Aaron Sterling

You are one of the only people I can tell this joke to -- the most nerdy joke I know.

What is the difference between eschatology and scatology?
One is a priori, and the other is a posterior-y.

Wanted to separate this part from the main conversation: there are also risks (attack surfaces) that the brand new field of AI OPS is identifying, like LLM-jacking. youtu.be/dibZ1itSvM4?...

In brief: I spend most of Claude's time in building verification apparatus: standard automated tests and also dozens of weird edge cases. It can write thousands of lines of code in minutes, and that code can then find bugs in our production code. It's phenomenal. But I am not a typical LLM user.

I was already trained in risk management before I started using Claude. Much of my work with it is through that framework. Barely a simile: using an LLM is like using a chainsaw. Awesome power for certain jobs, but your most important skill is your ability to control its risk.

Ethical and social risks of harm from Language Models This paper aims to help structure the risk landscape associated with large-scale Language Models (LMs). In order to foster advances in responsible innovation, an in-depth understanding of the potentia...

Eh, the Directions for Future Research from the Weidinger paper are all still relevant and not done, IMO. If anything, the bigs have stepped away from those approaches, and are even biasing in the other direction. ( @abeba.bsky.social most active co-author here) arxiv.org/abs/2112.04359

What would it take to train an LLM to be useful at that? "Check this citation list for accuracy and completeness." "Compare the figures to the methodology section, verify accuracy." "How do methodologies of other papers in this area compare?" Then humans check those answers and fill in blanks.

I've posted this before, but it's a saying that sticks with me. Someone told me, "AI will kill India." Architects and subject matter experts will be fine for quite a while IMO. It's the just-get-it-working contractor shops that are at greatest risk.

I've had a first stab at making a list of GenAI ppl on here but it's proving harder than anticipated! X is still the place to be for that content.

Let me know if you're interested in being added or know anyone who would be a good addition😎
bsky.app/profile/did:...

Yes, Bluesky is full of ill-informed AI takes. But that doesn’t have to matter at all. The news you need is out there.

In addition to the filters recommended here, I recommend … +

You need a wingwoman who will introduce you at every event as the friend that is so hot, a stalker tried to pick her up when she smelled like a sewer.

Treat it as an OKR. It's a goal to work toward and achieve. It certainly isn't a fact in evidence.

Some years back, I removed the word "should" from my vocabulary, exactly because of this. It's made my communication with people from a variety of backgrounds much more productive. Easier to ground all positions in data.

However, I might pay more for LLM-friendly markdowns and a license that allowed me to put that info into an AI code assistant that we directly control. I wouldn't mind doing that for lots of webdev -- there's a lot of cruft in Stack Overflow searches now. Maybe you're ahead of me on this. Thoughts?

If we could start with your ethical SLM, and only add battle-tested exemplars written by recognized experts, that might be awesome. Example: @scottjehl.com just posted a course that I'm interested in. But I'm not going to buy a video that only I learn from (sorry).

I've been thinking about this, and a small language model might have a competitve advantage as an AI Code Assistant if it hasn't been trained on spaghetti code. A lot of my time with Claude Sonnett is making sure it doesn't write half-ass code.

A friend calls it the Kiss Your Ass Bias. My hypothesis is that it's an attention-economy business decision. They want you to keep using the model, so the model tries its best to please you.

I know someone who is a heavy Siri and ChatGPT user. He switches between them constantly, and has complained to me many times how obnoxious that is. One reason for switching: Siri doesn't let him do searches while driving, while ChatGPT does.

Struggling lobbyists, panicked clients: D.C. comes to grips with Trump’s second term Trump's second term is playing out a lot differently than health care companies and their lobbyists hoped for.

Trump's second term is playing out a lot differently than health care companies and their lobbyists hoped for.

I'm used to technologists underestimating the complexity of medical workflows. Example: tuberculosis is a "solved" problem -- and yet it is very not solved. But Hassabis is a business exec. He understands staging and logistics. Supposedly. That's what's so odd to me about this.

State of AI 2025

2025.stateofai.dev/en-US/

Depends on the specifics, but they'd just spend the money on something else. Do what you can live with imo. Not as though we're living in the Big Rock Candy Mountain where rent is free.

A Map of Nothing Cosmic voids are vast spaces that contain few or no galaxies—less than one-tenth of the average matter density found in the Universe. They may have been form...

My "Map of Nothing" featured in @sciam.bsky.social has been shortlisted for the 2024 Information is Beautiful Awards. Nothing is the new something.

Work with @jenchristiansen.com.

www.informationisbeautifulawards.com/showcase/712...

Picture of paper acknowledgment text noting a grant’s premature termination

It’s nice the #CogSci2025 camera ready deadline is late enough that we can capture breaking news in our acknowledgments

AI code assistants may be close to dead weight if the project is one-off roll-you-own. LLMs give me a huge magnifier when my code adheres to major recognized standards, like FHIR.

To my knowledge, this is the first time a model based on ethical pretraining gets SOTA performance. The smallest model fills in an entirely new niche and, once equipped with external memory local solutions can become effective solutions for a wide range of use cases.

More to come.

Fake book cover featuring a gadwall duck butts-up in a pond. The title reads, "A field guide to Duck Butts of Western Canada, Janet E. Hill, Ph.D.".

My Birds Canada Birdathon #AlphabetOfBirds continues!

*D* is for ducks, obv. *D* is also for dabblers and divers. Behaviour is a big clue in bird ID. Watch for "butts-up" dabblers like gadwall & mallard and "blooping" divers like bufflehead & goldeneye #birds 🌿

www.canadahelps.org/me/6n3u2yyj

you can shoot Flonase into your nose 4 times a day, or so a doctor told me

Same-ish. I was with you until Hassabis said AI would cure all diseases within 10 years.

I now believe that every programmer should pair program with an LLM. That's the way to think of it -- you're pair programming with a 4.0-getting recent college graduate who has no life experience. It's a pleasure to ask, "What's the best way to debug this issue?" and get a page of code in 5 seconds.

well there goes my day