Evaluating Early-Stage Ventures

When someone pitches "real-time AI voice translation with zero latency":

Don't code. Ask:

• End-to-end latency budget humans tolerate? (~150-200ms)
• Where does latency accumulate? (capture → encode → network → inference → decode → playback)
• Which parts scale with model size vs hardware vs physics?

If they claim "zero latency" without edge compute or predictive buffering, this violates physics. Dead on arrival.

Phase 1: Technical Sniff Test (5 minutes)

• Identify the fundamental constraint (physics, information theory, economics)
• Map where complexity accumulates in their solution
• Check if they're solving a constraint or working around it
• Flag impossible claims vs. questionable assumptions

Phase 2: Problem vs Solution Framing (10 minutes)

• Extract the organizational pain, not just the capability claim
• Distinguish structural pain from preferences/complaints
• Test: Does this solve a bottleneck or add a nice-to-have?
• Verify: Can the problem compound if unsolved?

Phase 3: Second-Order Effects (15 minutes)

• Map first-order benefits
• Identify what changes when those benefits arrive
• Find where moats might shift or erode
• Test durability against competitive responses

Phase 4: Anti-Consensus Check (10 minutes)

• Identify the prevailing wisdom
• Find the artifact trail that contradicts it
• Distinguish early/wrong from early/right
• Assess timing vs. market readiness

Phase 5: Human Judgment Under Stress (20 minutes)

• Test founder response to being wrong
• Check artifact quality vs. presentation quality
• Look for one ugly assumption they acknowledge
• Verify execution evidence over charisma signals

Example 1: False Technical Confidence Input: "We've solved the cold start problem for recommendation engines" Output: "What's your precision@k for users with <5 interactions? How do you handle taste evolution? If you need collaborative filtering anyway, what exactly did you solve?"

Example 2: Problem-Space Clarity Input: "We autocomplete code better than GitHub Copilot" Better: "Senior engineers are bottlenecks; juniors can't unblock themselves" Analysis: First sells capability (competitive). Second sells organizational pain (compounds).

Example 3: Second-Order Thinking Input: "Cloud costs dropped, so our GPU-intensive product is now viable" Question: "If costs dropped for you, they dropped for incumbents too. What prevents them from copying your features cheaply now?"

Compress Technical Assessment

• Don't need to code to understand fundamental limits
• Focus on constraint analysis over implementation details
• Trust artifact trails over demo performance

Follow the Pain, Not the Solution

• Organizational bottlenecks > technical capabilities
• Structural problems > loud complaints
• Compounding issues > one-time fixes

Model the System, Not Just the Startup

• How does success change the game?
• What do competitors do when this works?
• Where do moats migrate under pressure?

Test Stress Response

• How do founders handle being wrong?
• What's the ugliest part they'll admit?
• Do they show work or just results?

Overconfidence in Pattern Matching

• "I've seen this before" ≠ "this can't work"
• Constraints can relax silently (regulation, hardware, cost)
• Sometimes the impossible becomes possible when one assumption changes

Mistaking Noise for Signal

• Twitter rage ≠ market demand
• Loud complaints ≠ structural pain
• Clean pitches are often suspicious (real opportunities have ugly parts)

Over-Modeling

• Perfect causal chains that predict nothing
• Second-order effects can reverse unexpectedly
• At some point, bet before the model converges

Charisma Pattern Matching

• Some founders rehearse the role without the substance
• Artifacts > vibes
• Scattered conversation + strong execution history = potential false negative

False Positive Traps

• Fast growth driven by temporary regulatory loopholes
• Product-market fit that depends on unsustainable unit economics
• Technical elegance without market timing

Intervention Damage

• Giving advice when founders need to discover their own signal
• Contaminating their learning with your assumptions
• Sometimes the highest-skill move is strategic silence