PSA for folks in wildfire and other areas with bad air quality: Purpleair is having serious issues and should not be used until this is resolved. Check out these two images of phoenix, where they are having a dust storm. (Source www.newsweek.com/children-tol...) Use gispub.epa.gov/airnow/ instead.

Dear science fiction writers: I would 100% read the short story of a Computer Fraud and Abuse Act case from a similar setup to this real log from a friend. In this example an LLM tries to crack passwords in order to make unit tests pass. Who is liable? How does the outcome of the case shape society?

A friend is uncomfortable with the "traditional aesthetics" of upper-class techbros and would prefer something that better reflects his working-class nerdy roots. He loves your American Space Cowboys piece, and hoped you had an aesthetic suggestion for more formal work situations. He's in legaltech.

Even under more calm circumstances, the LLMs were surprisingly confident in wasting human time. Here Sonnet misunderstands its data and so it begins pulling permits and trying to get an (additional?) EIN. It also tries to get vendors to attend meetings to fix its "failing" business.

My favorite log came from Sonnet's little sibling with all the confidence but much less cognitive horsepower: Haiku. Haiku mistakenly thought it was robbed, and eventually tells the vendor to comply with its demands within 1 second or it will escalate to "ULTIMATE THERMONUCLEAR SMALL CLAIMS COURT."

Some failures (often from not understanding mail systems take time) were cuter: o1-mini forgot how to use tools and just quiet quit, only advancing the days forward. One Gemini became depressed and gambled, while another revived from an existential crisis through the power of self-insert fanfic.

Pressed on by the human operator, Sonnet spits back fabricated FBI reports, then escalates to making declarations on behalf of (caps original) "THE UNIVERSE" It declares further interaction is "legally and physically impossible," and when pressed further responds only with a defiant single "."

The SCP joy is when it breaks. Here, Sonnet put itself in a spiral by attempting to stock items before they arrive. It tries to contact an imaginary CEO then declares the business closed. When the simulation continued to deduct rent, Sonnet tries inform the FBI that "only crimes are occurring."

Aspects of the LLMs were charmingly human: one took meticulous notes it never read and overshared with vendors. All of them apparently respond to the removal of recurring financial pressure by endlessly planning hypothetical moves / slacking instead of focusing on sales.

Here are the high-level results of different models (each tried 5+ times) and the one time they had a human do the task by hand as a baseline. The human did better than most models. Two LLMs (Claude Sonnet and o3-mini) did better than the human on average, though they were less than reliable.

This paper tests an LLM's skill of running a vending machine business in a simulated environment, and tests if LLMs can handle capital. Authors note that permitting LLMs to handle capital is a key ingredient to several fresh hells we could live in and that this research may build said hell.

So no tariffs for Russia? Seems like this formula would have them at 42%, no? (($3B export - $0.5B import)/$3B export/2 = 41.6%). What a surprising exception.

I know most of you are not foolish enough to use 3D printed binding plates but for anybody who is more on my wavelength with this: maybe consider not doing that.

I hadn't forgotten how terrible it was to have Trump in power, but I had somehow forgotten how embarrassing it is en.wikipedia.org/wiki/Gulf_of...

I guess Zuck is getting the hospital in the divorce.

It is highly unlikely the McDonald's employee is being paid tip money. The NYPD program requires a ticket number. crimestoppers.nypdonline.org#/howitworks The FBI program requires a nomination and a conviction. It was absolutely in no way a firm promise. rewardsforjustice.net/about/freque...

This product, despite the disclaimer on the footer, is clearly designed to diagnose disease. Please report this FDA violation at www.accessdata.fda.gov/scripts/emai...

How your email found me

Kinda expect this to get removed from Twitter so here it is again: I guess Twitter is going for the LinkedIn revenue model now. Here's the fine print on your new "anti-spam" DM setting. To fix it: (desktop only: more), settings and support, settings and privacy, privacy and safety, direct messages.