sobchak

If the company also trains AI, I’m guessing they’re also forcing use to gather training data. I.e. tracking feedback, corrections, etc.

I've been noticing a lot of counter-surveillance projects lately too. I.e. Flock-You, Chasing-You-Tail-NG, drone trackers. Most using ESP dev boards or rPis. The cardputer looks pretty cool too. Dunno if that stuff fits or not.

I think this entire war was so Trump and the admin could personally enrich themselves by taking control of Iran's oil in some way. They had success doing it Venezuela, and they were then convinced they could do it in Iran too. Attacking these countries would only produce limited, "well I agree, but don't like the way he's going about it" push back from establishment Dems and Western politicians in general. Israel "helping" is just a bonus, and provided further buy-in from the many Zionist politicians. They thought it would be a quick decapitation and puppeting like Venezuela was. I think Israel made things worse for the admin by bombing the people the admin had in mind for being the leaders.

I mean, you could. Just encode 100ms chunks or whatever into tokens then push them through the same model. I'm pretty sure that's what the claim to do (though with MoE/routing now, maybe).

According to their tech/marketing papers, it's supposedly multi-modal, encoding audio to tokens.

I do try to give it a chance and use it every once and a while (most recently Claude Code; last year, Cursor), and it has been my experience that it personally decreases my productivity and quality. I found that even CoPilot’s autocomplete would introduce bugs if I decided to “trust” it and try to work too fast without meticulously reviewing every token generated. I have seen people I work with use AI to quickly create decent looking prototypes (i.e. bog standard, boring design), but I think this is still detrimental because they lose the full benefit of exploratory programming (and of course, the prototypes just have all kinds of faked data and functionality, glaring security problems, bad architecture). I’ve also experienced people submitting nonsense vibe-coded pull requests that would break tons of things they shouldn’t have even touched for the issue. I could see a less interested or overworked reviewer letting stuff like that through, which is why I think we’re seeing all these failures and bugs at these big tech companies. So for me, at least, I haven’t seen the benefit. Using CoPilot in VSCode actually caused me to go back to using nvim and lsp plugins :)

Haven’t used Mint in a while, but I think it’s mostly a function of the kernel. IIRC, there’s a nice GUI for selecting what kernel you want on Mint. Would probably only make a difference if you have new hardware and need a newer kernel.

I think at least some are scams. I would think there are ways to design cameras to mostly mitigate this too. I’m guessing ALPRs have the ability to see license plates at night even though headlights are emitting a large amount of light over a broad spectrum (including IR).

A single Dynavap bowl of decent flower will get me feeling buzzed. 4 bowls, I’m glued to the couch for 2-4 hours. I don’t use very often though. Sometimes a few times a week, sometimes months without using.

Not enough room in the GPU machine for all the HDDs I needed. Also who got millions of photos at home? People working on biological datasets.

If the corporate VPNs keep logs and allow government access, they will be allowed. That’s how it works in some (authoritarian) countries.

I believe China does statistical analysis to do stuff like detecting and blocking VPNs, suspicious looking ssh traffic, etc from home Internet connections not going to an approved business. It’s my understanding it’s very hard to get around the GFW at the moment, and pretty complex stuff is needed to mask VPN traffic to make it look normal (Project X, Xray, Reality, etc).

Rsync, syncthing, backups, mp3s, photos, json files; idk, a lot of tasks involve large amounts of small files. I personally ran into this problem training models on millions of photos. My GPUs would only get up to 25% utilization with mirrored HDDs, so I had to switch to SSDs.

HDDs have horrible random access times, so if you need to process or just copy a lot of small files, say photos, there’s a significant penalty.

Alibaba has released Qwen models under Apache licenses (and they are some of the best models that can reasonably be ran locally). Some argue that models aren’t really open source unless the training code and datasets are made available though.

I don’t think it’s been proven that the nitrogen the beans/bacteria fix is available to the corn (before the bean plants die and decompose). Though, I have done this (along with squash), and it does seem to work pretty well. I think it gives you more vegetables per sq ft, than if you were to grow them all separately.

Humans lived in what could be described as a sort of primitive communism for most of the species history. Basically, the society needs to be decentralized. If you can keep it sufficiently non hierarchal, there isn’t a lot of power people can get over many others. A problem I see with this is defending against large, centralized, outside organizations. So, I guess you’d need some federation-like structures. Some communes are pretty democratic and decentralized. The Zapitista territories are the best example I know of, of a large non-hierarchal federation of communities.

Small local businesses fuck over their employees too. Capitalism incentives it. It also incentives monopolies. And it seems when the wealth disparity gets large enough, it captures government and starts transforming into fascism.

I think California recently passed the Digital Age Assurance Act, which was backed by Google,Meta, and OpenAI. I think it goes into effect in 2027.

I find it detrimental to my productivity when integrated into an editor/IDE. I’ve found the “autocomplete” causes subtle bugs that I end up overlooking because I’m trying to go fast and putting too much trust in the generated lines/snippets. Tracing down these bugs becomes a huge time-sink. I do use chatbots in the browser for various things; mostly as a kind of “search” for alternative ways of doing things, frameworks, libraries, and algorithms. Agentic vibe-coding is ok for small one-off tools/scripts you wouldn’t need to maintain, IMO.