@mihai.page
Building AGI with Privacy and Security at OpenAI. Previously: ML Supply chain security @ Google OSS Security Team (model signing, GUAC). Previously: TensorFlow Security & OSS (@ Google) Previously: Haskell+differential privacy+ML @ LeapYear
I've been using LLMs to answer questions and learn in the past weeks. I'm summarizing how Deep Research and Study Mode (Guided Learning / Study and Learn) from Gemini and ChatGPT can be used for this, without going into too many details. mihai.page/deep-researc...
Between this and the new patent for letting AI post after a person's death I'm utterly convinced that Zuckerberg is a sick ghoul
I tested 80 models on two simple grid-based problems, asking them to locate 2026 and compute the sum of the neighbors when the numbers are placed in a spiral on the grid. Results surprised me: models performed better on harder problem, but they also cheated. Read more at mihai.page/ai-2026-1/
Today was my first day at OpenAI. Excited about what's to come. Will still work on security and AI.
My mission evolves from "OSS must be secure, especially AI" to "Let's build AGI with security and privacy in mind". OSS is still part of my work, so let's keep in touch.
Image of 2 laptops, one clean, one with a lot of stickers on it (Google, OSV, GOSST, CoSAI, OpenSSF, AntennaPod, Gemini, DevConf, PyOpenSci, TensorFlow, HEIR, Scientific Python). On the side, there's a censored employee badge.
Yesterday was my last day at Google. It was a bittersweet departure, leaving a team I really enjoyed working with. GOSST's mission is extremely important and I still believe in it.
I'm currently running 80 different models for this task, some are slow, some are fast. It will take a while until I can get all results. But I'll publish them when ready and announce. Like and subscribe :P
One of my tasks for the puzzles I'm sending the AI this year has been featured in Kaggle Task Tuesday. You can see the task at www.kaggle.com/benchmarks/k... and you can read more about the puzzles at mihai.page/ai-2026-0/.
Use `--yolo-mode`, Ralph, Gas Town, or any other automated tools? Beware of agent loops that run out of control. See a less dangerous example on github.com/google-gemin...
step 1: vibe code slop in prod
step 2: ???
step 3: "we take security very seriously. Your data is now in our hacker's hands"
Mistral models are too slow and this is why this experiment will take much longer than expected
Mistral models are way too slow
step 1: vibe code slop in prod
step 2: ???
step 3: "we take security very seriously. Your data is now in our hacker's hands"
4 progress bar, one labelled "Total run progress", another "prompts_square", then one named "gtp-4o" and the final one "Model score on task". The 3rd on is at 30%, the rest at 0%.
Let the experiments begin:
Use `--yolo-mode`, Ralph, Gas Town, or any other automated tools? Beware of agent loops that run out of control. See a less dangerous example on github.com/google-gemin...
New year, new problems to test the LLMs on: arranging the numbers on a spiral, what is the sum of the neighbors of 2026? Read on for more details and some preliminary results, as well as how to suggest other LLMs to test: mihai.page/ai-2026-0/
New year, new problems to test the LLMs on: arranging the numbers on a spiral, what is the sum of the neighbors of 2026? Read on for more details and some preliminary results, as well as how to suggest other LLMs to test: mihai.page/ai-2026-0/
I cannot remember the rule of divisibility by .7., it's always so hard. This is the smallest prime where I have to actually do the division myself
The books are:
- A Million Random Digits with 100,000 Normal Deviates
- One Million Digits of Pi
They both have the same length (1 million characters), the same 10 characters (the digits themselves). The plot of the first one is very random, whereas the second one has formulas.
The books are:
- A Million Random Digits with 100,000 Normal Deviates
- One Million Digits of Pi
They both have the same length (1 million characters), the same 10 characters (the digits themselves). The plot of the first one is very random, whereas the second one has formulas.
The books are:
- A Million Random Digits with 100,000 Normal Deviates
- One Million Digits of Pi
They both have the same length (1 million characters), the same 10 characters (the digits themselves). The plot of the first one is very random, whereas the second one has formulas.
Two hints:
- The books have similar(ish) lengths
- The plot of one book is very formulaic, whereas the other one is very random
Can you name two books, both having the same 10 characters, but being very different in their plot otherwise?
There is not just one single solution. I'll post a hint later this week if no one answers by then, but until then please share widely to see if anyone can name them.
Can you name two books, both having the same 10 characters, but being very different in their plot otherwise?
There is not just one single solution. I'll post a hint later this week if no one answers by then, but until then please share widely to see if anyone can name them.
Yeah, translation is correct. I know the English one I pasted and then looked for the original but didn't see that it doesn't match perfectly
Today's lesson in the words of Pompeii: "Ξα½ ΟΞ±ΟΟΞ΅ΟΞΈΞ΅,β Ξ΅αΌΆΟΡν, βαΌ‘ΞΌαΏΞ½ α½ΟΡ΢ΟΟΞΌΞΞ½ΞΏΞΉΟ ΞΎΞ―ΟΞ· Ξ½ΟΞΌΞΏΟ Ο αΌΞ½Ξ±Ξ³ΞΉΞ½ΟΟΞΊΞΏΞ½ΟΞ΅Ο" (that is: "do not quote the laws to people holding swords", in an approximate translation)
Just posted my 2025 wrapped blog post (mihai.page/2025-wrapped/). I met all my goals for the year, time to make goals for 2026 and have 2026 be a year of doing, then talking about done. I have some concepts of some plans, so to say.
π‘ This weekβs OpenSSF Project Spotlight explores the Model Signing project with Mihai. Learn how verified model signatures help trainers & developers ensure their models havenβt been altered; and why lightweight, flexible signing beats container-bound approaches.
π youtu.be/P1AE23uZQ50?...
#AIML
New favourite quote π
"Any sufficiently well explained mathematics is indistinguishable from being obvious" - Matt parker standupmaths
On today's installment of "blog-every-day-until-xmas", I talk about how Anna Karenina applies to the basics of linear algebra (or am I?) mihai.page/anna-karenin...
On today's installment of "blog-every-day-until-xmas", I talk about how Anna Karenina applies to the basics of linear algebra (or am I?) mihai.page/anna-karenin...