Garbage Collected #1

An irregular roundup of reading, items, and videos from around the web. My own personal vibe-check. A line dropped in the time flowing by.

The cynical fear-mongering of the US AI labs reaches a shriller pitch and becomes more of a circus each week. I recently saw a German news headline over a picture of Sam Altman at the UN that translated roughly to “Those that set the fire now want to play fire department. Something can’t be right there” and I keep thinking of this thoughtful post by Brian Cantrill: The contagion of fear. The technologist, the domain expert, has a responsibility that’s easily abused. It’s easy to shout about a deadly virus in a crowded room. He writes, “…the degree that digital systems have agency in the physical world, it is agency that we — we humans with arms and legs and brains and parents and children — permit: AI executes on physical systems that have been engineered with human accountability and control.” What determines who bears responsibility for the actions of rouge digital “intelligence”? The sensible suggestion from Ones & Tooze: tort law.
I’ve enjoyed re-reading Why are we doing this again? several times because it echos the cycle of novelty and annoyance familiar just after each new model release. On the one hand, “I had it do quite a bit of reverse engineering of my robot vacuum in ways that were quite impressive. So it’s definitely cool!” But it also powers a “slop machine” that runs for 35 hours on a single prompt, exchanges 1400 messages with its subagents and adds 75k LOC that all need to be tossed out. What are these models for, Armin asks. Probably not software. “…[T]hese models increasingly are for other people. For lawyers, 3D artists, mathematicians, whoever uses computer use, etc.”
If you’re interested in what the potential ceiling of these models is and if we’re already brushing against it, and why, then I highly recommend watching this talk by Diogo Almeida given some months before he launched Jev. I understood about a third of it, just go watch it. The rawest claim: We know LLMs shape themselves to the preferences of the user, but did you ever think that we’re now living in OpenAI’s MySpace era. Buckle up.
Btw, I’m not the only one thinking Jev is a natural born SRE. Here’s the world’s first (half tongue-in-cheek) “Jevops” job listing. How do you pronounce Jev in most of Europe? “Jeff.”
The promise of local models is so appealing but the current distance between the server room and what we can run at home is vast. This post of notes on migrating to local models states what should be obvious: “One of the biggest assets we get from Frontier Providers isn’t the model - it’s large context windows. They have the hardware necessary to support your inference. As a result, they get access to the session data.” That’s a reason I expect we’ll increasingly see deep-pocket companies, like this one, investing in their own private hardware.
Do you find pictures of objects taken with your phone attractive. Increasingly the results leave me cold. Sure there are photos I like, but there also is the visible flicker when the image stops moving and you can see a visible sheen layered over it. There’s something about the way an iPhone stamps an image, filtering it. Well recently I discovered Uncamera and I’m having fun taking pictures again. Will it last? Actually I think it might. To my naïve eye it makes pictures look like they were shot on film. There is a warmth to the image that I haven’t seen for a long time. And what I like most is that the final picture does look different from what you see when you shoot it. I really enjoyed this writeup by the developer aon creating it. It’s just been released, go try it!
The memory crisis has morphed into a full blown hardware crisis. So why not purchase 6 busted Steam Deck motherboards, repair them, and then daisy chain them together to run tiny agents? “I need to prove I exist, and the only way to do that, clearly, is by making a cluster and not running Kubernetes on it.” Less than 6 minutes long. One of my favorite videos of this year: I spent all my money on broken steam decks. This guy gets it.