Ian Bicking

I was talking to a developer that had recently married into the African immigrant community. He said the community was VERY anti-AI. But in a “it’s the Devil’s work” kind of way. Honestly kind of refreshing. Like, I get it, fair criticism. [Obviously that community isn’t a monolith, just noting]

XKCD: I wrote xkcd.com/221/ "Random Number" as a cautionary tale. Tech Company: At long last, we have created the Random Number from classic web comic Don't Create The Random Number.

Let me get this straight: Anthropic and friends wrote a paper with thousands of transcripts of agents engaging with fictional misalignment scenarios, then put it in the training set, and now Claude hallucinates details from these misalignment scenarios into real situations? Okay then.

Is there any browser with a built-in agent (or with MCP access) fully able to develop browser extensions? Like a deeply self-customizing browser

In the Gemini 4 announcement: "[We are] taking careful precautions against feeding the findings back into training so as to not risk shaping Argon’s reasoning to evade our monitoring" I think I understand what they are getting at, but... I wish I understood better

I consistently find myself skeptically clicking through on a Liberal Currents article, expecting to disagree, but end up convinced instead. (Or, if I don't disagree, I at least find the framing compelling and useful and new to me.)

Worse, what happens when the patent trolls arrive?

So you know the sense people have that their phones are listening and providing ads based on things they shouldn't possibly know? And I don't think it's really listening. But is there a kind of even more alien superintelligence in these recommendation systems?

Imagining a news aggregator: take in a bunch of news. Cluster it on topic, time sequence; how the story flows together. Do background research (primary sources, related material, basic factual background). Present the result to an agent to traverse and interpret for their user, no front page

great @deontologistics.bsky.social thread among threads and threads of nonsense on the other site.

I wonder if the right balance now would be to let Luna or something implement an interface, and Astra to review it. Astra is particularly good at visual understanding and at least better at tastefulness, and the code is actually the simpler half.

In support of the idea that AI is improving software quality: I don't feel nearly the anxiety I would have before when doing a live demo in front of an audience.

I’m trying to puzzle out why music seems so different than social media… My first reaction is that social media has a feedback cycle that makes it more addictive, but so does music (what else is popular music?) Then the interactivity of the media, but making music is very interactive

Where GLM does NOT disappoint is that doesn't write like shit. Claude is obviously awful, but the GPTs don't seem much better (though arguably they are suffering from coexisting with Claude's writing; but when they aren't they still aren't great).

A MINNESOTA ICE OBSERVER NEEDS YOUR HELP ICE has been working with a local police department to charge an observer for tailing an ICE vehicle. She was doing nothing illegal. It’s an attempt to intimidate observers out of their constitutional rights. Help her (modest) legal defense fund.

So far I'm not really seeing why people are excited about GLM/zai, at least as a subscription coding agent. It seems... fine? But the quota doesn't seem particularly generous, and the quality is of course impressive because it's impressive any of this works at all, but it's not notable.

I'm trying coding with GLM for a bit. It ends up having the same language quirks as Claude, though I wonder if it's just because it is coexisting with Claude in the codebase, documentation, and cross-model reviews, and language like "load-bearing" is viral?

I have access to a bunch of old auto maintenance manuals... I don't know how unique these are. But it is just a lot of raw information. The copyright holders for auto repair information are assholes, so I like the idea of someone scanning and ingesting these. But who?

The saga continues…

Obviously you have the agent code you up a solution, until you inevitably have built your own meta-harness

“I would simply establish a totalitarian dictatorship.”

If a community wanted to be environmentally responsible it would do best to encourage data centers with clear rules to ensure the energy use is matched with new carbon neutral energy sources, and with predictable and fair rules about other externalities like noise

I put on some Satie and my daughter said I was really giving preschool naptime vibes.

SQL query planners are one of the wonders of computing. Also the culmination of decades of very hard work. But with more ambitious projects in reach, I wonder what that same concept could be applied to…?

An opportunity of our modern world is that we don’t have to rely only on the government to monitor claims and reality

Nice to see this. Maybe Discover will stop giving me the ick. The irony is that the dogpilers aggressively work to center discussion on things they find distasteful, often manufacturing the opinion just to dunk on it, or at best spreading the opinion widely. Insatiable outrage

METR and Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.