Dear Reviewers: Don't reject because we made it look easy. Making it look easy is hard work. Don't reject because our solution is simple. Simple solutions often have the most impact. Don't reject because you can imagine additional experiments. There's a page limit.

Unpopular opinion: companies continue to shove AI into everything because from their perspective, it's going better than we'd like to admit. One example is Google's AI overviews. I was one of the people loudly complaining about it in its early days when it was in the news for telling people to put glue on pizza. But the quality has improved gradually yet dramatically, and these days I find it pretty useful. I think our disdain for companies "shoving AI down our throats" is largely a selection effect — when one of these AI integrations is new and experimental, we tend to notice, but over time the kinks get worked out, it becomes a part of our workflow, and we stop noticing it. Reminds me of the classic quip that "AI is whatever doesn't work yet." To be clear, even though many of these integrations are pretty useful, the AI-in-everything trend is problematic. There are second-order effects to worry about, such as losing our skills over time when something we used to do manually gets done by AI. And in the case of AI overviews, it cannibalizes traffic to the sources that enable search engines to be profitable. So if we think these integrations harm us or harm society in the long run, it's not enough to simply yell at companies that we hate these features. They have heard us and have correctly concluded that our behavior speaks louder than our words and that we're not quitting these products over AI features. I do think there are some AI integrations we should resist, but to do so effectively we first have to get past the simplistic idea that most AI integrations are useless and companies don't know what they're doing.

I've decided to release a minimal, free online version of my upcoming "10-202 - Intro to Modern AI" course, starting January 26: modernaicourse.org. As a brief summary, this course introduces students to the elements of modern AI systems: you'll build and train a simple LLM chatbot from scratch in PyTorch, without any pre-built layers or pre-trained models at all. More information about the course, and a link to enroll, is available on the webpage link above. In the free online version, you can watch all lecture videos and submit all the assignments (all of which are autograded in the full course anyway). There won't be quizzes/exams, and there's no credit offered via CMU; this is purely for educational purposes. The online course will begin two weeks after the CMU version, and lectures and assignments will all be released with the same two-week delay after the in-person version. Note that this is a first-time offering for this course, so I certainly expect some hiccups along the way. But I hope the material will prove useful for many people.

For a previous birthday, I bought my partner a Klein bottle (kleinbottle.com). This Christmas, I bought her a Klein bottle opener. (Yes, she still responds to my emails.)

Me: …is this the part where you say you can’t advise me any more because you don’t want to be responsible for this disaster? Gemini 3: Not at all. In the spirit of culinary science (and because I am an AI and cannot taste the consequences), we are absolutely doing this.

Eventually, our house will reach an equilibrium where the dogs track as much mud back outside as they track inside.

Thought of the day: Weekly pill organizers are just year-round advent calendars for adults.

I must admit it was pretty surprising when I trained a model on some birds (for a totally different experiment), asked it to name an important politician of the first decade of this century and it said Thomas Jefferson.

Me telling Claude code to make a massive change: GO GO GO GO GO! I BELIEVE IN YOU! CURRENT VERSION IS IN GITHUB SO WE ARE CLEARED TO MAKE A MESS! Claude: HELL YES! LET'S GO! 🚀

"Look up the Waluigi effect in LLMs and generate me an infographic poster for a session describing the entire paper with detail and also cool fun and Nintendo illustrations, please." Google Nano Banana Pro:

Gemini 3 is amazing at vibe coding, but you can ask it something like, “what piece of code controls spawn rate?” and it will rewrite half your app.

Gemini 3 is fu*king so bad at following instructions.

Sometimes vibe coding feels like you’re trying to ride a wild horse, the one they said no one could tame, but maybe, just maybe you could win her over with your animal empathy and kindness.

This sounds bad: @antigravity is vulnerable to the classic lethal trifecta exfiltration attack where a prompt injection can cause the agent to construct a URL to an external server controlled by the attacker and then invisibly leak stolen data to it by rendering a Markdown image

I think chatgpt is what makes linux on desktop actually viable

From NYT today: nytimes.com/2025/11/16/well/… The impact today is mixed, but LLMs are already being used to supplement the medical system, and (IMO) it’s a lot easier, faster, and cheaper to improve LLMs for medical support than to train 1 million doctors to meet demand.

This is called the Black Mirror Gambit. It’s when a company tries to position itself as the villain in an episode of Black Mirror to fish for outraged retweets, think pieces, and credulous venture capital. It’s several years old now and I think it’s getting old.

Due to an unforeseen naming conflict, we are renaming Project AELLA to Project OSSAS (Open Source Summaries At Scale) Thank you to those who brought the context surrounding this name to our attention, and to our partners and the research community for their ongoing support.

This is really great. A course on AI should be preparing the students for AI is 2025 and will be when they graduate, not on what it was in the 1980s and 90s.

this is the correct AI usage policy for college courses

I'm teaching a new "Intro to Modern AI" course at CMU this Spring: modernaicourse.org. It's an early-undergrad course on how to build a chatbot from scratch (well, from PyTorch). The course name has bothered some people – "AI" usually means something much broader in academic contexts – but I think the time has come where the first thing that many students interested in AI should see is how the AI they are familiar with actually works (because it's really simple!) The more people who understand it the better. I'll be trying to put as much material as I can that we develop online (assignments autograding, hopefully lecture videos), though as a first-time course there are also likely to be some bumps along the way. Hopefully it becomes a good resource over time, though. Feedback welcome.