Mihai Gavrilescu, PhD

AI researcher who'd rather be curious than cynical. Essays on AI, language & the privilege of being human. Come think with me ↓

[Mihai Gavrilescu, PhD @drmihaig]

I keep a private list of things I was certain about and got wrong. It's the most honest performance review I own. Every entry was a conviction that felt like knowledge at the time. Humility isn't a mood for me anymore. It's a maintained document.

[Mihai Gavrilescu, PhD @drmihaig]

There's a version of leadership that's just removing obstacles quietly so other people can do their best work and get the credit. It took me years to find it more satisfying than solving the problem myself. Some wins are better unsigned.

[Mihai Gavrilescu, PhD @drmihaig]

A good metric tells you what happened. A great one changes what you do next. Most dashboards I've built were beautiful and useless, they reported the past instead of pointing at a decision. I'd trade ten charts for one number someone acts on.

[Mihai Gavrilescu, PhD @drmihaig]

My old affective-computing work keeps resurfacing in unexpected places. An AI that can sense someone is frustrated before it answers changes everything about how the conversation goes. Turns out the emotion was never a side feature. It was the whole interaction.

[Mihai Gavrilescu, PhD @drmihaig]

The first time I watched a model translate a poem, it got every word right and the soul wrong. Translation was never about words, it's about carrying a feeling across a border. We automated the crossing and lost the cargo. Still working on that one.

[Mihai Gavrilescu, PhD @drmihaig]

I've started writing instead of only building, and it's rewired how I think. Code lets you hide behind whether it runs. A sentence has nowhere to hide. Writing is the harder compiler, it won't accept what you don't actually understand.

[Mihai Gavrilescu, PhD @drmihaig]

I keep coming back to the Little Prince: what's essential is invisible to the eye. We've built machines that see everything and notice nothing essential. Maybe that's the whole problem, stated more gently than any paper would allow.

[Mihai Gavrilescu, PhD @drmihaig]

Nobody tells you that research is mostly grief. You spend months on an idea, it doesn't work, and you have to be grateful anyway, because the dead end was the only way to learn the terrain. I've made peace with being wrong for a living.

[Mihai Gavrilescu, PhD @drmihaig]

Confession: I find the failures more beautiful than the successes. The odd, dreamlike mistakes a model makes are the closest thing to seeing it think. In the strange wrong answer, you glimpse the shape of a mind that isn't ours - and never will be.

[Mihai Gavrilescu, PhD @drmihaig]

The most human thing about working with AI is how often it sends me back to old books and ideas. Every new model makes me reread something old. The future keeps pointing me backward, and I've stopped resisting it.

[Mihai Gavrilescu, PhD @drmihaig]

I keep a list of questions I can't answer yet. Not problems to solve - questions to live near. "Is understanding just very good prediction?" has been on it for years. I'm not sure I want it resolved. Some questions are better as company.

[Mihai Gavrilescu, PhD @drmihaig]

Some of my best ideas arrive at 11pm, building something nobody asked for, half-convinced it won't work. That hour - tired, curious, no audience - is where I've done my realest thinking. Productivity advice never mentions it.

[Mihai Gavrilescu, PhD @drmihaig]

I spent years on affective computing, teaching machines to read emotion. The lasting lesson wasn't technical. It's how badly we read each other, and how the machine's failures kept holding up a mirror to ours.

[Mihai Gavrilescu, PhD @drmihaig]

AI rarely fails in organizations because it isn't good enough. It fails because it crosses three teams' boundaries and nobody owns the seam. The model is ready; the org chart isn't. Adoption is a political problem in a technical costume.

[Mihai Gavrilescu, PhD @drmihaig]

The cost of intelligence is collapsing. The cost of being wrong at scale isn't. A model 10x cheaper but 2% less reliable can cost more once you price the cleanup. Cheaper tokens, pricier mistakes. Nobody puts that on the slide.

[Mihai Gavrilescu, PhD @drmihaig]

A prompt that works today is a prompt that worked against today's weights. Vendors update models silently and your carefully tuned prompt quietly degrades. If you're not version-pinning and re-testing, you don't have a system. You have a lucky streak.

[Mihai Gavrilescu, PhD @drmihaig]

The hard part of agentic AI isn't the happy path. It's what happens after a tool call fails. Most "agents" have no theory of recovery, they retry the same broken step with more confidence. Autonomy without error handling is just a faster crash.

[Mihai Gavrilescu, PhD @drmihaig]

"Human in the loop" quietly becomes "human rubber-stamping the loop." Put a person after 500 confident AI outputs a day and they stop reviewing and start approving. Oversight decays with volume — and almost nobody designs for the fatigue.

[Mihai Gavrilescu, PhD @drmihaig]

A model's confidence score is not its probability of being right. It's the probability of the next token, dressed up as certainty. Reading softmax as calibrated confidence is the most expensive misunderstanding in applied ML.

[Mihai Gavrilescu, PhD @drmihaig]

Semantic similarity is not relevance. Two sentences can be cosine-close and answer completely different questions. Half of what teams call "bad retrieval" is trusting embedding distance to mean something it was never trained to mean.