iiinova — where your digital self keeps evolving

Coraline !!!!!! @tech5l.bsky.social · 11m

This is a much scarier safety issue than effective altruist nonsense about super intelligence or RSI...

Coraline !!!!!! @tech5l.bsky.social · 14m

This is structurally inherent to the Transformer architecture and cannot ever be changed. The prescribing of a mind or "intelligence" to these things is very risky because people will start to trust it in undue matters.

100

Coraline !!!!!! @tech5l.bsky.social · 15m

Right. They have a weak understanding over residual stream layers, but this pertains to trying to figure out the connections of things (attention) and how these relate to statistical "knowledge" (the perceptron) in order to weight for next token (or multi token prediction). There's no mind.

110

Coraline !!!!!! @tech5l.bsky.social · 23m

I wish more of my peers understood this

000

Reposted by Coraline !!!!!!

Jak @jaksays.bsky.social · 5h

LLMs make the typing bit easy But the typing bit is the simple part of development - the hard parts are requirements gathering, problem analysis and verification, and as per this study those are parts LLMs actually make harder Those are also the only parts that add business value

2263

Coraline !!!!!! @tech5l.bsky.social · 36m

The attention mechanism is a cruel mistress

010

Coraline !!!!!! @tech5l.bsky.social · 40m

It's always the next model bro

000

Coraline !!!!!! @tech5l.bsky.social · 42m

But it's findings are extremely obvious to any SWE working in enterprise today lol. If anything it's much worse now.

000

Reposted by Coraline !!!!!!

Sean Barrett @nothings.bsky.social · 3h

2024: "This study doesn't include the newest models so its conclusions aren't valid." 2025: "This study doesn't include the newest models so its conclusions aren't valid." 2026: "This study doesn't include the newest models so its conclusions aren't valid."

1161

Coraline !!!!!! @tech5l.bsky.social · 46m

No answer I guess lol. Another vibe coder with zero computer science training...

010

Reposted by Coraline !!!!!!

Butlerian Jihadist of House Dornick @shimminykricket.blacksky.app · 8h

Basically AI allows software engineers to produce shitty code *faster* and in higher volumes which increases the amount time spent on code reviews and revisions nullifying the productivity gained from producing the code in the first place.

161267223

Coraline !!!!!! @tech5l.bsky.social · 1h

How am I being a prick. I'm a ML engineer and natural language transformers is my area of expertise. I'm trying to engage in the discussion earnestly

010

Coraline !!!!!! @tech5l.bsky.social · 1h

This benchmark doesn't measure hallucinations. It measures "correctness"

130

Coraline !!!!!! @tech5l.bsky.social · 1h

Not really a coherent question. Hallucination rate is softly correlated with model size but it's kind of tricky to measure to what extent or why. Some models achieve low hallucinations rate by using a refusal mechanism (Eg minimax)

100

Coraline !!!!!! @tech5l.bsky.social · 1h

Do you know what the difference is between an algorithmic process and a stochastic process? Lol

120

Coraline !!!!!! @tech5l.bsky.social · 1h

On here opus 5.5 has an hallucinations rate of 59%. What about that comes across as improvement to you? Do you think it's reliable if something is going to give you a wrong answer 59% of the time?

160

Coraline !!!!!! @tech5l.bsky.social · 1h

LLMs are not algorithmic in the strict sense lol.

100

Reposted by Coraline !!!!!!

Butlerian Jihadist of House Dornick @shimminykricket.blacksky.app · 8h

Harvard just dropped a study AI productivity in software engineering and it found the average increase in productivity from using Claude code is.....zero and for exactly the reasons you'd think fion.ac/jellyfish.pdf

3933561297

Reposted by Coraline !!!!!!

Emily M. Bender @emilymbender.bsky.social · 9h

Software engineering could and should be done with the same care as any other type of engineering. This is a post in honor of Margaret Hamilton (and also a subtweet of all vibe-coders everywhere).

Screencap

"“I fought to bring the software legitimacy so that it — and those building it — would be given its due respect, and thus I began to use the term ‘software engineering’ to distinguish it from hardware and other kinds of engineering, yet treat each type of engineering as part of the overall systems engineering process,” Hamilton told El Pais in 2018. “When I first started using this phrase, it was considered to be quite amusing. It was an ongoing joke for a long time. They liked to kid me about my radical ideas. Software eventually and necessarily gained the same respect as any other discipline.”"

Source: https://news.mit.edu/2026/margaret-hamilton-computing-pioneer-dies-1007

8662206

Coraline !!!!!! @tech5l.bsky.social · 9h

Even less so when everything is done by the chatbot and no one is checking anything

020

Coraline !!!!!! @tech5l.bsky.social · 9h

Lean? Not in the least bit.

120

Coraline !!!!!! @tech5l.bsky.social · 9h

This remains me of Mythos when we were being told it was so strong and powerful at hacking yet in reality it was absolutely nothing like it was sold to us. It's like that but in mathematics.

120

Coraline !!!!!! @tech5l.bsky.social · 9h

The problems discussed in the paper expand extremely when you're just fully automating the entire loop, deciding it's lean valid and throwing it on the pile of "solved problems". This is exactly what OpenAI are doing.

120

Coraline !!!!!! @tech5l.bsky.social · 11h

Oh my apologies. It's such an important time to be critical right now against this anti science onslaught from OpenAI!

140

Reposted by Coraline !!!!!!

Antonio E. Porreca 🐳 @aeporreca.org · 12h

The Association for Human Mathematics (AHM) statement on OpenAI’s October 6 release of mathematical documents just dropped! www.ahmath.org/statements

12300151

Reposted by Coraline !!!!!!

Kybosh @kybosh.bsky.social · 06/10/2026

I like seeing the dev of the SNES core on the Pocket actually telling one of these dudes off. Holy shit.

8 Closed
Experimental MSU-1 streami. #123
nasonliu wants to merge 2 commits into a
agg23 commented yesterday
If it works, that's pretty cool; I didn't think
would ever see this implemented (I also didn't
expect LLMs to exist).
However, this is not something you can PR
here, or anywhere. You sent me a nearly 50k
line diff, unasked for, completely LLM built, and
you somehow expect me to review it. There is
SO MUCH unrelated garbage here. Never
minding all of the docs and Python, you've
somehow touched basically all of the upstream
HDL, which already had MSU1 support, so idk
what in the world you're doing.
This kind of behavior is not acceptable. You are
infringing on the time and brainpower of people
by pushing "your" content that little/no of your
brainpower went into; completely unrequested.
I hope for the sake of all responsible
maintainers that you do not do such a thing
again.
Again, I think it's cool that you and your LLM
supposedly got this to work. It has no place in
my world.

4789542167

Coraline !!!!!! @tech5l.bsky.social · 07/10/2026

I hate them and the whole slop crusade they're doing in math which ultimately isn't contributing to human understanding in math or helping develop new directions. It's just clankers clanking and OpenAI desperately trying to save themselves from financial oblivion. Not to mention the theft

010

Coraline !!!!!! @tech5l.bsky.social · 07/10/2026

Fuck scientific standards we're vibe mathing

070

Coraline !!!!!! @tech5l.bsky.social · 07/10/2026

This is accurate but consider: pathetic attempt to prove your company is worth trillions as its financial future is extremely dubious

041

Coraline !!!!!! @tech5l.bsky.social · 07/10/2026

Lol I mean they're definitely not checking the value of the proofs in any depth. They're fully reliant on Lean

030

Coraline !!!!!! @tech5l.bsky.social · 07/10/2026

Before all this math slop l was actually a lean enjoyer but we've got so much relatively or completely useless AI Math results now because "lean valid" that I'm really not sure anymore that lean valid carries any substantial weight as to the ontological value of a proof

070

Coraline !!!!!! @tech5l.bsky.social · 06/10/2026

It really exposes the extremely deluded bubble they live in. Mfs actually believe LLMs are emerging deity like structures. They're so delusional

010

Coraline !!!!!! @tech5l.bsky.social · 06/10/2026

It's a little more recent but you could also argue very strongly that since the 2000s silicon valley culture around AI has effectively been Sci fi eugenics

000

Coraline !!!!!! @tech5l.bsky.social · 06/10/2026

As a ML person who's extremely sick of the silicon valley cultism, the horrible state of ML research and the ridiculous bubble around AI, thank you so much for writing this. I feel heard

010

Reposted by Coraline !!!!!!

Mike Cook @mtrc.bsky.social · 22/09/2026

New Post: Why I Love AI Next month it'll be twenty years since I went to university, learned to program, and studied AI. I wrote about why I loved it, why I still love it, and why, unfortunately, it no longer exists. www.possibilityspace.org/blog/posts/i...

32558165

Coraline !!!!!! @tech5l.bsky.social · 04/10/2026

Concerning that this needs to be stated at all. ELIZA effect has cooked 99% of discourse around NLP Transformers

000

Reposted by Coraline !!!!!!

Timnit Gebru @timnitgebru.blacksky.app · 03/10/2026

Even more depressing than the knowledge that congress is now overrun by effective altruist and overall TESCREAL cult member staffers, is the number of billionaires in these cults that are going to be created by the OpenAI and Anthropic IPOs.

619448

Coraline !!!!!! @tech5l.bsky.social · 04/10/2026

I am sceptical Yud knows any of these things to any real depth lel

050

Coraline !!!!!! @tech5l.bsky.social · 04/10/2026

This is much more reasonable to call brute forcing than "intelligence". The question of if LLMs "understand" is tricky. I would argue they only understand in that they have pairwise QKV clusters which can weight their probability distributions but they don't have an ontological sense of mind.

000

Coraline !!!!!! @tech5l.bsky.social · 04/10/2026

Well also you can brute force over lean and in theory use a soft form of Evolutionary algorithms. This is probably what they did for NS aside from also stealing the problem space lead from mathematicians. You can run millions of instances with lean and kill processes that eventually hit a wall

100