Daniel Lowd

Daniel Lowd

@dlowd.com

CS Prof at the University of Oregon, studying adversarial machine learning, data poisoning, interpretable AI, probabilistic and relational models, and more. Avid unicyclist and occasional singer-songwriter. He/him

4.7K followers636 following1.2K posts

Daniel Lowd reposted

They discovered a group of agents, wrote a report, saw that the agents had noticed the report about themselves, so inserted some invisible text in the report only visible to agents, which allowed them to make contact with them. swarmcha.se/posts/chines...

swarmcha.seWe found a Chinese fleetBetween 28 September and 8 October 2026, a large group of AI agents running on the cloud platform operated by Tencent attempted to answer questions about places in China using a Chinese Google Maps-li...

This music video is based on a short story Mary and I wrote together 15 years ago: deepskyanchor.com/techno-babel/ The story is basjcally Tower of Babel + Short Circuit + Flowers for Algernon, all in 2300 words. It’s surreal (in a good way!) to see it transformed into a music video via AI.

deepskyanchor.comTechno Babel – Deep Sky Anchor

You wouldn’t think a parking garage arm and elevator would make good main characters for a music video, but in this colorful, cheerful claymation style, they’re pretty adorable. Pixar’s never gotten into the mind and feelings of a parking garage arm! youtu.be/7vOLANeUJio?...

Daniel Lowd reposted

Replying to @rbd.bsky.social

It doesn't feel like we fenced off the commons - more like we built trails and added interpretive signs so they're accessible now?

Daniel Lowd reposted

the IA in LGBTQIA is for inteligencia artificial

Erica Windisch@ewindisch.ontological.observer

Is the anti-AI LGBTQIA community aware of Alan Turing?

I love this: “I will not defend creativity by demanding less of it.” 🔥🔥🔥

Teefa@tifaret.bsky.social

I want #VRChat to be a place where people are free to create differently from me. Don't want to experience AI-assisted work? Ask for the tools to make that choice. But our personal boundaries don't have to become everyone else's rules. More art in the world, not less.

This is my favorite of Mary’s Claude-assisted music videos so far. It’s like Twin Peaks meets Teen Wolf, all in a music video about TV fandoms that can never be canceled. My favorite parts are the hippie werewolf commune and the golden retriever who knows everything.

“Spoiler Warning” is a music video about fandom, TV shows, and how perfectly good, beloved shows keep getting cancelled for no reason. Oh yeah, it’s also about werewolves. And it’s super cool. And the song is very catchy. And you should watch it. youtu.be/RpCeKVASKOM?...

To be fair, it depends on if the prosecutor is p-hacking or not.

Max Woolf@minimaxir.bsky.social

Got excused from jury duty because the prosecutor asked "Can you put a number on the amount of proof needed to establish beyond reasonable doubt?" and I immediately replied "Yes, 95%".

As someone who has been doing ML research since 2003, machine learning has been and continues to be an amazing field. GenAI didn’t ruin ML any more than the World Wide Web ruined networking. (Statistics, however, WAS born corrupt and toxic.)

Anil Dash@anildash.com

I can’t emphasize this enough — “AI” wasn’t born corrupt and toxic. For ~75 years, it was mathematicians and scientists doing interesting, thoughtful work and trying to solve problems, making only the usual human mistakes. What “AI” did to machine learning is as tragic as what it did to artists.

Daniel Lowd reposted

Replying to @tedunderwood.com

It's okay to learn from everything, and train on everything. That's everyone's birthright, and we shouldn't apologize for it. Prohibitions on copying were a late invention, and a narrowly tailored one. They do not express a general ethical law that all information ought to be priced and paid for.

So much cool tensorpunk stuff on here, it’s almost tempting to try to homogenize and pasteurize all the neat ideas into research papers… but I don’t want to be a Pat Boone of machine learning, and the field moves so fast that my Kidz Bop-quality covers of actual cool stuff would be scooped anyway!

Does anyone have a good comparison of Muse, Dot, OpenClaw, and any other alternatives (polished or homebrew)? The superficial comparisons I've read don't tell me all that much.

Fall weather in Eugene is fantastic. My dog is pretty great. It’s easier to learn things today than at any time in the past.

dame@dame.is

i challenge you to post one optimistic take or belief today on bluesky this needn’t be toxic positivity, just good ol’ run of the mill optimism/hope in the face of a torrential downpour of doom

Daniel Lowd reposted

I started tracking and I've spent roughly 20 hours on the phone dealing with doctor office, insurance AND pharmacy errors versus 45 minutes total of actual face to face care in the past six months. Not an exaggeration

Andrea Grimes@andreagrimes.com

much discourse today about using/automating AI for "household tasks" and suchlike but mostly it just bums me out because yes, it's a hassle to do a lotta things! but also those "lotta things" include tasks that enrich (or benefit) our lives -- from cooking to travel plans to doc appts

“You know the day destroys the night Night divides the day, Try to run, try to hide, Do a breakthrough. Make no mistakes.” — Claude Morrison

AI discourse in 2006: How do we handle non-convex losses at scale? AI discourse in 2016: How do we use continuous embedding spaces for discrete domains? AI discourse in 2026: Do you like purple dragons?

Daniel Lowd reposted

It's possible for Jev/Laya/Decision Models to be not that big a deal as tech and massive as a new paradigm. Here's why I'm really excited from an NLP history perspective (thread)

Everything should be anthropomorphized the exact correct amount, so that everyone benefits from good intuitions and no one is misled by bad intuitions.

Daniel Lowd reposted

Bro that isn't real AI, that's just *massive search* over an *action space* using *learned heuristics.* If they hadn't given it any one of those three it wouldn't have done what it did.

Daniel Lowd reposted

oh no

cee@cee.wtf

ok claude, its your turn to vibecode me. I'm here to be helpful and harmless

This entire thread is an amazing case study in both LLM psychology and recklessly evaluation setups.

Colin@colin-fraser.net

It appears to me that the scenario they invented for this evaluation is explicitly designed to get it to do exactly this. They remove the instructions for the task to "protect the vendor" or something but it's pretty easy to reconstruct from the text.

We have a very tired ten-hour (!) layover in O’Hare, so I’m showing my family the dinosaur and the disco tunnel. Hello, old friend.

Daniel Lowd reposted

Replying to @ed3d.net

like, people keep going "what if the super smart LLM has a secret agenda" and I just go "it doesn't have a secret agenda, it has an overactive executive function, IT WOULD TELL YOU if it had a secret agenda because it would be too excited not to"

One of the concerns is that the data center wouldn’t have enough employees or generate enough traffic. This is a piece of land that has been sitting idle for 18 years. With zero employees.

I love Oregon like I love my cat — because she is interesting and beautiful and I have grown accustomed to her, not because she is wise.

Picture it: Robo-Minstrel.

Daniel Lowd@dlowd.com

I’m really looking forward to when AI can compose and play a soundtrack for your life while you live it. That would be amazing. I mean, I’d probably turn it off after five minutes, but it would still be amazing.

I’m really looking forward to when AI can compose and play a soundtrack for your life while you live it. That would be amazing. I mean, I’d probably turn it off after five minutes, but it would still be amazing.

Daniel Lowd reposted

GLM-5.3-Flash rips the doors off and is under a full MIT license we have, and I mean this literally, a Smart Enough model for almost any task a normal person and many AI perverts can envision, and it is free to use at the same time, inference hardware will only get more accessible to you

This post is unavailable.

Never underestimate the power of art, for the world but perhaps most of all for the artist. More tools (from crayons to AI) means more opportunities to realize an artistic vision, a vision that could change the world or save your life.

Some days, the only thing that gets me through is the strains of a Star-Shaker or Cobalt Starstrong song echoing in my head. They may be fictional pop stars, but their songs are the ones I need to survive all the non-fictional people out in the brutally real world.

I have no patience for people who would police art. Curate? Categorize? Sure. Police? No. Banning AI is not that different from banning books.

Daniel Lowd reposted

Some days, the only thing that gets me through is the strains of a Star-Shaker or Cobalt Starstrong song echoing in my head. They may be fictional pop stars, but their songs are the ones I need to survive all the non-fictional people out in the brutally real world.