Most writing about the future of foundation models is either utopian or apocalyptic, and both are useless for making decisions. The useful version is narrower: what do the signals visible today suggest about the next few years, and what should you do differently because of them. This article takes that approach. It is a thesis, not a prophecy, and it is grounded in trends you can already see rather than in speculation about artificial general intelligence.
The honest framing up front: nobody knows the long-term trajectory, and anyone who claims certainty is selling something. But the near-term direction is more legible than the discourse suggests, because it follows from current economics, current limitations, and the way the field has actually moved. Below are the shifts I am most confident about, the trade-offs they create, and what they mean for how you build and what skills hold value. For the fundamentals these predictions build on, The Complete Guide to Foundation Models is the reference.
Signal 1: Capability Gains Are Shifting From Raw Scale to Systems
For several years, the dominant story was that bigger models were better models, and progress meant more parameters and more training data. That curve has not stopped, but its slope has changed, and the marginal gains from pure scale have gotten more expensive to buy.
The observable shift is toward systems around the model rather than just larger models: retrieval that grounds output in real data, tool use that lets the model act rather than only generate, and orchestration that breaks hard problems into verifiable steps. The thesis: the next wave of practical capability comes more from how models are wired into systems than from the next size jump. This is good news for builders, because system design is something you control, unlike the size of the underlying model.
Signal 2: Small Models Are Quietly Winning Production
The headlines go to the largest flagship models, but the production reality is moving the other way. Smaller models keep getting more capable, and a small model that is fast and cheap enough to deploy at scale beats a large model that is too slow or expensive to use widely.
This creates a durable pattern: a tier of small, specialized models handling the bulk of routine work, with large models reserved for the genuinely hard cases. The economic logic is hard to fight — most production traffic does not need the biggest model, and routing it there wastes money and latency. Expect tooling and practice to increasingly assume this tiered, route-by-difficulty world. The selection logic for living in it is in A Framework for Foundation Models.
Signal 3: Agentic Systems Will Be Real and Overhyped at the Same Time
"Agents" — models that take actions, use tools, and pursue multi-step goals — are the current frontier and the current hype magnet. Both things are true: they are genuinely useful for some workflows and badly oversold for most.
Where they will work
Agentic systems work when the task decomposes into verifiable steps, the actions are reversible or low-stakes, and there is a way to check the result. Structured workflows with clear checkpoints — research, data gathering, code with tests — are fertile ground.
Where they will disappoint
They struggle when errors compound across steps, when the task requires judgment the model does not have, and when the cost of a wrong action is high. The naive vision of an agent autonomously running a complex business process unsupervised will keep disappointing, because a small per-step error rate compounds into frequent overall failure across many steps. The realistic future is agentic systems with humans at the checkpoints, not without them. The over-reliance risk this creates is covered in The Hidden Risks of Foundation Models (and How to Manage Them).
Signal 4: The Commodity and the Moat Are Separating
As multiple providers ship capable models, the raw model is becoming more of a commodity. What is not a commodity: the data you ground it on, the workflows you build around it, the evaluation that tells you it works, and the domain expertise that makes the output trustworthy.
The thesis for organizations is that competitive advantage migrates away from "we have access to a good model" — everyone will — toward "we have wired a good model into our specific problem better than anyone else." That favors teams who invest in the surrounding system and their own proprietary data over teams who treat the model as the whole product. For individuals, it favors the applied skills covered in Foundation Models as a Career Skill: Why It Matters and How to Build It.
Signal 5: Reliability and Governance Move From Optional to Required
Early foundation-model adoption tolerated a lot of duct tape because the stakes were low and the novelty was high. That tolerance is shrinking. As these systems handle more consequential work, the expectations around reliability, evaluation, data handling, and governance are hardening.
The practical implication: the practices that are currently "nice to have" — regression evals, monitoring, data governance, human review on high-stakes output — become table stakes. Teams that built these in early will look prescient; teams that skipped them will spend the next phase retrofitting under pressure. The playbook for getting ahead of this is in Foundation Models: Best Practices That Actually Work.
What This Means for You
A grounded view of the future should change what you do now. Three concrete implications:
- Invest in systems and evaluation, not in chasing the latest model. The model you build on will change; the surrounding system and your eval discipline are the durable assets.
- Design for a tiered, route-by-difficulty world. Architect so you can swap models and route traffic by difficulty, because that is where production is heading.
- Build the governance now while it is cheap. Evals, monitoring, and data handling are far cheaper to build in early than to retrofit once the stakes rise.
The future of foundation models is less about a dramatic leap and more about these capabilities becoming embedded, governed, and economically optimized infrastructure. The teams that treat it that way will be ready; the ones waiting for a magic model will keep being surprised.
Frequently Asked Questions
Will foundation models keep getting dramatically better?
They will keep improving, but the near-term gains are shifting from raw scale toward how models are wired into systems. Expect steady capability growth plus large practical gains from retrieval, tools, and orchestration rather than only from bigger models.
Are agents going to automate everything soon?
No. Agentic systems are genuinely useful where tasks decompose into verifiable steps with reversible actions, and they disappoint where errors compound or judgment is required. The realistic future keeps humans at the checkpoints rather than removing them.
Should I bet on big models or small ones?
Both, in tiers. Small models will handle the bulk of routine production work because of cost and latency, with large models reserved for hard cases. Architect so you can route by difficulty rather than committing to one size.
What becomes a competitive advantage if models are commodities?
Your proprietary data, the workflows you build, your evaluation discipline, and your domain expertise. As model access commoditizes, the moat moves to how well you wire a capable model into your specific problem.
What should I build now to be ready?
Evaluation, monitoring, data governance, and provider-abstracted architecture. These are cheap to build early and expensive to retrofit, and they are exactly what becomes mandatory as the stakes rise.
Key Takeaways
- Near-term progress is shifting from raw model scale to systems built around models.
- Small, cheap models are winning production; design for a tiered, route-by-difficulty world.
- Agentic systems are real where tasks are verifiable and oversold where errors compound; keep humans at the checkpoints.
- As models commoditize, advantage moves to proprietary data, workflows, evaluation, and domain expertise.
- Build governance and evaluation now while it is cheap, because it becomes table stakes as stakes rise.