There were plenty of releases, demos, and announcements. Most of them blurred together. Models got slightly better. Interfaces got slightly cleaner. Costs came down in uneven ways. A lot of energy went into packaging rather than changing how work actually gets done.

The biggest shift wasn’t a single model or feature. It was how people used the tools. Some treated AI as a shortcut and got shallow results. Others spent time setting up workflows, guardrails, and habits. The gap between those two groups widened over the year.

What didn’t really happen:
Fully reliable agents. Clean handoffs between tools. Evaluation that most teams trust. Clear answers on safety beyond policy pages. Those problems are still open.

Looking ahead to next year, I’m watching a few things closely.

First, AI moving deeper into existing software instead of launching as standalone products. Less “new app,” more quiet replacement of steps people already do.

Second, more attention on evaluation and failure modes. Not because of public pressure, but because teams shipping real systems can’t avoid it anymore.

Third, a slowdown in novelty. Fewer surprises. More follow-through.

For readers trying to position themselves:
Spend less time chasing updates. Spend more time building something small that you can run every week. If a tool doesn’t fit into that loop, it probably won’t matter to you for long.

That’s my read going into next year.

ChatGPT introduced a feature that shows eligible users a summary of how they used the product over the past year. The recap includes usage patterns and activity breakdowns. Availability is limited to certain regions and accounts.

AI is shifting from generic assistant → personal system.
The more context it has on how you think and work, the more leverage you get.

ChatGPT now lets users adjust enthusiasm, warmth, and emoji use.

The settings apply across conversations and can be changed from the personalisation menu.

OpenAI posted a role focused on capability evaluations, threat modeling, and mitigation work for frontier models, the role is called "Head of Preparedness”

This signals that evaluation and safety work is being tied directly to shipping decisions.

Cursor acquired Graphite, a company that builds an internal developer portal focused on performance and workflow visibility. Financial terms were not disclosed.

Takeaway: This continues the move toward bundling tools and workflow into one surface.

Alibaba’s Qwen team released a new image generation model that allows users to edit specific layers of an image without regenerating the entire output.

Generation without editing is starting to feel incomplete, people want to edit the image outputs just like they edit text output by prompting again and again.

IBM released a video describing its expectations for AI development, including work on multi-agent systems, on-device reasoning, robotics, and quantum-related research. No timelines were provided.