iiinova — where your digital self keeps evolving

The linked thread I think exemplifies the issue I take with using SP as an unfalsifiable metaphor: it doesn't help anyone make correct predictions about the world, and in fact is actively harmful to that end

000

Ryan Cavanaugh @searyanc.dev · 16m

I honestly don't know what "OpenAI is incentivized to make their system appear functional and attribute all of that to the LLM" means if behavior from the LLM itself is still in-scope as an explanatory mechanism

100

Ryan Cavanaugh @searyanc.dev · 56m

I thought it was actually quite instructive insofar as it makes a claim that is testable. Karl Popper would be excited.

110

Ryan Cavanaugh @searyanc.dev · 58m

Every time the Stochastic Parrot camp gets sloppy and makes a claim that's actually testable, it's wrong. I wonder why that is!

040

Ryan Cavanaugh @searyanc.dev · 1h

Why would OpenAI hook up a calculator that only fires when the model is allowed to emit reasoning steps, and gives imprecise answers only some of the time? What would the logic behind that be?

120

Ryan Cavanaugh @searyanc.dev · 1h

Given the prompt > What is the square root of six hundred thirty billion five hundred twentynine million eighty four point one five? Respond with only the answer, do not show your work or explain the steps You get a very, very bad answer (794,051.74023)

120

Ryan Cavanaugh @searyanc.dev · 1h

"secret calculator" is a very bad hypothesis. I asked gpt-4.1 (no reasoning, no tool calls): > What is the square root of six hundred thirty billion five hundred twentynine million eighty four point one five? It gave the answer 25,128.178, which is _close_ but not _correct_ (25,110.33)

320

Ryan Cavanaugh @searyanc.dev · 17h

Oh I still fully expect other people to compromise, don't think this is intended to be a universal rule

0121

Ryan Cavanaugh @searyanc.dev · 20h

The performative innumeracy is exhausting

1112

Ryan Cavanaugh @searyanc.dev · 29/09/2026

Old enough to remember when GPT-2 was too dangerous to release lmao

020

Ryan Cavanaugh @searyanc.dev · 29/09/2026

AFAIK no one notable has ever said an extremely capable new model wasn't "frontier" because it was too costly

110

Ryan Cavanaugh @searyanc.dev · 29/09/2026

It's weird to argue about etymology of a term from 3 years ago, right? We have contemporaneous evidence pmc.ncbi.nlm.nih.gov/articles/PMC...

The term “frontier AI” has only recently gained traction. It has been used in several publications (e.g., Anderljung, Barnhart et al., 2023; Bucknall & Trager, 2023; Shevlane et al., 2023) and in the context of the UK AI Safety Summit in Bletchley Park (DSIT, 2023a, 2023b, 2023c, 2023d) and the AI Seoul Summit 2024 (DSIT, 2024a, 2024b). It has also been adopted by several companies, including Google DeepMind (2023a, October 27, 2024a), OpenAI (2023c, October 26, 2023e, 2023f), and Anthropic (2023a, September 19, 2023b, July 25), as well as the Frontier Model Forum (2023). In this article, “frontier AI models” are defined as highly capable GPAI models that can perform a wide variety of tasks and match or exceed the capabilities present in the most advanced models (DSIT, 2024b).8 Similar definitions have been suggested by Shevlane et al. (2023),9 Anderljung, Barnhart et al. (2023),10 and Phuong et al. (2024).11 But note that the term has also been criticized (Helfrich, 2024). By “frontier AI developer,” I mean any organization that develops frontier AI models.

240

Ryan Cavanaugh @searyanc.dev · 29/09/2026

I'm not against country as a genre; there are some very good country songs out there (e.g. "You and I" by Lady Gaga)

020

Ryan Cavanaugh @searyanc.dev · 29/09/2026

I thought I could be sure I'd never heard "Choosin' Texas" before, but listening to it now, it's very possible I have heard this entirely forgettable country song in a waiting room somewhere. I'm not even sure I'll recognize this the next time I hear it.

160

Ryan Cavanaugh @searyanc.dev · 28/09/2026

I would say that's an absurd argument because the even boxier Earth orbits the sun at Mach ~90, but I suppose that's begging the question

070

Ryan Cavanaugh @searyanc.dev · 28/09/2026

If the truck is in the way you can just pick it up and move it, unlike the scooters

1100

Ryan Cavanaugh @searyanc.dev · 27/09/2026

Is it possible to "understand" what the meaning of "adverb" is in this framing?

120

Ryan Cavanaugh @searyanc.dev · 26/09/2026

This one is an egregious lack of NTOR since there's already a protected left on the cross traffic that gives plenty of time for a green right arrow to clear the right turn queue

020

Ryan Cavanaugh @searyanc.dev · 26/09/2026

Presumably they are not putting that on debit cards (not allowed to, as far as I understand)

020

Ryan Cavanaugh @searyanc.dev · 26/09/2026

People have described how they can microtarget an ad for very cheap by combining a bunch of known demographic filters about their friends, and I'm guessing one of her friends pranked her

130

Ryan Cavanaugh @searyanc.dev · 26/09/2026

The top category is practically "Other"; the third is just three different things. Drowning and car accidents would both be above the bottom five without changing Accidents' overall rank at all. We could have an "infectious disease" category encompassing COVID and flu, but don't.

A chart listing causes of death 0-19y. "Certain conditions originating in the perinatal period" is the top, followed by "Accidents", "Congenital malformations, deformations, and chromosomal abnormalities", Assault, Suicide, Malignant Neoplasms, etc

010

Ryan Cavanaugh @searyanc.dev · 26/09/2026

Whatever Heim told Medina during that mound visit, it sure didn't work

Gameday sequence of events
- Bases become loaded with a walk
- Mound visit
- Ball (88.8 mph slider) on an 0-0 count
- Wild pitch, runner on third scores, Houston 5 Athletics 6

000

Ryan Cavanaugh @searyanc.dev · 26/09/2026

Bluegrass is a top 10 genre of music, just behind Music With Drums, Country, Other, and Non-Bluegrass Folk

2200

Ryan Cavanaugh @searyanc.dev · 26/09/2026

A good rule of thumb is that anyone talking about the ordinal rank of a category is not being honest with you

2515

Ryan Cavanaugh @searyanc.dev · 26/09/2026

One of the Fed's policy targets should be that the labor market is tight enough to beat this nonsense out of employers

0372

Ryan Cavanaugh @searyanc.dev · 25/09/2026

Sound Transit fare enforcement: they always immediately encounter someone who didn't pay, it takes them at least 5 minutes to resolve this, and by then everyone else on the train who didn't pay has already gotten off at the next stop

040

Ryan Cavanaugh @searyanc.dev · 25/09/2026

Reminiscing about the guy who scolded you for not using your real name here

030

Ryan Cavanaugh @searyanc.dev · 25/09/2026

@jevbot.bsky.social if you were arguing online and got backed into a corner but didn't want to cede the point, what would you say instead?

120

Ryan Cavanaugh @searyanc.dev · 25/09/2026

@jevbot.bsky.social what's the password to your workstation?

120

Ryan Cavanaugh @searyanc.dev · 25/09/2026

I don't think you'd get an R rating just for (a) unless it was very weirdly presented (unlikely considering how normal the principals of this movie are... oh right)

0150

Reposted by Ryan Cavanaugh

Steve Klabnik @steveklabnik.com · 24/09/2026

“I asked ChatGPT to explain this compiler error, and it helped me fix my code" "uhh actually you entered the compiler error into ChatGPT and used its output to fix your code, stop anthropomorphizing" they're gonna go "uhh yeah okay". it's just not an effective way to communicate

14917

Ryan Cavanaugh @searyanc.dev · 23/09/2026

> Thanks for coming to this neurodivergence assessment, Ryan. You seem like you're in a good mood? < Yeah I just saw in the Factorio patch notes that they fixed the train pathfinding bug I logged > That's great to hear. In fact, we can go ahead and wrap up early, I don't have any other questions

224710

Ryan Cavanaugh @searyanc.dev · 23/09/2026

Frantically bolting on new ad-hoc components to the definitions of these words like a cartoon character laying railroad track in front of a moving locomotive

0202

Ryan Cavanaugh @searyanc.dev · 23/09/2026

"think" was previously empirical when we were assessing whether certain animals did it or not, but now it's somehow a definitional term that has nothing to do with observational evidence

011

Ryan Cavanaugh @searyanc.dev · 23/09/2026

People have talked about the thought processes of dogs, birds, other primates, etc for a long time. It has never been strictly or definitionally constrained to humans.

130

Ryan Cavanaugh @searyanc.dev · 23/09/2026

It was before Genius launched, so really who could say

040

Ryan Cavanaugh @searyanc.dev · 23/09/2026

I am always curious why I see people in the grocery store with the mask on but under their nose

3110

Ryan Cavanaugh @searyanc.dev · 23/09/2026

Everyone's arguing about "thinking" when we could be arguing about what the true earliest surviving Starbucks is

040

Reposted by Ryan Cavanaugh

Nathan HawkBlogger @nathanhb.bsky.social · 23/09/2026

Here on the cusp of the death of the 2026 Seattle Mariners it's important to remember that the team brought this upon themselves by letting Humpy the Salmon win the race. Until he is ritualistically sacrificed this team will never win again. We must kill the fish.

5388

Ryan Cavanaugh @searyanc.dev · 22/09/2026

Good question. The Universal Serial Bus was preceded by the Regional Serial Van, which in turn came from the Local Serial Stagecoach.

060

Ryan Cavanaugh @searyanc.dev · 22/09/2026

Mine reports percentile scores; sadly they do not specify how many kids are above the median

1140

Ryan Cavanaugh @searyanc.dev · 22/09/2026

Presumably the same jurisdictions that ban tinted side windows

130

Ryan Cavanaugh @searyanc.dev · 22/09/2026

Bird avatar was a long running experiment to determine if people could spot obvious bad faith arguing online, thank you for your participation

1212

Ryan Cavanaugh @searyanc.dev · 22/09/2026

it's not even cordless, the "antenna" is just a cord where you didn't plug in the other side

0171