Jordan Thayer
@jordanthayer.com
130 followers 270 following 130 posts
AI researcher, consultant, professional speaker, video and tabletop gamer, and trans woman. https://jordanthayer.com/
Posts Media Videos Starter Packs
Pinned
Hi, I'm Jordan. I did my PhD in heuristic search, one of the less famous subfields of AI. You use it to plan and control robots.

These days, I'm a developer and consultant. I make AI systems and speak on what AI is and how to use it.

I'm also an avid gamer, reader, and outspoken trans lady.
Hey, good news, I'll be giving a tutorial on off the shelf AI tools (the non-generative stuff) at @techbash.bsky.social this year. We'll talk about machine learning, optimization, planning systems, and a little SAT solving. Looking forward to seeing some of you there!
From now on, whenever I embark on something particularly amazing or silly, this is how I will be referring to it.
Agentic coding tools (and LLMs more broadly) have all the energy of an undergrad who is more interested in a passing grade than actually correct systems or mastery of the underlying concepts, and I am sick to death of pretending that's not the case.
Thanks to everyone who attended @kcdc.bsky.social and a special thanks to the folks that put it on. That quality of conference is a huge amount of labor, and it was an amazing event.

If I didn't see you there this year, here's hoping I see you there next year!
Just a couple more days. I'm driving out tomorrow, but my talk about highway maintenance will have to wait until Thursday. In the interim, I'll be speaking about data structures and Big O notation and wandering around the hallway track. Please come say hello!
I suppose I should use the little title card generators they gave me too. I'll be giving that Data Structures talk on Wednesday. I'm really excited about it. Still can't believe it gets accepted, so quick, before they realize their mistakes.
In a week I'll be out @kcdc.bsky.social. I'm giving a workshop on data structures and a talk on how AI planning approaches can be used to plan for highway maintenance. There's a big slate of content I'm really excited for. You can get tickets here: www.kcdc.info/tickets

Hope to see you there!
In less contentious and depressing conversations, I gave an internal talk on how I use LLMs for the work I do that's not writing source code, like blogging, presentations, and planning how to work events: youtu.be/e5wCxM8zA60?...
LLMs for Not Coding Brownbag
YouTube video by Jordan Thayer
youtu.be
I took some time out of writing about AI in our industry to write about marginalization. It's got a super uplifting title, but I promise it's not just venting or bitching. It's an anectdote laden discussion of being othered and advice on how to avoid that.

jordanthayer.com/every-day-br...
Every Day Brings New Horrors – Jordan Thayer
jordanthayer.com
So, I also got rolled up in this. I have a fantasy novel and a set of open letters about the trans experience published under a pen name that are no longer discoverable.

Many of the things delisted here weren't anywhere near pornographic, they were just queer or discussed heavy topics.
Most AI “research tools” aren’t helping you research. They’re just helping you produce. (Un?)Fortunately, real research isn’t about output, it's about building expertise.

I wrote a slightly more nuanced version of this that you might care to read:
jordanthayer.com/ai-research-...
AI Research Tools Miss the Point of Research – Jordan Thayer
jordanthayer.com
Once again, if someone is looking for a gently used AI expert, my home has made it abundantly clear that my family, my tax dollars, and my capabilities are unwelcome here.
Today I fulfill a lifelong dream. I've wanted to see weird Al live since I was 7. I used to record myself singing his songs because my folks forbade me from buying the albums.
Got a visit in to the Vonnegut museum today. I'd hoped for more correspondence, but it's still lovely
I misread a calendar, and so now I have an hour to read white papers on LLM evaluation techniques. Under a tree. Next to a pond. There's a fountain!

You can take the girl out of academia, but you can't take academia out of the girl I guess.
The longer you struggle to find information, the more ad impressions they get.
A colleague suggested that I produce a sheet of pop culture references for all of my talks so that he can stop asking "Wait, what is that from?" and get back to focusing on the educational content.

Is this a thing people would actually want? I feel terrible bringing leave behinds no one takes.
So there I was, chatting with my friend from Germany about how we're processing the political killings in the US this weekend when this lumbers across the sky...

I suspect I will spend the rest of my life unpacking this moment.
I know it's a bit belated, but thanks to the folks that came out to Stir Trek (if you missed it, it's a lovely show, you should go!) and to the folks that swang by the talk.
This Friday, I'll be talking about AI (specifically combinatorial optimization and machine learning) were used to improve planning for highway maintenance @stirtrek.com

Looking forward to seeing you all there!
For everyone thinking the first point "They aren't good akshuly" is full of shit, I encourage you to use an LLM in a domain where you are an expert for any length of time.

I find that the moment you're capable of being critical of the output, the curtain gets torn aside, and Oz presents himself.
LLMs effectively lean into our social conditioning. It's impolite to tell someone that their work is shit. You assume they're doing their best and you mitigate harm.

I've yet to see a deployment that doesn't hinge on exactly this. Then vigilance fatigue sets in and everything goes to hell. 4/5
Systems that simulate speech mess with our perception. Only humans speak, so we ascribe human qualities to them, and critically, social niceties. If it speaks, it deserves grace and the benefit of the doubt. This is a common theme in tech, going back to at least ELIZA. (3/n)
Almost every place where you'd use an LLM, there's a better bespoke AI technique waiting in the wings. It's just got higher development costs and barriers to entry for development and use.

The _only_ reason that's true is because a bunch of people stole material for training their llms. (2/n)
I think "They suck and aren't useful" is pretty right, actually. It's a matter of having standards.

Interaction with LLMs are an exercise in excusing incorrect, irrelevant, or mediocre output.

When building with them, the bulk of the effort is safeguarding output against those three things. (1/n)
There are a lot of critiques of LLMs that I agree with but "they suck and aren't useful" doesn't really hold water.

I understand people not using them because of social, economic, and environmental concerns. And I also understand people using them because they can be very useful.

Thoughts?