What's Coming Next

What's Coming Next

We're in a period of rapid, compounding progress in AI. Predicting exactly what comes next is impossible — the field has repeatedly surprised even its leading researchers. But there are clear trajectories worth understanding, both to prepare for what's coming and to think critically about claims you'll encounter.

What's Happened Since 2024

Multimodality is now standard: frontier models natively handle text, images, audio, and video as inputs and, increasingly, outputs. What was "coming soon" in 2024 is now the baseline expectation.

Context windows reached ~1M tokens: what counted as frontier in 2024 is now standard across major providers, paired with better memory systems so AI can maintain context across longer sessions and more of a "memory" of prior work.

Reasoning models became the default: rather than a separate experimental tier, models that spend extra "thinking" time before answering — with adjustable reasoning effort — are now the norm across frontier labs, not a niche add-on.

Agents moved from demos to production: companies now routinely deploy AI agents for real-world tasks — coding, research, customer service, data analysis. The patterns are established, and reliability keeps improving.

Open-weight models closed much of the gap: models from labs including DeepSeek, Qwen, and others now compete closely with frontier closed models on many benchmarks. Running capable AI on-device is increasingly mainstream.

Genuinely Forward-Looking Trends

Scientific acceleration: AI that can genuinely assist with scientific research — hypothesis generation, literature synthesis, experimental design, data analysis. There's early evidence in biology, chemistry, and materials science; the open question is whether this extends to genuinely novel discoveries rather than accelerating existing workflows.

Robotics integration: language model reasoning combined with physical robots. The cognitive flexibility of LLMs + physical manipulation capabilities. Still hard, but improving.

Long-horizon autonomy: agents that can reliably pursue complex, multi-day goals with minimal supervision, rather than today's shorter, closely-watched tasks.

Economic and societal integration: AI becomes infrastructure rather than a feature — embedded in workflows, decision processes, and systems as thoroughly as databases and networks are today.

How to Think About AI Progress

A few principles for navigating claims about what AI can and will do:

Distinguish demos from deployment. Impressive demos appear regularly. Reliable deployment in production is harder, slower, and less glamorous.

Watch the benchmarks, but understand their limits. AI benchmarks get saturated quickly. New benchmarks get created. Benchmark performance and real-world usefulness are related but not identical.

Scepticism and engagement both matter. Dismissing AI progress as hype leads to being caught off guard. Uncritically believing every capability claim leads to poor decisions. The right position is informed engagement.

Your most valuable response is to build and use. The people who understand AI best are those who build with it and use it regularly — encountering real limitations, discovering real capabilities, developing genuine intuitions.

Dysgu exists to help you build those intuitions. You've covered a lot of ground. The next step is applying it.