Do AI companies work?
Venture-backed AI labs racing to build ever-larger language models face mounting skepticism over whether their economics are sustainable or defensible. Commenters note that training costs are enormous, models are quickly commoditized by open-source and cheaper “good enough” alternatives, and clear moats beyond brand, data access, and integration remain elusive. Many compare the moment to the dot-com or rideshare eras: real technological value likely exists, but most current players may burn cash chasing scale while eventual winners emerge in infrastructure, specialized applications, or adjacent sectors like energy.
VC Hype, Historical Parallels, and Bubble Dynamics
- Many compare the current AI wave to ride‑sharing/food‑delivery booms: massive VC subsidy, goodies for users, then consolidation and “enshittification” plus destruction of incumbents.
- Some argue this “slime mold” style capital allocation is useful exploration; others say it’s toxic, creates unsustainable competitors, and leaves users worse off once subsidies end.
- Several suggest founders and VCs mostly aim for hype-driven exits, not durable businesses.
Business Models, Moats, and Commoditization
- Core worry: LLMs are becoming commodities. Models are interchangeable, input is just text, and switching vendors can be relatively cheap at the API level.
- Proposed moats:
- Process complexity and accumulated research/software, analogous to search engines’ ranking systems.
- User data and feedback loops for continuous improvement.
- High integration/switching costs in real deployments.
- Brand, UX, ecosystem, and enterprise relationships.
- Others counter that inference is cheap, open models are “good enough,” and price pressure will push LLMs toward utility‑like margins.
Open Source vs Frontier Models
- Open models (Llama, etc.) are seen as rapidly catching up, often at much smaller sizes, especially when combined with fine‑tuning, LoRAs, and “activation engineering.”
- View A: this keeps big spenders perpetually within ~6–18 months of being cloned, undermining multi‑billion‑dollar moats.
- View B: over time, frontier labs’ private research codebases and data access will form a barrier that late entrants can’t quickly cross.
AGI, Superintelligence, and Skepticism
- Enthusiasts claim we’re near an “AGI landslide,” driven by scaling, national‑security pressure, and potential recursive self‑improvement.
- Skeptics see current systems as “crappy chatbots” or sophisticated autocomplete: impressive but far from human‑like, still brittle, bad at long‑horizon tasks, math, and grounding.
- There’s disagreement over whether LLMs are on the right path to AGI or a powerful but local maximum.
Use Cases, ROI, and Practical Limits
- Strongest consensus value today:
- Coding assistants and developer tools.
- Customer support, drafting emails, summarization, translation, and knowledge retrieval.
- ROI is questioned: many see near‑zero or modest gains, especially in customer service and generic chatbots, relative to enormous capex.
- Several note the “bottleneck” is not smarter models but product design: getting AI to reliably do real work given human communication limits and verification needs.
UX, Branding, and Differentiation
- Multiple comments argue the real competition will be on UX, personality, integrations, and vertical solutions, not raw model quality.
- Current text‑box interfaces and prompt fiddling are seen as hostile to mainstream users; there’s a call for stronger product, design, and brand thinking to build lasting user loyalty.