shape your business both today and in the future
This Week's Insights
Multi-agent systems are becoming critical for scaling AI applications
Swarm simplifies workflows through functional decomposition and modularity
Whether to bring AI to your workflow or integrate workflows into AI systems
Centralized services offer ease, but decentralized solutions offer flexibility
AI tools are advancing biotech but require balancing innovation with safety
🐝 Agents All The Way Down
Following OpenAI's latest cookbook, Swarm has been all the rage – and not without at least some merit. The underlying ideas aren’t ground-breaking; the beauty lies in their intuitive, simple application.
Many of us have built our own Swarm implementations. Whether borrowing from the fields of distributed computing, workflow orchestration, or domain driven design, the blueprint has been there all along. Scale with complexity by decomposing your applications along functional lines, and tuning each individually. Separate your orchestrator from your executors too for an added boost. Pretty fundamental stuff!
But in a space so dominated by hype and an endless conveyor belt of VC-backed tools looking for use-cases, the fundamentals are so often lost. So let's do two things.
First, let's appreciate that Swarm is a pattern – not a tool. Sans dependencies, sans breaking changes.
Second, let's reorient ourselves on the direction we're heading: and that's towards multi-agent architectures with multi-modal dialogue and multi-model inference. Prompt Engineering is the bottle neck right now and architectures like Swarm help us get past it by facilitating better testing.
As a staunch advocate of bridging the gaps between GenAI and both classical Software Engineering and ML, I am pleased that the space is having its own Unix moment.
The question is, how long until the monorepo movement finds its twin in a monoagent movement?
🎨 Claude, Canvas, and (De)Centralization
The world of LLM SaaS is heading in two competing directions.
One direction is towards decentralization. You can access your LLM in more areas than ever before: Claude Dev and Amazon Q bring it to VSCode; NotebookLM and other GCP services bring GenAI to all of Google Workspaces; and numerous others
The other, centralization. Increasing investment in the B2C LLMaaS offerings of OpenAI, Anthropic, Gemini, and others exert a gravitational pull on the users.
The launch of OpenAI’s Canvas, just like Claude’s Artifacts, is a pull toward centralization.
The question is which is right for you? Do you bring an LLM to your workflow, or bring your workflow to an LLM?
It depends (of course). And based on the major LLM companies’ dual B2B and B2C approach, they know it too.
While Canvas makes it easy to iteratively generate an artifact, whether a blog post or a code snippet, it leaves a lot to be desired: Cursor's multi-file support; Artifacts' execution preview capabilities; Projects' global context. The new interface streamlines existing functionality rather than creating anything new and the underlying model doesn't best its competitors.
Canvas is a pull towards Centralization – just not a big enough one.
📰 Your AI News Brief
Stay updated with the latest developments in artificial intelligence.
Anthropic CEO’s must-read essay envisions an AI-driven future, balancing bold predictions with strategies for managing risks and equitable implementation.
Anthropic's Contextual Retrieval boosts RAG by preserving context, promising scalability and integration in diverse enterprises.
Zyphra's Zamba2-7B outperforms top models, offering efficiency and scalability, but companies must optimize infrastructure to harness its full potential.