Daniel Van Zant

When I find an interesting article/paper/etc I don't read it right away, I save it in a folder on my computer. Once I have enough, I run a small script to get it printed as a glossy magazine with an AI-designed cover (here's the latest) and read through everything at my leisure.

I am presenting online for the Wolfram Institute at 2pm ET today. Will be talking about an AI system that I am building and how it might be used to validate abstract theories with purely computational "Wolframian-style" experiments. m.youtube.com/watch?v=S9qz...

Got to have a great conversation with my advisor Dr. Elan Barenholtz on his theory of language! youtu.be/P-yfQLDM5pA?...

Setting up a weapons shop inside Rome's walls makes you a respected merchant. Outside the walls, you're seen as a barbarian threat, but you're free from Rome's taxes and politics. Research has the same geography. It's worth knowing which side of the wall serves your purposes.

One minute clip from a short talk with Dr. Elan Barenholtz about how he uses AI in his work. Check out the full video here: www.danielvanzant.com/p/an-extra-c... . Also worth checking out an interview he did with Curt Jaimungal of Theories of Everything: curtjaimungal.substack.com/p/the-theory...

For anyone interested in this question this guy: substack.com/@cyberneticist sent me this excellent resource: people.eecs.berkeley.edu/~jrs/highd/ over on Substack.

Being unbiased in science communication seems almost impossible. I admire scientists who, instead of pretending to be unbiased, work to be transparent about their biases. I.e. "I came up with this theory so I prefer it," "my field looks at this thing this particular way," etc.

5 years since our first pilot, and 25,000 participants later, I'm super happy this work with Makaela Nartker, @chazfirestone.bsky.social and Howard Egeth on inattentional blindness is now out in eLife! A little 🧵 of what we found... 1/12 elifesciences.org/articles/100...

Wikipedia is also a good evaluator of how much I can trust the LLM. If the topic has multiple fleshed-out Wikipedia articles with lots of sources then the LLM is usually mostly right. If there's just one two-paragraph article that cites one paper then usually the LLM output has many hallucinations.

There's so much polarization around LLMs. They are way overhyped, I agree. But I also use them semi-regularly now. Here's a thread of genuine use cases where I find them helpful. Please add your own!

Demo of an AI tool that displays relevant information live as you are having a technical conversation or presentation. The display is based on either a query you gave it beforehand where it can grab and process relevant documents from the internet. www.danielvanzant.com/p/augmenting...

asteriskmag.com/issues/09/gr... Very cool article going into detail on what the best strategies might be for humans to take their first steps into becoming a multi-planet species. Always excited about anything on how we could get one step closer to Star Trek.

I will be presenting my work building an artificial intelligence cognitive co-pilot to help neuroscientists create better theories next week, April 9th, 10am EST. Anyone is welcome to join the Zoom: fau-edu.zoom.us/j/3806783624... . The talk will also be recorded and posted on Youtube.

www.newscientist.com/article/2472... UK tech secretary's official chatgpt records were requested and given under freedom of information(FOI) laws. Under most FOI laws, text conversations between two people on are exempt. Should conversations with a chatbot be treated the same?

Have been enjoying participating in these "salons" which are long-form semi-structured conversations with a neat group of scientists and thinkers. I have appeared in these two: youtu.be/lzLg4lEnQuY?... youtu.be/TN4aMXWpnoI?... and there are others on that Youtube channel.

arxiv.org/abs/2502.203... Maybe LLM's are using symbolic mechanisms built through transformers? Raw transformers are only the "machine code" explanation for how LLM's incredible performance. I really enjoy papers like this that form theories about what the "software" might be.

open.substack.com/pub/elanbare... Start of a series my advisor is writing on his theory of cognition. Even though I disagree with some parts of it, this theory is well-argued, well-evidenced, compelling and accessibly written. It deserves serious thought and consideration.

Paper comprehensively benchmarking different models and methods for RAG use-cases. If you're wondering "what is the best reranker for X" or "what is the best architecture for Y type of questions" there is probably a table that answers that question. arxiv.org/pdf/2502.202...

Excellent interview from the leader of a research org that is managing to “think big” without becoming based purely around ideas and do applied work without becoming completely profit-focused. I love seeing stuff like this out there in the world. substack.com/home/post/p-...

youtu.be/miLqkSU_qG4?... An awesome conversation I got to have with Dr. Elan Barenholtz on one of my favorite YouTube channel. We talked about what the structure, successes, and failings of large language models could mean for cognition more widely among other things.

A short piece I just wrote on a really important trade-off in state of the art RAG: substack.com/home/post/p-...

Have been trying to understand what it means for a system to be Turing-complete (which in theoretical computer science acts as a sort of "check" for programming languages or physical computing systems). For a system to be turing-complete it must be able to compute an arbitrary

Anytime a researcher who is an expert in their field writes a book with the intention of being accessible, I immediately add it my reading list. Chaitin, Ericsson, Penrose, and Schwarzlose come to mind as folks who are examples of this. Any that I should add to my list?