Aanchal
Mittal
Notes on AI, LLMs, and how the field is actually changing.
agentic AI engineer @ Nokia. yes i talk to LLMs more than most people. no i'm not worried about it. this site is where the thoughts land before they get away.
Everything, newest first
7 items-
The bugs were old. The clock is new.
I went back through four months of AI security incidents, from Black October's PHP2Shell to OpenAI's 100+ notification letters to an Indian certificate authority whose private key was the number 4. Almost none of it is a new kind of hole. It's old holes on a much faster clock, and the industries holding the most sensitive data are the slowest to move.
-
What happens after you hit send
An AI model reads your whole prompt in one fast pass, then writes its answer one small piece at a time. If you're the one putting AI into a real product, that split explains the latency complaints, the bill, and where the ROI actually lives.
-
Context engineering · the context pile
Build an agent, race it against a typical one on the same bug, and see why the model re-reading its whole memory every turn is the thing that actually matters.
-
The AI panic is real. The timing is a little sus.
Three frontier labs, one three-week window, and a lot of very detailed safety reports. Most of the incidents are boring identity and config failures wearing an AI costume. The Hugging Face one isn't. And FelonyBench keeps score.
-
Reasoning off the page: what looped transformers actually change (and what they don't)
Everyone's suddenly worried recurrent depth is hiding reasoning inside GPT-6 Astra. I spent a week reading the papers. The concern is real for the trajectory, but the 'we've regressed to black box AI' framing overstates the evidence.
-
The agentic AI stack
Three planes, nine layers. Data (substrate), Control (reasoning), Governance (oversight). Click a layer to see why it matters and what it directly touches.
-
Looped transformer inference
How a recurrent-depth transformer runs the same block R times per output token. The loop, the halt gate, and where the reasoning actually happens.