Preview: This week, we unpack what self-modifying systems and better image pipelines reveal about where AI is really heading.

Editor’s note: This is the last issue for 2025. I’m taking a short break over the holidays and will be back in your inbox on January 8. Thanks for reading, thinking, and experimenting along the way. See you in the new year.

AI in Action

When AI can edit itself [Research/Governance]

“AI rewrites its own code” makes for a great headline. It also skips the most important part: control. Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory recently demonstrated a research prototype that can automatically test small changes to how it executes tasks and keep only what improves performance under human-defined criteria. The system does not invent new goals, redesign its architecture, or decide what success looks like. Humans still define the objectives and the rules for evaluation. What is new is the feedback loop, a tightly bounded way for the system to propose limited execution-level tweaks, test them, and retain only what works better.

So why does this matter outside a research lab? It points to a different kind of AI progress, less about sudden intelligence leaps and more about software that becomes more reliable over time. Systems like this could reduce repeated mistakes, adapt faster to edge cases, and improve between major updates rather than waiting for manual fixes. That shift is emerging now as AI moves from experiments into real workflows, from chat demos into tools people rely on every day. However, the same mechanism introduces a new constraint. If systems can propose changes quickly, the hardest problem becomes deciding which changes are safe, correct, and worth keeping. Evaluation becomes the bottleneck, and measuring outcomes clearly starts to matter more than generating new ideas.

That is why the real implication of this research is governance by design. Even limited self-modification changes what needs to be controlled. The upside for everyday users would be quieter technology: fewer brittle AI moments and tools that become more predictable over time. The risk is quieter too. Change can happen without being obvious, making it harder to understand what shifted, why it shifted, and who is accountable when something goes wrong. As AI becomes more adaptive, trust will depend less on raw intelligence and more on how change is measured, constrained, and controlled.

Bits of Brilliance

Why does AI images suddenly look real? [AI application]

If AI images suddenly feel better to you lately, you’re not imagining it. Try ChatGPT’s new image function or tools like Nano Banana and NanoBanana Pro, and the difference is hard to miss. Images look more realistic, more consistent, and noticeably faster to generate. It’s tempting to assume the models suddenly became more creative. However, the real shift is more practical. Today’s image systems improved because they learned to build pictures in a smarter order. Older approaches tried to adjust millions of tiny dots all at once, which made them slow and fragile. That’s why the same problems kept showing up: strange hands, uncanny faces, lighting that didn’t quite add up. So what actually changed?

Newer systems now start with a simplified internal rough draft of the image. They focus first on the big structure, composition, shapes, and lighting, before filling in texture and fine details. This approach is faster and more stable, which helps explain why NanoBanana Pro feels quick and why ChatGPT’s images tend to follow detailed prompts more closely. A lot of progress happened quietly under the hood, but once these systems crossed a reliability threshold, users felt it all at once. The images didn’t just improve. They stopped falling apart.

That last point matters more than it sounds. AI images feel better today not because they are wildly more imaginative, but because they fail less and stay consistent from one attempt to the next. Reliability is what turns image generation from a novelty into something people can actually use for real work, whether that’s a slide graphic, a marketing mockup, or a quick concept sketch. And the lesson goes beyond images. In AI, progress often doesn’t look like something new. It looks like something that finally works.

Curiosity in Clicks

Which AI image generator do you prefer? [Experiment]

Go into “image” in ChatGPT and Nano Banana / NanoBanana Pro, put in the same prompt below:

“A natural-light portrait photo of a girl sitting by a window, soft shadows, realistic skin texture, subtle imperfections, shallow depth of field, candid expression, no stylization, no filters, with christmas decor in the background.“

Compare the results. Look closely at faces, hands, lighting, and small details. How ling did it take to generate this image? Which image feels more natural? Which do you like more? Mine are below. Guess which ones from Gemini, and which one’s from ChatGPT?

_{Byte-Sized Intelligence is a personal newsletter created for educational and informational purposes only. The content reflects the personal views of the author and does not represent the opinions of any employer or affiliated organization. This publication does not offer financial, investment, legal, or professional advice. Any references to tools, technologies, or companies are for illustrative purposes only and do not constitute endorsements. Readers should independently verify any information before acting on it. All AI-generated content or tool usage should be approached critically. Always apply human judgment and discretion when using or interpreting AI outputs.}