Parv Mahajan

The current SOTA model was released without safety evals

2026-03-07T00:00:00+00:00

OpenAI released GPT-5.4 Thinking and GPT-5.4 Pro on March 5, 2026. GPT-5.4 Pro is likely the best model in the world for many catastrophic risk-relevant tasks, including biological research R&D, orchestrating cyberoffense operations, and computer use. GPT-5.4 Pro has no system card, and to our best knowledge, Pro has been released without any safety evals. We argue this has occurred at least once before, with GPT-5.2 Pro, and provide recommendations for how a team could conduct fast, independent risk assessments of models post-deployment.

This week will not be abnormal

2026-03-03T00:00:00+00:00

For many, the last 7 days have felt particularly eventful. This feels like the first week in a while where the all-encompassing feeling of accelerating change has screeched out of SF and into dining rooms, offices, and barracks around the country. I'm starting to think this is what the takeoff feels like. And if this is what the takeoff feels like, we have to adapt to this feeling.

I got interviewed by the AJC, and internalizing AI unemployment

2026-01-08T00:00:00+00:00

In October, I gave a short talk in an intro seminar class to over 300 freshmen studying Computer Science at Georgia Tech titled "Transformative AI, risk, and you." The talk asked them to extrapolate AI progress and internalize that AI will very soon impact their careers and lives in ways (and perhaps extremely negative ways) they should start thinking about now.

It seemed well-received, and recently I was interviewed by Jason Armesto at the AJC about the talk and how college students are thinking about job loss from AI.

Everything that's in my backpack

2026-01-05T00:00:00+00:00

As I gear up for a couple days of travel, I'm revisiting what's in my backpack! Heavily inspired by Aaron Bergman, here's my preparedness list including protein bars, first aid supplies, Narcan, N95 masks, and various travel essentials.

Turning 20 while the world turns upside-down

2025-12-21T00:00:00+00:00

I turn 20 in January, and the world looks very strange. Probably, lots of jobs will be gone soon. Probably, the smartest thing alive will no longer be human. Probably, things will change very quickly. Maybe, one of those things is whether or not we're still here.

This moment seems very fragile, and perhaps more than most moments will never happen again. I want to capture a little bit of what it feels like to be alive right now.

Seriously, use text expansions

2025-12-09T00:00:00+00:00

You probably type some things a lot, especially if you do non-trivial amounts of administrative or communicative work. Save that time by using text expansions, shortcuts triggered by a short phrase, through something like AutoHotKey (Windows) or Alfred (Mac).

TastyBench: Toward Measuring Research Taste in LLMs

2025-12-02T00:00:00+00:00

An early-stage research update detailing work done alongside Yixiong Hao and Yilin Huang. It's important to benchmark frontier models on non-engineering skills required for AI R&D in order to comprehensively understand progress towards full automation in frontier labs. One of these skills is research taste, which includes the ability to choose good projects.