Alessandro Bahgat
1,854
posts
@abahgat
VP Software at Quilt. Professional coach for Eng leaders. 12y
@Google. Building with AI, LLMs & open source. Still building side projects at night.
Redwood City, CA
Joined March 2008
Would be interesting to see if the net result of this is that the legibility of the typical paper were to go up correspondingly, as much as I have some reservations on the whole
Prompt: “Write a paper to publish on arXiv” 30% of academic papers appear to be made by AI. The share of new arXiv papers flagged as machine-written, 2021 to 2026, at a threshold calibrated so pre-ChatGPT papers flag at 0.4%.
Built this Scope 3 prototype (Scopewise) over the weekend for a distributed #EarthDay hackathon nudging agents from my phone between family activities. If one parent with fragments of weekend time can get here, what's a full sustainability team's excuse for not even prototyping?
I spent a few hours over the last couple of weekends getting local AI models to play Zork, a 40-years old text adventure. On paper, another toy project. In practice, a great lesson on agent evals.
You know it's 2026 when building a B2C app takes less time than summarizing what you built in Google Slides.
I’ve been working with MCP recently and noticed something surprising. Even when models know a new standard exists, they are often haunted by the statistical momentum of their own training data. I call it "The Ghost in the Training Set." A few thoughts on managing this below.
