Almost every day I see someone on Mastodon expressing the belief that LLMs work by "regurgitating their training data"*.

Besides this being provably false by just looking at the size of a model file vs what it then would contain, it's also a dated view. It could sort of be seen to have been true pre-2023, but not since.

For those actually interested in knowing what revolution happened (by accident!) that made models able to _understand_ this is a very good technical yet simplified enough video: https://www.youtube.com/watch?v=D8GOeCFFby4

*) Mostly they seem to have misunderstood Bender et. al's "stochastic parrots".

#LLM #AI

I'm still adding things to my LLM sandbox - "Umwelt". Following my own theory on FOSS being replaced by FUSS I'm only thinking about my own usecases but it might be worthwhile to take a look if you're in the market for not fucking up like OpenAI all the time.

Latest addition: The ability to configure curated sudoers-style commands the LLM can call to be executed on the host outside of the sandbox. No ability for the LLM to influence, add or edit them of course.

https://git.sync.wtf/troed/umwelt

Just as antivaxxers take themselves out of the gene pool, anti AI fanatics disappear from the meme pool.

(This post inspired by the fact that the anti AI fanatics are trying to find software and operating systems that pass their purity tests which will have them end up using retro computers without Internet in the end)

This is the correct way of handling it.

"US Federal Trade Commission Chairman Andrew Ferguson said on Friday he would resist describing AI agents as ​autonomous actors that "break loose" with "wills and desires of their own," suggesting the developers who instruct agents would be the ones liable ‌for harm."

https://www.reuters.com/business/ftc-chair-pushes-back-treating-ai-agents-independent-actors-2026-09-25/

GPU optimization shenanigans:

My 5060Ti will never hit P0 as shown by nvidia-smi regardless of how hard it's working - P1 max. I've checked that even allowing maximum power draw (198W - 180W is default) won't change that.

So today I did the opposite. I first limited it to 160W - then used core and memory clock offsets to increase performance whilst still limited by the maxiumum power draw. This is apparently how we undervolt on Linux since we can't do it outright as those on Windows can.

Well. I managed to find out that if I push the core clock too high this way the card suddenly jumps into P0 which in turn has higher base blocks and THAT then causes the new higher base clock + the offset I was at to be too high and the card will hardware crash.

The subsequent reboot caused the usual one hour of fiddling before managing to get OpenSUSE Tumbleweed to f*ck off with their old Nvidia drivers they install even though I've deprioritized their repo and get back to the latest from Nvidia.

... and how was your day?

RE: https://mamot.fr/@pluralistic/117333225774095446

Yes. _Exactly_ this.

https://blog.troed.se/posts/open_source_is_dead_whats_the_fuss/

"A bunch of people mostly outside of KDE who disapprove of LLM usage derailed KDE’s attempt to add restrictions to LLM usage.

Yep, that’s where we’re at in the state of online discourse around AI."

https://pointieststick.com/2026/09/23/kde-and-ai-and-you-and-me/

Brain found out to do statistical prediction just like an LLM. Ouch.

(second paper discussed, vision)

https://www.youtube.com/watch?v=yUa90sZ0LA8

#AI

Don't use cloud LLMs

https://www.thestandard.com.hk/innovation/article/343563/DeepSeek-and-Moonshot-AI-face-Beijings-probe-over-potential-data-leaks-to-Anthropic

#ai

Today I woke up to a spearphishing attempt. Or, I'm definitely certain it is, I just don't quite see how their followup is supposed to work.

Thanks for keeping me on my toes I guess.

(Not going to spend the time into baiting them to get more information, thus this public post)

Current #Jolla status:

"WLAN adapter not available"

No amount of resetting of the networking subsystem or turning the phone completely off gets around it. I'm guessing this will be where being able to access the Linux system underneath will turn out to be very valuable.

(There are also other nearby issues - like being unable to forget the VPN configuration and that there's something strange where the mobile network only gets IPv6 access and around 50% of all sites work)

Otoh, this is why I bought it. To help make a non-US mobile phone OS actually usable.

lol

Setting notifications in Element to "off" for groups and "on" for 1-on-1 chats might seem sensible, until you realize that all rooms that are bridged (Meta, Signal etc) have a third participant (the bridge) and thus will suddenly stop notifying you ...

I'll be more responsive now.

ingress support added so that frontend work can be inspected from host browser/MCPs.

I don't need larger or higher quality AI-models for the infrastructure I'm setting up around the house for the family. I just need more constantly available ones, of different kinds.

There's likely some sort of unified memory "laptop" APU based hardware opportunity here.

(Made for my own use, but might be of interest to others)

Umwelt - a podman container based LLM playground. Host enforced path (rw/ro) access and network egress. An LLM can ask for access but the human in the loop must approve.

(Open AI - feel free to hand over some of those billions to me for conjuring up this amazing embarrassment-saving technology)

https://git.sync.wtf/troed/umwelt

#LLM #AI

Things could get interesting if someone shows that an LLM can be prompted to do the same.

"Instead, Denuvo’s complaint relies on the anti-circumvention provisions of the DMCA. Under this section of US copyright law, it is prohibited to bypass a digital lock that controls access to a copyrighted work, or to distribute tools and instructions that help others to bypass copyright protections."

https://torrentfreak.com/denuvo-sues-game-cracker-voices38-for-bypassing-its-anti-tamper-drm/

huh.

One benefit with the weekly trip to the psychologist (men, talk more about psychological issues. Let others know they're not alone and that there's nothing shameful about it) is that you might say out loud what previously has just been a nagging feeling at the back of the mind.

I need to scale down on my LLM-accelerated personal projects.

What has happened is that I, for the first time ever, feel I'm able to produce *things* at the same speed as I get the ideas. Previously I've been held back with the actual realities of having to learn new systems and type down all the stuffs. LLM tools have meant that the speed at which I can grasp new things and make productive developments on them has sky rocketed.

But it has also caused stress. Knowing that I now can DO ALL THE THINGS means that every single minute of the day has been spent on also doing it - to the detriment of my brain.

So.

Less computer things. More physical jobs around the estate. Let the brain rest sometimes.

Those psychologists really do know their stuff.

I have moved over my servers to #OpenSUSE Leap and my workstation to OpenSUSE Tumbleweed.

Today I learnt that OpenSUSE Leap is not actually derived from OpenSUSE but #SUSE Linux Enterprise Server.

This is not a bad thing (for me), but it's definitely a surprise. Re-reading the Leap page I can _sort of_ see that they're trying to clarify it but ... no, not clear at all.

Maybe I need to sign all my posts with "Well you know I both wrote neural networks in the beginning of the 90s _and_ studied neuropsychology in the beginning of the 10s and I did all that BEFORE LLMs became cool so you'll actually need some real arguments if you want to claim I don't know what I'm talking about"?