1. Training data provenance became a standard diligence requirement in AI investment primarily following:
Correct — the NYT and Getty lawsuits created the template. Investors began requiring provenance documentation and legal exposure assessments as standard diligence deliverables.
The catalyst was the copyright lawsuits — NYT v. OpenAI/Microsoft and Getty v. Stability AI — which created a legal template investors applied across their AI diligence. Not the EU AI Act, FTC, or a Sequoia memo.
2. An investor-grade network effect argument differs from a generic claim by including:
Correct.
The three required elements are: the specific mechanism (how it works), the quantified gap (how much better), and the time barrier (why a competitor can't catch up quickly).
3. OpenAI's governance crisis in November 2023 revealed that Microsoft, despite $13 billion in commitments, had no mechanism to override the board's decision to fire Sam Altman. The structural reason was:
Correct. The nonprofit-controlled capped-profit structure is why standard VC governance rights — board seats, protective provisions — didn't exist in the expected form for Microsoft.
The correct answer is OpenAI's unique non-standard structure: a nonprofit-controlled capped-profit LLC that wasn't subject to normal C-Corp investor governance mechanics.
4. Which aspect of Stability AI's investor relations practice contributed to the board conflict and CEO Emad Mostaque's March 2024 resignation?
Correct. Mostaque's public statements about Stability AI's technical roadmap consistently outpaced actual execution. Investors who believed they'd funded a specific vision encountered a different reality in board meetings — creating an irreconcilable expectations gap.
Not quite. The core issue was a public-statements expectations gap — Mostaque made ambitious public roadmap claims that didn't match execution reality. The resulting conflict between stated vision and board-level reality contributed directly to his departure.
5. Inflection AI raised $1.3 billion in June 2023 at a $4 billion valuation. By March 2024, most of its team had joined Microsoft for approximately $650 million. Regulators investigated this as a potential de facto acquisition primarily because of concerns about:
Correct. The UK CMA and EU regulators examined whether the structure evaded merger control — effective acquisition of the team and technology without filing for antitrust approval.
The correct answer relates to regulatory evasion — the talent/licensing deal effectively acquired the company's assets without triggering formal merger review requirements.
6. Which three domains does Lesson 1 identify as AI's native strengths that, when built into a problem frame, pre-empt the "why AI?" objection?
Correct. Scale, speed, and pattern complexity are AI's native domains — defining the problem in these terms makes the AI solution feel self-evident rather than imposed.
Review Lesson 1. The three domains are scale, speed, and pattern complexity — the areas where AI structurally outperforms human effort.
7. Under full-ratchet anti-dilution, a company's down round at 60% of the original valuation causes the prior investor's conversion price to:
Correct. Full ratchet is the most aggressive mechanism: the conversion price drops entirely to the new down-round price regardless of round size.
Full ratchet resets the conversion price completely to the new lower price. The weighted average description applies to broad-based weighted average anti-dilution, not full ratchet.
8. Which language pattern signals structural differentiation over current-state advantage?
Correct.
Structural language explains why an advantage persists over the investment horizon. A regulatory timeline creates a forward-looking barrier that cannot be overcome with compute budget — benchmark leads and first-mover claims erode quickly.
9. A drag-along provision that requires only "a majority of preferred shareholders" without requiring common shareholder approval creates which founder risk?
Correct. A preferred-only drag-along can force founders and employees (common shareholders) into a sale they oppose, with no common vote required.
The correct risk: without a required common shareholder vote, preferred investors can drag founders and employees into an approved sale regardless of their opposition or economic interest.
10. In AI venture investing, "benchmark overfitting" refers to which practice that prompted sophisticated investors to require third-party model evaluations?
Correct. Benchmark overfitting is fine-tuning models to excel on known public benchmarks (MMLU, HumanEval) without genuine capability gains. Investors like Coatue now require blinded third-party evaluations on unseen benchmarks as a result.
Not quite. Benchmark overfitting is when AI companies fine-tune models specifically to score well on public evaluation datasets without real underlying capability improvement — prompting investors to require independent, blinded model evaluations.
11. A technical reviewer who discovers a failure mode that was NOT disclosed in a model card will most likely conclude:
Correct. Undisclosed failure modes damage credibility not just technically but organizationally — they suggest either insufficient testing or intentional omission.
Incorrect. Undisclosed failures signal to reviewers that the founder lacks rigor or transparency — both are serious trust problems in an investment context.
12. An investor-grade data moat presentation addresses four dimensions. Which set is correct?
Correct.
The four investor dimensions are origin, scale, velocity, and materiality — each testing a different aspect of whether the data asset is genuinely defensible.
13. Cohere's competitive moat in 2022–2023 was primarily based on:
Correct.
Cohere's moat was deployment architecture — on-premises capability satisfied data sovereignty requirements for regulated enterprises that OpenAI's API-only model structurally could not serve.
14. When Drew Houston pitched Dropbox, what made his opening line effective according to the lesson?
Correct. "People lose their files" is universally felt and immediately recognisable — no market-size slide needed to convince the room this problem exists.
Houston opened with a universal, daily problem statement — no credentials, no market size, no company name first.
15. Which of the following is the most durable form of AI moat according to investor frameworks from a16z and Bessemer?
Correct. Data moats are considered the most durable because they compound — more usage generates more data, which improves the model, which attracts more usage. Model benchmarks are temporary; proprietary data is not.
Data moats are most durable. Model performance is temporary — new foundation models can close performance gaps — but proprietary data accumulated over years cannot be replicated quickly.
16. Which investor metric best captures the compounding quality of an AI SaaS customer base?
Correct. NRR captures whether existing customers are expanding their spend faster than they churn — a company with 130%+ NRR compounds revenue from its installed base independent of new customer acquisition.
NRR is the most powerful quality metric because it measures expansion within existing customers — a company with 130% NRR grows revenue 30% annually from existing accounts alone, demonstrating compounding economics independent of new sales.
17. GitHub Copilot's market position over competitors using similar models was primarily built on:
Correct.
Copilot's advantage was distribution (existing GitHub users) and workflow integration depth (habitual IDE embedding) — not model quality, pricing, or exclusivity agreements.
18. Which provision in an AI term sheet most directly determines founder proceeds in an acqui-hire exit at $150M where the liquidation preference stack is $120M?
Correct. At $150M with a $120M stack, the participating vs. non-participating distinction determines whether founders receive all $30M remaining or share it pro-rata with participating preferred investors.
The key determinant is participating vs. non-participating preferred. With $30M above the preference stack, non-participating preferred leaves all $30M for common; participating preferred splits that $30M between investors and common pro-rata.
19. The BigScience BLOOM model introduced the RAIL license. What distinguishes RAIL from Creative Commons licenses?
Correct. RAIL (Responsible AI License) governs model use — what downstream users can do with the model — rather than data licensing terms.
Incorrect. RAIL's key distinction is that it applies at the model level, restricting certain uses of the model itself regardless of training data licensing.
20. A "structural scarcity premium" in AI valuation refers to:
Correct. Structural scarcity reflects temporary supply constraints that can become permanent moats. Mistral's European frontier AI position illustrates this: no additional engineering spend could quickly produce a credible alternative European frontier AI lab.
Structural scarcity premium is the value placed on a position (geographic, regulatory, technical) that cannot be quickly reproduced by a well-resourced competitor. Mistral's status as the credible European frontier AI lab exemplifies this — the demand existed but the supply could not be quickly created.