Summary: A technical edge in AI is a “business card”, not a moat: once users touch it, competitors have about six months to catch up, in the good case. Selling tokens with public benchmarks invites endless one-upping and one-click switching (OpenRouter is the worst case). The durable path is a “good enough” product whose value lives beyond the specific model, like Spotify knowing your music, or a car that fits your family rather than one with the highest top speed. The frontier labs are in a money-losing arms race whose trillion-dollar IPOs strain growth math.
Key ideas
- “Good enough” wins: nobody switches between Waze and Google Maps for five seconds, or buys a car for its 240 km/h top speed (explainable—164-gilad-levi-continual-learning, gilad-levi).
- Benchmarks as a product are the worst business: like Spotify publishing evals against Apple Music. Spotify instead says “we already learned you” (explainable—164-gilad-levi-continual-learning, gilad-levi).
- Model one-upmanship is tiny and weekly (Codex vs Claude Code flip-flopping), and AI products are technology-based, not market-based, so few last more than a week (explainable—164-gilad-levi-continual-learning, gilad-levi).
- Founders: if you don’t know your moat yet, say so and name it as a core problem; “everything’s fine” lowers credibility (explainable—164-gilad-levi-continual-learning, gilad-levi).
- Labs fight fires (Gemini, Anthropic, OpenAI, DeepSeek releases) and cut product lines. OpenAI said in 2023 it would try many lines and see what sticks; Anthropic cut far fewer (explainable—164-gilad-levi-continual-learning, gilad-levi).
- Lose-now-earn-later works (Lime, Sequoia-style) but has never been tried past a trillion dollars. Google IPO’d at ~2T IPO in a market roughly 4× that can’t. Public pension money then carries the bet (explainable—164-gilad-levi-continual-learning, gilad-levi).
- Big incumbents aren’t dying: Microsoft has compliance, Apple has MacBooks (~5.5T) is the backbone (explainable—164-gilad-levi-continual-learning, gilad-levi).
- Building on model vendors means competing with them: vendors move into every vertical (code, legal, design). The test is whether you create enough alpha to justify your margin on tokens (ignore-instructions—26-ai-cyber-pavel-gurevich, pavel-gurevich).
- Agent companies reaching escape velocity (coding, voice) train a fallback model of their own: better economics they can pass to customers, which frontier labs can’t match at ~$1 per million tokens (ignore-instructions—26-ai-cyber-pavel-gurevich, pavel-gurevich).
- Frontier LLMs have “zero distribution moat”: no migration cost and precise benchmarks mean users switch weekly, so labs fight brutally; even toilet paper has more stickiness. Technology that learns its user continually is sticky by design (hidden-layers—manifold-gilad-levi, gilad-levi).
- Frontier models commoditize their users: everyone using the same model gets the same results as their competitors. Differentiation needs access to the model’s guts (inject your data, fine-tune) (hidden-layers—bria-misha-feinstein, misha-feinstein).
Disagreements & open questions
- Will one frontier lab collapse? The host expects one to; Levi doubts the causality but agrees the market is moving in an unhealthy direction (explainable—164-gilad-levi-continual-learning).
Takeaways
- When building on AI, write down what value remains if a competitor swaps in a better model tomorrow (explainable—164-gilad-levi-continual-learning)
Related
ai-hype · enterprise-ai-adoption · model-selection · ai-cybersecurity · solo-builders · image-generation