Limits of AI & the future
What today's AI cannot do — not because of bugs, but because of fundamental properties of how it's built. Plus a sober look at where it's likely heading.
Limits that aren't going away soon
Properties of next-token prediction trained on human-generated text. Not bugs to be patched.
- Long-horizon planning. Multi-month plans with checkpoints and contingencies — current models drift, forget commitments, lose track of state. Reasoning models help on bounded problems, but unbounded long horizons are genuinely hard.
- True novelty. Excellent at recombining what's been written into something fresh. Much weaker at producing genuinely new ideas — theorems, theories, discoveries — that weren't latent in training data.
- Calibrated uncertainty. Models speak confidently whether or not they know. They don't have introspective access to their probability mass in a way that translates cleanly to "I don't know."
- Stable values across contexts. A model trained to be helpful/harmless/honest can be coaxed into bypassing those values with the right framing (jailbreaks). It has trained policies, not deep values.
- Real physical grounding. LLMs know physics from text about physics, not from being in a body. Robotics is closing the gap slowly.
What's improving fast
Things that look like fundamental limits but are moving rapidly: hallucinations on common facts (RAG, longer context, better grounding); math and code (reasoning models); multimodal understanding (vision becoming default); long context (4k → 200k → 2M+ in a few years); speed and cost (today's frontier is next year's "tiny model").
The pattern: things with a verifiable signal (math correctness, code correctness, retrieval accuracy) move fast. Things without one (creativity, stable values, calibrated uncertainty) move slowly.
Real-world examples
Where things are heading
Predictions about specific dates are usually wrong. Trends are more reliable:
- Reasoning becomes default, not an opt-in mode.
- Multi-agent and computer-use mature: flaky today → reliable for narrow workflows.
- Smaller models close the gap on common tasks. A laptop-sized model in 2027 may equal today's frontier on everyday work.
- On-device AI becomes ubiquitous (privacy + latency wins).
- Real-time multimodal becomes the default UI — voice + vision + text in one stream.
- Tooling stabilises around standards (MCP is an early example).
- Regulatory frameworks emerge (EU AI Act, US executive orders, Chinese rules) — divergent, real, shaping product decisions.
What you won't see soon: AGI in any rigorous sense, AI replacing all knowledge work, AI "becoming conscious." These framings sell magazines but obscure the actual progress.
If you've made it this far, you have the framework to read papers, evaluate new tools, build features that survive contact with reality, and push back on the hype. That's most of what anyone needs in 2026.
Check your understanding
- 1. Which of these is a 'fundamental' limitation of today's LLMs (not just a temporary engineering issue)?
- 2. What does the 'Bitter Lesson' argue, in one sentence?
- 3. What's the right framing for predictions about AI's future?
Found this useful? Share it with someone learning AI.
Further reading
- Rich Sutton — The Bitter Lesson (2019) — required reading; 1,000 words.
- Karpathy — On the future of LLMs — practical perspective from a leading researcher.
- Yoshua Bengio — Pause Giant AI Experiments? — careful framing of risks worth taking seriously.
- Anthropic — Core Views on AI Safety — one frontier lab's published position.
- AI Index Report (Stanford HAI) — annual data-driven snapshot of the state of AI.
- 80,000 Hours — AI safety — accessible summary of the case for caring about long-term AI risks.
You've finished the Advanced track. Congratulations. The field will keep moving — but now you have the framework to follow along, build with it, and push back on the hype.