LLenny's Podcast
← All frameworks
InnovationMike Krieger (co-founder of Instagram)

The Three-Part AI Product Utility Equation

A useful AI product needs model intelligence, context/memory, and application/UI to all converge.

Difficulty
Easy
Time to result
~weeks to results
Steps
4
Confidence
92%

Krieger's 'fake equation' for whether an AI product is actually useful breaks utility into three factors that must all be present simultaneously: raw model intelligence, the right context and memory, and the application layer plus UI. A world-class model with no context, or great context with no discoverable interface, still yields a poor product. It's a diagnostic for locating where your AI product is weakest.

Origin

Mike Krieger's self-described 'fake equation' (as the non-researcher in the room) for reasoning about AI product utility; he uses it to explain why Anthropic built MCP to attack the middle term.

Core principles

  • 01All three factors must converge — the product is only as strong as its weakest term
  • 02Model intelligence is increasingly a solved/shared commodity across labs
  • 03Context and memory are often the real bottleneck, not raw model capability
  • 04The application and UI layer determines whether integrations are discoverable and repeatable

How to run it

  1. 1

    Assess model intelligence

    Confirm the underlying model is capable enough for the task. For frontier tasks this is often already sufficient and not your constraint.

    Pro tip If competitors using the same off-the-shelf model could build your feature, intelligence is not your differentiator.

  2. 2

    Wire in the right context and memory

    Feed the model the specific documents, conversations, and data it needs rather than relying on generic web knowledge. This is usually where quality is won or lost.

    Pro tip Krieger flags this middle term as what MCP exists to solve — pulling context from Slack, Drive, and internal docs.

    Watch out Generic answers vs. good answers is 'entirely' the difference between having the right context and not.

  3. 3

    Make the application layer discoverable and composable

    Ensure integrations are discoverable, the UI is right, and users can build repeatable workflows around the capability.

    Pro tip Expose primitives as MCP endpoints so the model can act on them, not just talk about them.

  4. 4

    Diagnose the weakest term

    When a product underperforms, identify which of the three factors is failing rather than reflexively demanding a smarter model.

    Watch out Teams over-index on model intelligence and neglect context and UI, which are frequently the actual gap.

In the wild

Product strategy answer, with and without context

Asked to discuss Anthropic's product strategy, a model with only web access gives a weak generic answer; the same model given internal strategy docs plus MCP access to Slack and Google Drive gives a sharp, grounded one.

Krieger uses this to show the context/memory term alone flips a bad answer into a good one, holding model and UI constant.

Common mistakes

Treating a smarter model as the fix for every product gap

Many failures come from missing context or a poor UI layer, not insufficient intelligence — swapping in a better model won't help.

Building integrations that aren't discoverable or repeatable

Even with great context, if users can't find or reliably reuse an integration, the product's utility collapses at the application term.

Is it for you?

Best for

Product and engineering leaders building applications on top of foundation models who need to diagnose why an AI feature feels underwhelming.

Not ideal for

Pure research on model capabilities, where the application and context layers are out of scope.

From the transcript

my like fake equation for like utility of AI products uh it's three-part

54:30

One is model intelligence. The the second part is context and memory. And the third part is like applications and UI. And you need all…

54:30

the difference between like the right context and not it's like entirely the the the difference between like a good answer and a and a…

55:30

From the episode

Anthropic’s CPO on what comes next

Mike Krieger (co-founder of Instagram)