cthos 🐱

https://mastodon.cthos.dev/@cthos

Hi! Professionally I'm a developer (unfortunately) and a fledgling writer (hooray). Personally I work on Tabletop RPG stuff. I love tinkering with new tech, futzing with project management stuff, playing video games (FFXIV, for one), cats, and learning (human/cat) languages. I run authomnibus.com [he/him]

Website: https://alextheward.com
Bluesky: @cthos.dev
Keyoxide: https://keyoxide.org/CBEE4111467A891E67B633C70E0414CBF9A9EAEB
Itch.io: https://cthos.itch.io

Public Key

npub1swmr5e4dt9zrwkvx3d4jqc60jgzuc95jn6z6j4pq95ejgh6nglksq84je6

NIP-05 Address

cthos_at_mastodon.cthos.dev@momostr.pink

Profile Code

nprofile1qqsg8d36v6k4j3phtxrgk6eqvd8eypwvz6ffapdf2ssz6veytaf50mgpz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0dswjkewp

Publishing to

nostr.mom

nos.lol

relay.damus.io

Author Public Key

npub1swmr5e4dt9zrwkvx3d4jqc60jgzuc95jn6z6j4pq95ejgh6nglksq84je6

Show more details

Last Notes

2026-09-19 01:28:12 UTC

Alright, I should know better to engage with certain types of discourse :tm: but I need to talk about something in particular.
Datacenters. Because apparently I need to point out that the size and scale of AI datacenters *in particular* are categorically different from traditional data centers.
For one, the things that are going into them are different. In traditional webapp contexts you don't need a warehouse full of Nvidia H100s that pull 1000W *by themselves* (depending on the version)

2026-09-14 03:40:17 UTC

Saw this posted on Bsky, bringing it over here.
"She made it increasingly clear, however, that she undervalued this discipline and was accustomed to getting what she wanted for free. Her reasoning was that she worked full time and didn’t have the money to spend on the study of suspense thrillers, even though she wanted to write and sell one for profit."
https://ruadanbooks.com/an-editor-on-generative-ai/

2026-07-11 01:42:17 UTC

- reply

*clicks through to the developer's personal website, sees "AI Consulting" in big letters*
Yeahhhhh, instincts were spot on there.

2026-07-11 01:33:52 UTC

I didn't realize it until just now but there's a particular *aesthetic* to vibe coded websites that I keep running across and that's.... weird.
Kinda very similar to how every blog started converging on the same design to game Google search rankings.

2026-05-18 02:43:08 UTC

- reply

One thing potentially worth mentioning... passkeys.
They're stored in Vaultwarden and will sync to the clients, but you can't use them via the web vault.
A forked client could still use them, though (and because I know someone will bring up attestation - the fork could still send bitwarden's aaguid and sites shouldn't be using attestation anyway, etc etc).

2026-04-13 21:09:15 UTC

- reply

I have a couple of ecoflows and have had them for several years and am happy with them!
I also have a couple of the smaller Anker ones and they all work off of the same solar panels.

2026-03-25 20:53:54 UTC

So, two articles from Ars Technica today that are both basically Google propaganda:
- We quantized the models with *no performance loss* teehee, LLMs are great
- Quantum is here, imminently, prepare now!
Like, you're just advertising for "Google is still a growth company, trust us" at this point.
Edit: That's a bit overly harsh because they both do call out "this is what Google is saying" but come on, the timing of this is "we're prepping for the next bubble"

2026-03-18 03:35:18 UTC

RE: https://vmst.io/@jalefkowit/116246587865814045
Exhibit I do not know what number I am up to that markdown files are not “programming” the LLM
#note1fka…lfvd

2026-03-04 22:31:26 UTC

I am once again begging you to understand when you say shit like "It takes a lot of skill to make a prompt that doesn't generate garbage code" is that PROMPTS DO NOT INSTRUCT THE COMPUTER THEY ADJUST STATISTICS.
IT IS A STOCHASTIC PROCESS!
You have no methodology to say that "it really works now!" What you're saying is "it *feels like it works*".

2026-01-07 17:25:39 UTC

- reply

The version with the typo in it was also good because it was making something imperfectly (I thought you did it on purpose).
Anyway thank you for this series, it's so good.

2025-12-23 00:44:43 UTC

- reply

Thank @npub1hzv…lzy3 for that one, I added it after feedback.

2025-12-23 00:39:46 UTC

I wrote a bit about the "one good use of LLMs" (namely, a bunch of different "one good uses") I keep seeing and why those aren't good, actually.
https://alextheward.com/blog/llms_are_a_cognitohazard/
A lot of this boils down to "if you care about accuracy and learning, you shouldn't be using an LLM for anything".

2025-12-19 23:42:39 UTC

LLMs are melting too many software developers’ brains.

2025-11-20 21:47:30 UTC

- reply

I use Aves Libre (there are two versions, Libre's on F-Droid): https://github.com/deckerst/aves

2025-06-18 16:15:13 UTC

- reply

The fact that the GenAI industry isn't providing actual data about their energy usage should make you _very_ skeptical about their claims of "it's fine".
And I'm a little disappointed in seeing folks I generally respect being like "oh no, it's not as much as you think" - y'all need to show me hard data.
Also, I would fucking hope a text extruder uses less energy than serving 4k video, the fact that it presently *isn't* is a travesty - nevermind its other travesties.

2025-06-18 16:13:23 UTC

So, re: LLMs and the "jury is out on energy usage", I just want to point out a couple of historical things.
When studies came out linking smoking to increased risk of cancer, the Tobacco industry paid for and produced studies that said "nuh uh".
When Obesity and heart disease rates increased in the United States and the problem was linked to sugar intake, the sugar industry produced studies saying it was fats.
https://pmc.ncbi.nlm.nih.gov/articles/PMC3490543/#sec4

2025-06-05 17:05:24 UTC

- reply

What happens when your $20/mo junior intern suddenly costs $200?
What happens when you lose power every other week because that datacenter down the road is straining the grid? Did your "agentic AI" help you then?

2025-06-05 17:04:42 UTC

- reply

There's some argument here to be made for "but work just keeps piling more things on me, and I have to keep up" and I sit in a place of privilege to be able to say "no, maybe you don't", but like the net outcome of this is going to be bad for a lot of people.
And I refuse to concede that OpenAI and their ilk are going to come out ahead on this (for one, the economics of it still make absolutely zero sense).

2025-06-05 17:03:17 UTC

- reply

I also don't concede that they actually "work" as their supporters describe. Like sure, it's churning out code-like text at a faster pace than I could, but I'm pretty dang fast at the writing part and I actually understand what I'm doing most of the time.
And when I'm not, I learn through the process and thus get better. Sure, I've written some very bad code, but I also have learned how to *not* do that in the process.

2025-06-05 17:01:42 UTC

- reply

This is the same thing that every tech hype cycle does, every single time. Just flood the zone with shit and see if it sticks.
My own sample set is likewise biased, every time I've "tried" one of these things, it just doesn't do what it says because I only test it on subjects I'm well versed in. It still makes tons of mistakes.
But there's no ethical argument to using these things for "fun" or "productivity".

2025-06-05 16:58:28 UTC

I should probably quote boost this one on its own too, because I also have *thoughts*. I think the point of the AI boosters is to make it difficult to refute their arguments, but they also keep changing their arguments (gish gallop style!): "Oh you're not using the latest highest-paid models! Have fun staying poor!"
This is designed to make their arguments hard to refute, because there's no stable baseline for refutation. The output just *feels* good to them.
https://mastodon.social/@glyph/114629092986679419

2025-03-16 17:54:54 UTC

Did you know that @npub1ulc…x3hg has an offline archive that you can install via @npub1le8…4nlt ?
Now you do! Now's a great time to have an offline copy of your repair guides.

2024-09-12 18:41:53 UTC

- reply

You didn't specify *which* scale.

2024-08-28 16:02:41 UTC

- reply

I appreciated the info dump in blog format, that was a good read though I’m sorry both that you’re burnt out and that social media is social media-ing.