Application Over Training
The Strategic Shift from Model Development to Product Development
The Shift
“Application Over Training” describes a fundamental strategic pivot in the AI industry: the recognition that raw model capability is becoming less important than what you build on top of models.
This represents a departure from the “scale is all you need” narrative that dominated 2022-2024, where companies competed primarily on model size, benchmark scores, and training compute.
Key Drivers
1. Model Commoditization
Frontier models from OpenAI, Google, Anthropic, and open-source providers have achieved rough capability parity. When GPT-4, Gemini, and Claude can all handle most tasks comparably, the model itself becomes less of a competitive moat.
2. Diminishing Returns on Training
The era of dramatic, across-the-board improvements with each model generation appears to be ending. Improvements are now incremental and domain-specific rather than revolutionary.
3. No Clear Path to AGI
The “straight shot to AGI” narrative has lost credibility. Companies that bet everything on achieving AGI first are pivoting to more practical, near-term value creation.
4. Enterprise Demand
The $37.5B enterprise AI market (projected for 2026) doesn’t primarily care about benchmark scores - it cares about whether AI solves real business problems reliably.
Who’s Saying This
Sam Altman (OpenAI):
“It is not a training problem. It is an application problem. It’s not about the model’s intelligence. It’s about building the applications to get the most intelligence out of them.”
Alex Kantrowitz (Big Technology):
“It absolutely isn’t who has the better model right now. What matters is what you do with that model and how you distribute it.”
Ranjan Roy (Writer.com):
“For a long time, I’ve been on team product. And it’s nice that Sam is finally coming around.”
Implications
For AI Companies
- Build products, not just models - Invest in UX, integration, and domain expertise
- Enterprise sales capability - Build the infrastructure to sell to businesses
- Vertical focus - Deep expertise in specific industries may beat general capability
For Enterprises
- Less model lock-in - Multiple viable models means more negotiating power
- Focus on use cases - Prioritize solving real problems over chasing the “best” model
- Build vs. buy - With commoditized models, custom applications become more feasible
For Researchers
- Application research matters - How to effectively use models is as important as improving them
- Domain expertise valued - Understanding healthcare, legal, finance beats pure ML expertise
- User research - Understanding what people actually need from AI
Timeline
| Date | Event |
|---|---|
| 2024 Q2 | Google Gemini achieves GPT-4 parity |
| 2024 Q3 | Open-source models approach frontier capability |
| 2025 Q4 | Altman declares “application problem, not training problem” |
| 2026 | Enterprise AI projected to hit $37.5B |
Related Reading
- Enterprise AI - Where the application focus is most evident
- Model Commoditization - The dynamic enabling this shift
- Sam Altman - The most prominent voice for this shift
Expert Mentions
Sam Altman
It is not a training problem. It is an application problem. It's not about the model's intelligence. It's about building the applications to get the most intelligence out of them.
Ranjan Roy
For a long time, I've been on team product. And it's nice that Sam is finally coming around.
Alex Kantrowitz
It absolutely isn't who has the better model right now. I think right now what matters is what you do with that model and how you distribute it.
Sam Altman
The models will get good everywhere, but a lot of the reasons that people use a product, consumer or enterprise, have much more to do than just with the model.
Sam Altman
Bolting AI onto the existing way of doing things, I don't think is going to work as well as redesigning stuff in the AI first world.