Welcome to November, and welcome to the Stubsack, your first port of call for learning fresh Awful youāll near-instantly regret.
Any awful.systems sub may be subsneered in this subthread, techtakes or no.
If your sneer seems higher quality than you thought, feel free to cutānāpaste it into its own post ā thereās no quota for posting and the bar really isnāt that high.
The post Xitter web has spawned soo many āesotericā right wing freaks, but thereās no appropriate sneer-space for them. Iām talking redscare-ish, reality challenged āculture criticsā who write about everything but understand nothing. Iām talking about reply-guys who make the same 6 tweets about the same 3 subjects. Theyāre inescapable at this point, yet I donāt see them mocked (as much as they should be)
Like, there was one dude a while back who insisted that women couldnāt be surgeons because they didnāt believe in the moon or in stars? I think each and every one of these guys is uniquely fucked up and if I canāt escape them, I would love to sneer at them.
(Credit and/or blame to David Gerard for starting this.)
Iām sure that treating people like a dog that was tricked by wrapping medicine in a bit of bologna will go over well.
If only the media would report it as āAltman is yapping some shit againā instead of breathlessly repeating his words like heās a holy prophet
(Death comes for XSLT)[https://dbushell.com/2026/11/03/xslt-rip/] (article about google killing XSLT, or gunning for it, not my usual fair but it is bad, and I didnāt see much attention pointed at it so far).
LW and the more credulous parts of the general techsphere are going gaga over this
https://rikiyat.github.io/blog/weight-space-programming/
Is this the long foretold, mythical RSI???
RSI = recursive self-improvement, not repetitive strain injury
He links to comments on the orange site which are pretty predictable
Wait will this mean that poisoning LLMs will be even harder to detect and prevent?
lol, I saw that going aflame among the promptfondlers nearest me, too
model auto-tune. theyāre getting hype for model auto-tune. I expect weāll little outside breathless tests of autotuned models for a few days
At least weāve moved past
classifiers that weāve had for 15 yearsJev
so thereās a new nber econometrics paper out, covering our favourite industry
it has some bangers in it
Paid adoption is substantially more common among younger and higher-income consumers: those earning more than $200,000 per year are 1.7 times as likely to pay as those earning less than $50,000, and those younger than 55 are 1.8 times as likely as those aged 55 or older
starting strong!
This force creates a dynamic incentive to set lower prices today: lower prices attract more users, improving future model quality and raising future demand
I donāt entirely follow how setting lower prices today improves future model quality but sure Iāll go along with the ride
Despite this limited horizontal differentiation, we find some degree of vertical differentiation across providers, consistent with observed market shares. At the end of our sample, the average consumer was willing to pay a $10 monthly premium for OpenAI products relative to Anthropic products. Part of this difference reflects model quality: our estimates indicate that consumers are willing to pay more for higher-quality models. For instance, moving from GPT-4ās capability level to GPT-5.5 Proās raises average willingness to pay for OpenAI products by approximately $25 per month
We also use the model to examine how consumer preferences for AI services shape competition. We begin by considering substitutability across services. By the end of our sample, all pairs of AI services are substitutes on average
ādo you want tea or coffee?ā āyesā
OpenAI has the highest retention of all AI products we study, with 77% of people still paying after one month, 65% after three months, and 55% after six months. Anthropic and xAI are lower, but similar to each other: around 65%, 50%, and 40% after one, three, and six months,
in the words of basement jaxx, down down downdown down downdown downdown
chat is it good when your world changing ai god keeps not getting picked for anyoneās team during lunch football
I donāt entirely follow how setting lower prices today improves future model quality but sure Iāll go along with the ride
More people using it, more training data fed into it.
77% of people still paying after one month
That seems a bit low.
More people using it, more training data fed into it
yes but thatās not actually the hard and costly part. the hard and costly part is the actual training action, which runs in the hundred of millions if not more
which, yāknow, when lower pricesā¦ā¦
subscriptions that are less sticky than a YMCA membership
this combined with the anthropic amended s-1 (which I very briefly dug into here, will dig more later), the numbers are all just about as bad as Iāve been thinking theyād be
Anthropic released amended IPO-aaand its even worse!
The artificial intelligence company said it generated $33.24 billion in revenue for the nine months ended Sept. 30, up roughly 1,200% from the same period last year. It reported a $104 billion net loss during the period, after losing $25.1 billion over the same stretch in 2025.
āb-but the convertible notes!!!ā
Remember the episode of DS9 where the authors agree that creating a holographic simulation of someone without their permission for sexual purposes is shameful? Ten years later in The Singularity is Near, Kurzweil seems to say that in the glorious virtual future this will be routine and accepted (its a section with a lot of dialogue between characters with years in their names, not digging out a copy and checking the page).
Does Kurzweil also have a bad take about people getting Tuvixād
Kurzweil actually liked the warp 10 lizard sex episode ā pass it on
Does he dance with the devil In The Pale Moonlight?
Hey wait a minute that episode was iconic for being actually good.
I find it so so funny that the Proton AI manifesto asks why the productivity is not increasing despite LLM usage, and then itās all likeā¦ youāre not prompting hard enough!
I take it as a positive sign when even the biggest AI pushers realize itās not working and turn to denial.
I am not a Proton customer. My impression is that they offer semi-secure/private email and VPN access.
What the fuck do they need GenAI for??
To extract more value from their engineers. The controversy is about the CEO having AI psychosis and wanting to forego stuff like pen testing to speed up feature delivery, not over anything customer facing.
Oh my god
So why, at Proton, does it not feel like the revolution has come? Why has productivity not increased dramatically if everyone is using LLMs all the time already?
The first is a change in engineering culture and engineer mentality. Your job has fundamentally changed. It is no longer to write code, and LLMs are not here to help you write better code faster. Your job is to help the LLM write better code faster.
The second reason we have not been able to take full advantage of the AI revolution is that we have been almost completely stuck at phase 1 ā using LLMs to speed up tasks as part of existing workflows. But writing code is not the only part of the software development lifecycle, and what happens when you eliminate a bottleneck? Another takes it is place.
The third reason is a collective action problem. Using LLMs effectively to write code is something an individual engineer can do for themselvesā¦ But to retool the company at large requires amounts of both organization and resourcing that we simply have not invested in yet.
Itās always vague talk about how they need to rearchitect their workflows to better integrate blah blah blah. Thereās never a specific model they want to move to or a specific plan or even a single detail about what this should look like other than āmore LLMs!ā
I propose the reason for this is that weāve already seen what this kind of model looks like with Gas Town or whatever heās calling it now and itās both blatantly stupid and incredibly expensive so any CTO who honestly tried to move a real company towards that model would laughed out of the room or sued for breach of fiduciary duty.
Why has productivity not increased dramatically if everyone is using LLMs all the time already?
Honestly, the manifesto should have stopped here. They admitted defeat right out of the gate. But the manifesto must continue, because the LLM cannot fail, it can only be failed. I wonder if this report is starting to get under their skin.
The first is a change in engineering culture and engineer mentality. Your job has fundamentally changed. It is no longer to write code, and LLMs are not here to help you write better code faster. Your job is to help the LLM write better code faster.
This is not even a meaningful distinction. If I use an IDE command to change the name of a method everywhere itās used, did I change all the names or did the IDE change all the names? What difference does it make?
I can only see this as them ordering engineers to use the AI more and go faster. The manifesto talks a good game about how āexpertiseā and āgood engineering practicesā are not dying, but telling engineers to offload more of their work to AI has the unavoidable tradeoff of making these things worse.
As a side question, how is it that all of these fuckers always want to use the precious gift of AI to go faster, not to, say, give their engineers more time to improve the quality of their software? Almost like they donāt actually give a shit about good engineering practices!
The second reason we have not been able to take full advantage of the AI revolution is that we have been almost completely stuck at phase 1 ā using LLMs to speed up tasks as part of existing workflows. But writing code is not the only part of the software development lifecycle, and what happens when you eliminate a bottleneck? Another takes it is place.
At least theyāre realizing that maybe one of the anti-AI talking points has some merit to it. But they donāt seem to realize why they need stuff like code review and CI/CD: the AI is unreliable and these processes are needed to reduce the probability of a major fuckup reaching production. All they think is, hey, more opportunities to use AI!
(āAnother takes it is place.ā I had to look at the original text to make sure. At least we know the CTO wrote this himself after many drinks at 3 am. Damn, why are all those people on the internet so mean to us? Iām gonna show them!)
The third reason is a collective action problem. Using LLMs effectively to write code is something an individual engineer can do for themselvesā¦ But to retool the company at large requires amounts of both organization and resourcing that we simply have not invested in yet.
What do I even say to this? Once again the problem is that there isnāt enough AI, and the solution is to buy more AI. Weāre gonna buy so much more AI that weāre retooling the entire company and shoving it up every bodily orifice.
If you know there are always new bottlenecks tha keep slowing things down why try at all? When do you go āwow none of this helped and we wasted a lot of energy on spinning our wheelsā
(Or perhaps you dont know what the actual bottlenecks are, a manager should know better).
Another entry in the weird āGPU worldā writing competition (What kind of wonderful happy go lucky future is brought about by LLMs like those of today being ubiquitous and constantly used but unaccountably not leading to the Great Singularity)
https://www.lesswrong.com/posts/siTsb48Wr5vWDGxLd/gpu-world-submission
This one is less messed up than the last. Instead of weird sex and graverobbing it featuresā¦ uh, a guy being paid by an LLM to braid string as busywork?
Get into the Cheesecake Factory or get left behind!
If you donāt eat cheesecake youāll be left behind, Luddite
Like it or not, the cheesecake is here to stay
I left the cheesecake in the fridge and it looks like it is trying to take over my fridge and creating bioweapons.
Please fund my cheescake institute.
And no obv this is too dangerous to let out of the fridge.
Please join me on the Cheesecake Truck
Man, Cheesecake no longer sounds like a real word and Iām real hungry now. Clearly the cheesecakes are too powerful and simply thinking about them or talking about them for too long will allow them to tempt even the most stout-hearted to open the fridge.
Will it crash if the white house gets de-Baghdad-ified?
I am just now putting together that this was a reference to obamaās recent remarks. Well, given how democrats usually operate, the cheesecake will continue thriving, and theyāll just make the white house look like a cheesecake factory designed by martha stewart
Iām looking it up and they have locations internationally, including in China. So I would say they are fairly regime tolerant. We just have to hope that Cheesecake Factory Alignment is solvable if not already solved
de-Baāathified. Baghdad (or any city) is larger than a single notorious denizen.
Jepsen releases a barely contained sneer https://jepsen.io/analyses/particledb-1.1.4
damn, I thought my sneers were on point, but clearly I must up my vip game. seems like boots optimized for rapid heel-turn vipping may be the meta
itās a sneer that recommends the users move off the product, and it was paid by the company that prompted the product into existence.
looking through the timeline of GitHub issues they had jepsen booked for a good long while too
I approve of their very helpful and comprehensive documentation for LLMs: https://jepsen.io/llm/index
looks like iocaine with a papers corpus (although I guess kyle couldāve made his own thing too)
My educated guess it is iocaine, as Kyleās been using it for other things too.
hah, nice. yeah the template layout immediately made me think that, but your word has far more weight :)
Itās quite something. Found deep in the discussion section:
Inserts ignored provided columns, creating rows best described as ādeeply cursedā Ā¹Ā²
- Jepsen encourages the consistency model research community to come up with more official terms for this.
