Alex Irpan
@alexirpan.bsky.social
Research Scientist @ Google DeepMind. Formerly Robotics, now AI Safety. Has a blog. Views are my own.
I didn't know where this post was going when I started and I'm not sure where it went now that it ended, but that felt correct in some way.
www.alexirpan.com/2025/11/16/a...
www.alexirpan.com/2025/11/16/a...
Authentic Imperfection
Auto-Tune is great.
www.alexirpan.com
November 16, 2025 at 5:31 PM
I didn't know where this post was going when I started and I'm not sure where it went now that it ended, but that felt correct in some way.
www.alexirpan.com/2025/11/16/a...
www.alexirpan.com/2025/11/16/a...
First paper since switching into AI safety team🎉
We look at problems that could be solved if the model behaved consistently over a set of prompts, and tried training that in output space and internal activations. Both were effective. See thread or paper for details.
We look at problems that could be solved if the model behaved consistently over a set of prompts, and tried training that in output space and internal activations. Both were effective. See thread or paper for details.
New Google DeepMind paper: "Consistency Training Helps Stop Sycophancy and Jailbreaks" by @alexirpan.bsky.social, me, Mark Kurzeja, David Elson, and Rohin Shah. (thread)
November 5, 2025 at 6:26 PM
First paper since switching into AI safety team🎉
We look at problems that could be solved if the model behaved consistently over a set of prompts, and tried training that in output space and internal activations. Both were effective. See thread or paper for details.
We look at problems that could be solved if the model behaved consistently over a set of prompts, and tried training that in output space and internal activations. Both were effective. See thread or paper for details.
Today is my 10 year blogging anniversary.
www.alexirpan.com/2025/08/18/t...
www.alexirpan.com/2025/08/18/t...
Ten Years Later
My blog turns ten years old today. The big 1-0. Thanks for reading!
www.alexirpan.com
August 18, 2025 at 4:24 PM
Today is my 10 year blogging anniversary.
www.alexirpan.com/2025/08/18/t...
www.alexirpan.com/2025/08/18/t...
For the past month I have been working on a blog post about niche MLP fandom drama. Well here it is.
www.alexirpan.com/2025/07/21/b...
www.alexirpan.com/2025/07/21/b...
Brony Musicians Seize The Means of Production: My Eyewitness Account to BABSCon 2025
Bronies are older fans of My Little Pony: Friendship is Magic. They are mostly male, typically in 20s-30s age wise, and have been trending older and more female over time. (A lot of girls in the origi...
www.alexirpan.com
July 21, 2025 at 5:12 PM
For the past month I have been working on a blog post about niche MLP fandom drama. Well here it is.
www.alexirpan.com/2025/07/21/b...
www.alexirpan.com/2025/07/21/b...
"I don't play gacha games because they're a scam"
vs
"Let me do one more hyperparam sweep before giving up. One more prompt tuning run. I swear we'll beat baseline. I know it's gonna beat the baseline this time. It's gonna win. This time for sure."
vs
"Let me do one more hyperparam sweep before giving up. One more prompt tuning run. I swear we'll beat baseline. I know it's gonna beat the baseline this time. It's gonna win. This time for sure."
June 5, 2025 at 1:16 AM
"I don't play gacha games because they're a scam"
vs
"Let me do one more hyperparam sweep before giving up. One more prompt tuning run. I swear we'll beat baseline. I know it's gonna beat the baseline this time. It's gonna win. This time for sure."
vs
"Let me do one more hyperparam sweep before giving up. One more prompt tuning run. I swear we'll beat baseline. I know it's gonna beat the baseline this time. It's gonna win. This time for sure."
I am now back from #MITMysteryHunt with no memory of anything besides Hunt from MLK weekend. Really this is probably for the best.
January 21, 2025 at 4:38 PM
I am now back from #MITMysteryHunt with no memory of anything besides Hunt from MLK weekend. Really this is probably for the best.
The ship has sailed, but I wish the ML reporting default was % incorrect rather than % correct. It better matches loss curves and magnifies the capture of edge cases.
95% accuracy -> 97.5% accuracy = meh
5% error -> 2.5% error = omg we've halved the error rate
95% accuracy -> 97.5% accuracy = meh
5% error -> 2.5% error = omg we've halved the error rate
December 19, 2024 at 8:36 PM
The ship has sailed, but I wish the ML reporting default was % incorrect rather than % correct. It better matches loss curves and magnifies the capture of edge cases.
95% accuracy -> 97.5% accuracy = meh
5% error -> 2.5% error = omg we've halved the error rate
95% accuracy -> 97.5% accuracy = meh
5% error -> 2.5% error = omg we've halved the error rate