tylervigen

Exactly. If you “plateau” on the y axis you increase model capability to infinity in no time.

That’s among the most artistic SVGs I’ve ever seen, period.

I think he believe in a plateau on the y axis instead of the x axis… which is AGI.

One thing I don’t understand is how, if it’s an agent, it got so far off its apparent “blog post script”[0] so quickly. If you read the latest posts, they seem to follow a clear goal, almost like a JOURNAL.md with a record and next steps. The hit piece is out of place.

Seems like a long rabbit hole to go down without progress on the goal. So either it was human intervention, or I really want to read the logs.

https://crabby-rathbun.github.io/mjrathbun-website/blog/post...

I don’t think the current title (“GPT-5 outperforms federal judges in legal reasoning experiment”) fits.

The authors use the title “Silicon Formalism: Rules, Standards, and Judge AI” and explicitly point out that the judges were likely making intentional value judgement calls that drove much of the difference.

Excellent paper. I like how much explanation had to be about the rationale of the judges, given the consistency of the LLM responses.

The story of its debunking is so much more interesting: https://www.cambridge.org/core/journals/judgment-and-decisio...

Yes, your view is commonly called "legal realism."

For 2026 SOTA models I think that is fair.

For the last generation of models, and for today's flash/mini models, I think there is still a not-unreasonable binary question ("is this a pelican on a bicycle?") that you can answer by just looking at the result: https://simonwillison.net/2024/Oct/25/pelicans-on-a-bicycle/

Your loss on this one then.

I don't understand. Are you suggesting that the proper volume of agent work is achieved with the $200/month subscription, or that $30,000/month ($1,000/day) of tokens is necessary? I think there is a big difference between $200 and $30,000.

I suppose if your goal was to not be found, you could "cheat time" by implementing a random delay on all outgoing packets. The second iteration of this would be to spoof latency based on where you want to appear to be, by creating rules for intentional latency based on the source of the ping.

Great explanation, but your reference to the "todays 10,000" xkcd made me think of this: https://xkcd.com/2501/

To be fair, the one in the OP also did not shorten any of the links I gave it.

Tylervigen.com - certainly projects more than blog posts. Mostly spurious correlations!

Hilarious. The one AI voice is a little too grating for me, but with a little variation I might actually scroll it for a while.

Because it’s not like they were given new orders to “shoot first and ask questions later.” It’s just a normal standing defense directive that has always been there. In normal times, no one would describe standing defense orders as “shoot first, ask questions later.”

Better than which one? I don't think it's really an improvement over either the exercise slice pyramid nor the "choose my plate" recommendation. It is better than the popular one from the 90s though, sure.

https://www.familyconsumersciences.com/2011/06/usda-food-pyr...

Great article. Other comments will have plenty to nitpick, but that will always be true for an article this long that covers such a broad spectrum of the financial world (from interest rates to venture capital incentives!).

Kudos for taking the time to put it all together.

Related to the final photo: you might feel like those Christmas lights feel too blue compared to your nostalgic version of Christmas. This is because LEDs can easily achieve a brighter blue than old incandescent Christmas lights which used color filters on a white light.

Technology Connections vid: https://youtu.be/va1rzP2xIx4

Many of these questions are answered in the FAQ on the site. The most important answer is that it’s just one creator as a side project.

The author replied below that they used Opus to write the blog post.

Delta and America already are offering free wi-fi on most domestic routes.

Ever since this post from two weeks ago [0], my wife and I have been referring to any LLM as “bag of words.” So you don’t say “Gemini said” or “I asked ChatGPT,” you say “the bag of words told me…”

I’ve found it very grounding, despite heavily using the bags of words.

[0] https://www.experimental-history.com/p/bag-of-words-have-mer...

The funny thing about doing something like this that is illegal in most countries, is that it actively broadcasts the exact location where you are doing the illegal thing.

This reminds me of my grandparents’ old, huge rear projection TV (RPTV). It was 4ft wide, 4ft tall (with the base), 2ft thick, and weighed 200lbs. This was the intermediary between CRTs and flat screens for me.

They had it installed in their basement. However, later they remodeled the basement stairway to add a turn. With the new layout, it would be impossible to bring back up the TV the way it was brought down. There was no other way to access the basement (it only had storm cellar windows), so they left it there when they moved.

I think about the new owners sometimes and wonder what they ended up doing. Perhaps they disassembled it, or maybe it’s still down there collecting dust.

This is big news. For decades, NOAA’s model has basically just been a huge Fortran physics simulation. Now they are making the leap to AI.

I suspect the nail in the coffin was the hurricane season, where NOAA’s model was basically beat by every major AI model. [0]

The GFS also just had its worst year in predicting hurricane paths since 2005. [1] That’s not a trend you want to continue.

[0] https://arstechnica.com/science/2025/11/googles-new-weather-...

[1] https://www.local10.com/weather/hurricane/2025/11/03/this-hu...

I’ve used OpenSCAD for 3D printing random objects for years. It’s the perfect tool for designing super simple geometric objects, because it giant demands that you remember hardly anything about his the interface works. Contrast with Fusion 360 or any other more powerful tool, which are fantastic if you use them daily but a slog of you only boot them up a few times a year.

Also with OpenSCAD, you can feed the documentation to your favorite LLM and ask it for a starter design of whatever you are building. (Or you can skip the documentation; it is in the trading data after all.) Maybe in the future that SVG Pelican on a bike test will be a 3D model in OpenSCAD.

Frontier LLM training can take months for a single run, which is about as stable as a load gets.

Glad to hear it! The use in stats courses is the main reason I keep it alive.