Six months (and a few days) ago, a day after leaving Google, I was entering the OpenAI office for my orientation as a new hire. Today I am posting a retrospective of the first 6 months. mihai.page/six-months-a...

Six months at OpenAI

This weeks marks six months since I joined OpenAI and I have several things to say

mihai.page

There's a certain math meme that surfaces periodically on the internet (or one of its variants). It's nothing more but just computing some sums. So, here I ramble about it. mihai.page/ramblings-on...

Ramblings on a trivial math pattern

There is a certain category of internet memes starting from a pattern involving multiplication of numbers in base 10. Here, we prove that this is something true in any base, there is nothing magical.

mihai.page

A while ago I did a little bit of a podcast episode on the OpenSSF model signing specification (OMS) and the AI/ML working group. It's now up at openssf.org/podcast/2026...

Dear @caltrain.com, the wifi always drops in between Palo Alto and Menlo Park stops. Every day, every hour of the day, for months.

Even with the Trump intervention to cheat out of a red card, the US got owned today. So hard that people would say "stop the count" once again. Gg Belgium! Good game

Dear @caltrain.com , I took 3 trains today, northbound and southbound, in the morning and in the evening. None of them had Wi-Fi working, for the entire duration of the trips. Unacceptable in the land of AI and tech startups.

In the AI era, if you have a bug in production code you actually have at least two. One bug is in the test suite, it should have surfaced the issue when the agent wrote the code.

"If they ever tell my story let them say that I walked with giants."--Troy I'm humbled&excited that the model we released last week is now trending as#2 on HuggingFace,between giant models like DeepSeek,Qwen,and Kimi. Excited that the community finds this useful; looking forward for more work here

SHIPPED: My first release at @OpenAI with a ton of work from an awesome team: a privacy filter model that is small enough it can run in the browser while also pushing the frontier in the space. Also CLI tool to use model. Both under Apache 2.0. Read more: openai.com/index/introd...

Introducing OpenAI Privacy Filter

OpenAI Privacy Filter is an open-weight model for detecting and redacting personally identifiable information (PII) in text with state-of-the-art accuracy

openai.com

I am really excited to read this blog and see this news! OSS-CRS to use AI to find and fix vulnerabilities while treating maintainers humanly, not swarming them with issues and PRs!

OpenSSF@openssf.org · 4mo ago

Welcoming OSS-CRS to the #OpenSSFCommunity Following the success of DARPA’s AI Cyber Challenge (AIxCC), we are thrilled to announce that OSS-CRS is joining the OpenSSF under the AI / ML Security Working Group. openssf.org/blog/2026/04...

The #OpenSSF Ambassador Program is now accepting applications for its first cohort. We are looking for 10-15 advocates to lead local initiatives, mentor developers, and represent the Open Source Security Foundation worldwide. openssf.org/community/op...

After significant deliberations today I decided to get out of the tech job industry. I just sent the "an update on Mihai" email. Starting tonight you can find me as a park ranger in Yosemite. That would be a career that cannot be impacted by AI related layoffs. It's 4/1, folks.

When I was preparing the problems I used to test the 80 LLMs at the end of January I made a mistake. An LLM showed me where I was wrong. mihai.page/ai-2026-2/

I was wrong and the AI corrected me

It turns out that I made a mistake in this year's puzzles I asked AIs to solve. It was an AI that showed me the mistake, and I will fix this in this post

mihai.page

Pi Day is almost ending (already did if you look at UTC time, but there are a few more hours in US), so I just squeezed in an article about some geometries where pi is 2, 3 or 4. There were hints on my blog about this, but now this is done and I have more questions for follow-ups mihai.page/pi-2026/

Pi day 2026: Other values of pi

It's Pi Day today, so we ask: can pi be 2, 3, or 4? We find simple worlds where the answer is yes, worlds that we already talked about in this blog

mihai.page

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...

Learning with AI via deep research and study mode

TODO

mihai.page

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

Tim A.@flipperpa.bsky.social · 6mo ago

Instagram's boss just claimed that 16 hours of daily use isn't addiction. That's every waking hour, glued to a device, on a stupid app engineered to be addicting. What planet are these people from? www.bbc.com/news/article...

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/

Testing 80 LLMs on spatial reasoning on grids

How do LLMs see 2D grids? Would it be harder for them to work on square grids or hexagonal ones? I'm expanding the Kaggle benchmark mentioned in the last article and I'm testing 80 different LLMs. The...

mihai.page

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.

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/.

Kaggle's Task Tuesday Leaderboard | Kaggle

Celebrating the Kaggle team’s favorite community-created Tasks, every Tuesday!

kaggle.com

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...

allow exit and quit commands without leading slash (/) · Issue #16723 · google-gemini/gemini-cli

What would you like to be added? This is feature request - someone can point it as bug as well but I will not. Ask / Request 👉🏽 Implement exit and quit as standard commands with a simple confirmati...

github.com

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 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"

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...

allow exit and quit commands without leading slash (/) · Issue #16723 · google-gemini/gemini-cli

What would you like to be added? This is feature request - someone can point it as bug as well but I will not. Ask / Request 👉🏽 Implement exit and quit as standard commands with a simple confirmati...

github.com

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/

Introducing my first benchmark of AI for 2026

Just like last year, on this special day, I create a new benchmark to test LLMs on different puzzles. This will not be the only benchmark I run this year, but it might be the only math related one.

mihai.page

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/

Introducing my first benchmark of AI for 2026

Just like last year, on this special day, I create a new benchmark to test LLMs on different puzzles. This will not be the only benchmark I run this year, but it might be the only math related one.

mihai.page

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.

Mihai Maruseac@mihai.page · 7mo ago

Two hints: - The books have similar(ish) lengths - The plot of one book is very formulaic, whereas the other one is very random

Two hints: - The books have similar(ish) lengths - The plot of one book is very formulaic, whereas the other one is very random

Mihai Maruseac@mihai.page · 7mo ago

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.