What’s up for today?

> A new technique to 'trace' LLM 'circuits' ala Anthropic

> Graph theory, the human brain and what you can learn from it

Read time: 3 mins

Circuit tracing - graphs for interpretability

By now, you've probably heard about Anthropic's two very new and very interesting papers (1 and 2) that examine interpretability of the Claude 3.5 Haiku models. Their findings are of course, very interesting, but their cross-disciplinary inspiration even more so.

The what?

In plain English: Anthropic has created a way to look inside their AI's "thought process" and map out exactly how it gets from your question to its answer - like seeing the step-by-step reasoning behind Claude's responses.

They developed a new technique called 'circuit tracing'; a method to reverse-engineer how Claude 3.5 Haiku processes information. They visualize these circuits as graphs that one can interact with i.e., nodes and edges mapping from prompt to output token through the model's various layers i.e., interpretability visualized!

The so what?

For technical founders, the techniques used in paper 1 are a blueprint to:

• Debug and pinpoint why models hallucinate or make errors by analyzing failure pathways. Working on critical applications in healthcare, finance, or customer support? This could be a game-changer when you need high reliability.
• Identify redundant circuits for pruning.
• Demonstrate model transparency to regulators using causal attribution graphs.

For SR&ED-focused teams:

• If you're working in a related space i.e., creating novel methods to “measure and modify AI cognition” - this is a great example that aligns with SR&ED’s “technological uncertainty” criteria.
• IP potential: Anthropic’s architecture for their interpretability tooling can become defensible infrastructure when it's time to file a patent. So can yours.

But!

The results in the papers carry important caveats...I'll save the details for another day, but here they are briefly.

​

​

Because inspiration can strike anywhere

The human brain is an intricate network of regions that specialize in different activities, making linear analysis very tricky. In the mid-2000s graph theory applications to the human brain took off (this paper summarizes key concepts well).

The what?

Brain regions are like nodes and neurons firing together are like edges. Watching how the graph changes when you say, draw a picture or go to sleep reveals a lot about your health, how you think, and how you develop over time!

The so what?

Use it in your own innovation

• Think like a network: Map your customer interactions with product or individuals e.g., identify key influencers (hubs in graph theory)
• Embrace interdisciplinary tools: Graph theory brought math and neuroscience together. What seemingly unconnected field can your product benefit from?
• Look for patterns: Look for “small-world” or hub-like structures in your processes or systems—these often indicate areas of high impact or efficiency.

​

How was this edition of Product for Founders?