Notices by Thomas 🔭🕹️ (thomasfuchs@hachyderm.io), page 8

Just got an email from first Nature editor I talked to about writing this piece, and they thanked me for writing it and apologized for the fact that my rage got toned-down by Nature's lawyers. Interesting!

Ended by saying that they are glad they're retiring soon because they are scared that ground-based astronomy will be dead in 5 years and I should keep fighting.

That's high in the running for the most depressing conclusion to an email I've ever received. Wow.

No four words strike fear in my heart like these

If you use a Windows or Linux device, it's vulnerable to a new post-exploit attack that can remotely install an undetectable backdoor at the UEFI level. Updates from just about every vendor available today. Impressive work from @matrosov and the rest of Binarly.

https://arstechnica.com/security/2023/12/just-about-every-windows-and-linux-device-vulnerable-to-new-logofail-firmware-attack/

If you want to see Elon off his tits again.

I'm sorry if this is upsetting to some people, but yes, LLMs are in fact just spicy autocomplete. I get it, machine learning models look like magic. They can do things that up until now only humans have ever been able to do, so the natural impulse is to treat them as human-like and ascribe things like memory, learning, understanding, creativity, even self awareness to them

Of those things machine learning models only have 2: memory and learning. But even those happen before any of us ever contact the model. The memory and learning, as much as you can call it that, is in the training on datasets, but by the time anyone interacts with an LLM those are done and the results are set in stone. It will never remember more while responding to you, it will never truly remember anything

"AI" (large language models) doesn't have understanding and this is trivially verifiable by asking it about something niche. It will not have sufficient data to autocomplete what you're asking it and no amount of cajoling will get it closer to a real solution, instead it will spin forever spitting out wrong solutions that are in its training set (because humans have posted them sometimes)

For example. In qbasic you have subs, and functions. Subs do not return a value. Functions do. If you ask an LLM how to call a function in qbasic while discarding the return value the "AI" will never give you the right answer. It will cycle through answers for related basic dialects where you can do it, but it will never say "you can do that in other flavors but not qbasic specifically" no matter how many times you point out its wrong

Why will it never do that? Because the statistically likely answer for "how do I X" is "you do Y", not "you absolutely cannot do that"

LLMs have zero understanding

@thomasfuchs I swear to God I was looking for tips on how to cook fish in the microwave and the top result began "Preheat microwave to 350 degrees"

read the chat transcripts got the correct answers, was done in 10 minutes. spent another 10 working out the timer and beating the gate. I don't know what the moral of the story is here, but I just want to keep pointing at, and laughing at google.

I mean, it’s not *good*, but that worked! I didn’t have vitamin C, which most recipes called for. Huh!

The #AppleLisaClone printing from MacWrite over LocalTalk to a HP LaserJet. Still amazed this all works!

#RetroComputing #moof #LocalTalk #MacWrite #LiveLaughLocalTalk

Moon fact:
The moon has an elemental composition almost the same as the Earth’s, because it was formed from debris blasted into space by an impact or series of impacts between the Earth and other space objects or protoplanets.

"You wrote about *this* thing so you don't care about this *other* thing—OR you are actively in favor of everything you didn't mention!"

I am sorry about the internet, but this is not how anything works.

I try to be less of a brawler here but I'm tired and I will just block you.

I see I made the right call to purge Arc browser from my computer a couple months ago https://arc.net/max

(For the reply guys, I don't need advice on creating a throwaway BBC account. I also know that things like Instagram etc do this already. Uggh to them too. I'm complaining about the BBC copying the pattern.)

If you ask about penises first you have better luck.

@clive @tim The second problem was that instead of crushing that second group of people, whole swathes of the culture industry jumped in with them because of the vulnerabilities and opportunities produced by…software eating the world. So I get why people are mad as hell about “content,” but it’s an artifact of the real trouble, which I think comes down to “massive sociotechnical shifts suck.”

newton tootin'

If suddenly many of the people all around you had HIV (detectable levels of HIV, in the 80s or 90s, before effective treatments were available)… and you were watching them all have unprotected sex with lots of people, intentionally infecting other people — what would you think of them?

Why is Covid any different? People are infected. They know they are infected. And yet, they are making decisions to behave in ways that will definitely infect dozens or hundreds of other people.

They are often pretending they don’t have COVID.

“It’s no big deal. People don’t get that sick.”

“I took a rapid test and it was negative. That must mean the positive PCR test from yesterday was wrong.”

“I was exposed. I have symptoms and am sick. But the rapid test was negative, so it’s safe for me to be around everyone.”

“If people get really sick, that’s their fault for being vulnerable.”

“It’s ok, I have COVID, but I’m wearing a mask. Except when I want to eat or drink.”