Future & Ethics
Forward-looking pieces — where the field is heading, what's getting harder, and the alignment / safety / hallucination problems that aren't going away.
Generative AI is moving fast enough that any "future" page is wrong within a year. These writeups try to be wrong slowly: they name the trends with the most inertia (agents, multimodality, on-device, compute as the gating factor), the open problems that won't resolve in the next 18 months (hallucinations, alignment, evaluation), and the practical responses that work today regardless of where the frontier ends up.
Ethics here isn't a section appended at the end — it's the failure modes that turn a research artifact into a liability. Hallucinations and alignment are *engineering* problems, not philosophy.
Key concepts
- Capability and reliability are different curves — the second one is what determines product viability
- Hallucinations are intrinsic to next-token training — reducing them requires architectural or post-training changes, not just better prompts
- Alignment is graded, not solved — every model release is a new local optimum on the helpful / harmless / honest triangle
- The compute frontier is a real wall — frontier-training costs are doubling roughly every six months
- On-device inference is the under-appreciated bet — it changes both economics and capability ceiling
Reference template
// Reflection-page H2 structure
## Summary
## What's changing
## Open problems
## Risks and mitigations
## What to watch
## Related materialAdapt to your problem; the structure is the load-bearing part.
Common pitfalls
- Predicting capability based on extrapolation — non-linear capability jumps will keep surprising even careful forecasters
- Treating safety as a marketing concern — the failure modes are real, measurable, and getting more consequential
- Underweighting the cost shape — most "AGI" predictions ignore that frontier compute is finite
- Reading hype as research — the gap between demo and product is wider in AI than almost any other field
Related topics
Items (4)
- The Future of Generative AI
Where the field is heading in 2026: agents, reasoning, on-device, multimodality, and the compute wall everyone is staring at.
Reflection Foundational - The Way Forward
What to learn next, in what order, and how to keep up when the field reinvents itself every six months.
Reflection Foundational - AI Safety and Alignment
RLHF, constitutional AI, red-teaming, refusal training. The engineering practices behind not-shipping-something-harmful.
Reflection Intermediate - Hallucinations and the Evaluation Problem
Why models confidently make things up, what causes it, what reduces it, and how to measure progress on a moving target.
Reflection Intermediate