Truly, the ability to create is wasted on us
Text
You know we actually had agi in the 1700s, they were called, well I can’t say the word, but they got recognised as people and are now colloquially referred to as black. The US had a whole ass civil war about it. AI companies just want slaves without the social and legal implications
Text
Use a chatbot to plan your day? Replacing it with a tarot deck might help.
new essay on substack, in which I explain why I occasionally use tarot to make decisions - even though I do not believe in tarot or any kind of divination
Text
The burning question:
How stupid is an AI model running locally compared to AI models running on the supercluster servers of Big Tech companies?
Text
Don’t be naive. You’ll never have an AI model running locally on your machine that’s as intelligent as an AI model running on a Big Tech super server. Big Tech companies don’t even disclose the number of parameters in their proprietary models.
Quote
LLMs are trained on violent content. They are trained on child pornography. They are trained on racism, misogyny, homophobia, ableism, nationalism, anti-Semitism – they are trained on the Internet. The well is poisoned, and while they’ve erected a sign to say “warning!” and maybe even a little fence (mostly so they can’t be sued), that doesn’t change that the waters remain poisonous.
Text
Smart Has Been Defeated by Foolish
I have only two questions:
1. How has posting LLM slop become the new influencer flex?
2. If LLM slop is the new flex; can we ban all influencers who do it from the internet?
Text
Not AGI Touted by People Who Don’t Understand
I appreciate how the definition of AGI has been reduced down to be almost meaningless here. OpenAI likes to do that. They want to make sure that their statistical chatbots might be able to be defined as AGI someday—doesn’t matter that AGI doesn’t really mean general intelligence in their definition.
Also, how is that linked Git project AGI? I can’t sleep because I am so shocked by how dumb this definition of AGI is.
I appreciate how this guy thinks counting goes 1,2,5,3,4.
“Level 3: Agents: Systems that can act autonomously on behalf of a user over extended periods, making decisions without constant oversight.”
To say we have reached level 3 is pretty generous. LLMs mess up a lot. If you like having your stuff broken I guess we have. To say that LLMs are reliable for this application is just a woefully silly thing to say.
Honestly given that LLMs don’t actually reason I don’t think we have reached level 2. Again boosters see what the LLM slop salesmen tell them to see. To reason you need actual concepts and the ability to symbolicate. LLMs can’t do either since they operate at an only a semantic level.
“Level 4: Innovators: AI that can aid in invention and generate new, original ideas or solutions.”
We aren’t even in the neighborhood of this. Remember this author thinks he saw AGI. I ask how can something innovative arrive from the collection of the most likely words? If I take the average of the average how will that end up being extraordinarily? This is just typical LLM hype slop trotted out by the boosters who don’t understand what the hell is going on.
An LLM has never done better than to plagiarize the work of a human or to plagiarize the work of another bot that plagiarized a human. This is just such anti intellectual nonsense to think that an LLM can innovate.
For anyone concerned—no AGI isn’t here. This guy talking about it has mistaken stupid for smart and truth for obvious lies.
Text
Toward Fully Verifiable, Reproducible AI
A case for applying the Bitcoin and Free Software Ethos’ — Full Transparency and “Rules Without Rulers” — to Language Model Development.
The Problem Nobody Is Naming Clearly
The AI field has a trust problem, and it’s not the one most people are talking about.
The common framing is: can we trust what AI says? But there’s a deeper, structural question underneath it: can we trust how AI was…
Text
If Only This Actually Worked
So if LLMs actually worked well enough to be worth this, then this would be a great area of endeavor. I love the idea of local, and given who you are giving your data to otherwise–it’s a no brainer trying to avoid it. I just don’t think LLMs actually merit the effort–especially given the universal downsides to any LLM.
“Current capability Research fetches full page content from Reddit and seller blogs, synthesizes without fabricating prices. Coding handles full-stack across Python, TS, Go. Takeaway Orchestration, routing, and memory matter more than raw model size. The right quantization (Q3_K_L vs Q4) can change throughput without a big quality drop. Still experimenting Retrieval quality, model selection heuristics, timeout/retry for long tasks.”
I just have trouble imagining how this can possibly be true given the model sizes are all 30B or below. I guess with a sufficient level of context control maybe, but this must be delicate and slow as anything.
Coding with such small models is virtually useless; given the quality of the code. For the rest of the tasks you’d have to chunk them in to such small bits to have any hope of accuracy. The author is using a 9B parameter model of research–which even using RAG is going to be pretty poor. Nemotron 3 Nano 30B is not especially great at RAG, with nowhere close to 100% accuracy. I guess if your questions are simple enough maybe, but then it begs the question why bother with this LLM noise?
Text
Text
Text
What’s the etiquette when you spot AI art on Tumblr? Just keep scrolling?
In general I feel like AI use should be challenged and discouraged by default, so that we can develop a social norm against using it. But accusing someone’s art of being AI is dicey, right? Like, if that is in fact their real art then imagine how devastating it would be to have someone look at your hard work and say that a machine made it.
I don’t know what to do about this.
Text
On AI-generated writing
epistemic status: thinking out loud
I recently read a fairly long nonfiction post that was written using AI. I didn’t know this when I began reading it (or I wouldn’t have read it) and didn’t pick up on it being AI from any of the content; I only know because the author explicitly said so. It used a bunch of their ideas and was mostly a tool for organization and polish rather than for original thought, but at the end of the day it wasn’t really their words.
And what happened when I read the admission was fascinating. I was angry. I had liked the post quite a lot and now I felt cheated except… how had I been cheated? The quality of the writing didn’t change because I knew it was AI-generated, although doubtless now if I went back to it I’d be able to find little nitpicks that I’d never care about normally. The quality of the message didn’t change because it was AI-generated; the author had carefully edited and removed any inaccuracies or similar.
So why the anger? This doesn’t feel to me like a rational response. My stated objection when questioned on LLM writing has always been that it’s unreliable, not any belief in the essential soulfulness of human writing, especially writing meant to convey a specific message.
If you write something using a LLM and edit it and factcheck it you can produce something that feels entirely like normal human writing.
But again, if I’d known it was AI beforehand I wouldn’t have read it. Now that I know that it was I’m less inclined to bring up its ideas to other people, and there’s a zero chance I’ll reblog it. Even though nothing about the actual content changes from the knowledge that it was AI-generated.
At the end of the day, I would have preferred a less polished, less coherent post to the one I read, and that strikes me as entirely absurd, since the point of the post was to make an argument, not to tell a story, evoke an emotion, or similar.
I assume the author used AI because they weren’t confident in their own ability to write a strong, coherent, logically-argued post on their own (they commented that with editing it took longer than it would have to just write out a stream of consciousness post, which I would, absurdly, have preferred). And since they’re not someone I think I’ve ever encountered before, I don’t know whether this is an accurate assessment.
There still is something to the “if you couldn’t bother to write it why should I bother to read it?” argument, but although the author didn’t bother to write it, they did bother to edit it, to correct it, shape it, etc. And of course the initial ideas were theirs. This wasn’t a zero-effort post, and to say that it was the wrong effort is a value judgment I don’t feel justified in making, although I do make it.
This post wasn’t written with AI, but if I took it and plugged it into a LLM and asked it to construct a more coherent, well-argued piece it probably could have. And that writing might have been more useful than this stream of consciousness. But at the end of the day, I still prefer this post.
Text
Ájtís skacok, szevasztok!
Közszolg!
Amúgy meg búmer-alert, meg abszolút hozzá nem értő-alert, csak hogy tudjátok honnan indulok…
Adott egy bazi szenzitív meló: kutatási projektek etikai elbírálása. Fel kell tölteni egy platformra (2 faktoros autentikáció után) minimum 3-féle doksit, de sokszor több minden kell. Az elbírálás többszintű:
1. megvan-e minden doksi
- az összes, ami kell, ami ugye nem mindig 3;
- az alkalmazott templét, ez alapesetben ez 1-féle, de lehet több is;
- alá van-e írva
2. Ha már átnézzük a doksikat az első körhöz, a tartalmát is ellenőrizzük
- koherensen van-e minden doksi kitöltve
- a választott kutatási terv elemei etikusak-e és megfelelnek e a GDPR-nak.
Az esetek 72%-ában visszaküldjük a cuccot javításra, mert tartalmilag nem oké. De sokszor hiányzik egy doksi vagy nem olyan, mint kéne vagy nem annyi az informed consent form, amennyinek lennie kéne (1, pedig külön kell minden kutatási aktivitáshoz 1, meg ha más a target group).
Ez azt jelenti, hogy
1. 1,7x annyit melózunk egy kérelemmel, mint ideális lenne, a többi kérelem addig felhalmozódik, gyűlik az át nem nézettek listája
2. Szar az anyag, amivel dolgozunk, fel kéne javítani a beadott kérelmek minőségét, hogy hatékonyabban menjen ez. Persze lusták a diákok, merthogy az ő kérelmeiket nézzük, nem olvassák át azt, hogy mit kell beadni, de őszintén a témavezetők sem segítenek sokat, és a kutatásmódszertan órán csak én foglalkozom ezzel a témával, a másik 2-300 diák csak akkor látja ezt, mikor szakdolgozni kezd.
3. De mivel 1. pont, nincs időnk a 2. ponton dolgozni. Ezért most kérünk még manpowert, hogy foglalkozzanak az 1.ponttal, mi meg a ketteskén.
Na erre a férjem aszonta, hogy hülyék meg bambák meg búmerek vagyunk, adjuk oda az AI-nak a melót.
De én nem hiszem, hogy a külön platformra (Labservant a neve amúgy, elég gáz platform, de erre az etikai cuccra lett kifejlesztve) feltöltött anyagokhoz okés, hogy hozzáférjen valami kommersz LLM (nem saját fejlesztésű, belsős machine learning cucc, az más történet lenne), mivel ezeken nevek, cégnevek, GDPR-szenzitív adatok vannak.
De lécci mondjátok meg, hogy van-e olyan LLM, ami megbízható? Tehát nem haluzik, nem adja át az adatokat 3. félnek, stb., ami át tudná venni ezt a feladatot.
Kapcsolódó kérdés: van-e olyan megbízható LLM, amivel kutatási anyagot lehet pontosan transzkribálni (audio -> text)?