/tech/ - Technology

>>34065
>how fine the line really is between science and religion, between technology and idolatry.
much to ponder here, not for me mind you, i'm really stupid. but scientists think they have neatly compartmentalized science from religion, or even ideology in general for that matter and it turns out, not only have they not, they're insane zealots.

>>34063
>Making your model open weight would actually make it cheaper for you since some fraction (extremely small but whatever) of people will run it themselves on their RTX 6000s (consumer version of fucking Blackwells) instead of groyping your servers.
Yes but you also are, in essence, subsidizing training and research costs to your competitors and compute providers.

>>34068
>Yes but you also are, in essence, subsidizing training and research costs to your competitors and compute providers.

"in essence" is doing some very heavy lifting there. that's like saying flowers are subsidizing food costs to bees by allowing them to harvest their pollen. innovation is symbiotic, it requires free sharing and copying of ideas, otherwise it dies just like flowers die when there no bees to spread their pollen and bees die when there are no flowers to eat pollen from.

if the wright brothers had allowed other people to improve on their ideas to design and build better airplanes then they could have taken their improvements and used them to build better wright brand airplanes, but instead they just became patent trolls and sued anyone who tried to build an airplane. where is the wright company now? where are all the supersonic metal-body mono-wing jetliners with the wright logo on them? this is why things like capitalism and intellectual property are retarded.

>>34070
>"in essence" is doing some very heavy lifting there.
it is not. it describes what is happening.

>where is the wright company now
it merged with curtiss and exists as curtiss-wright today, it publicly trades as CW in the NYSE. i feel like you could have chosen a better example, maybe? you seem confused about what businesses do and seem to have deluded yourself into thinking high flyer might be above profit motive.

>>34070
>that's like saying flowers are subsidizing food costs to bees by allowing them to harvest their pollen. innovation is symbiotic
>this is why things like capitalism and intellectual property are retarded.
I don't want to samefag but your post is so baffling because it's the most egregious example I've seen this week of someone confusing what is vs what it ought to be. I feel there are more and more people like you on this board, posting these sort of confused tiktok comment section illiterate pabulum that adds nothing.

>>34072
>I feel there are more and more people like you on this board

yeah that's what happens, time passes, events unfold, old people leave, new people arrive, attitudes change, cultures evolve. all this dynamism is kind of the point of reading online discussion boards, if you prefer static text content then maybe you should unplug your router and read a word document or something.

>>34063
>RTX 6000s
You don't need an RTX 6000 to run open weights models. They come in all sizes and can be run on any computer. The smaller ones are just dumber.

>>34075
high level pabulum here

>>34076
I'm aware but I'm talking about the newest DeepSeek models here.

PSA: the only "non-corporate" AI cybersecurity scare report, which people pointed to as evidence that it's not all marketing stunts, turns out to be a sloppy work produced by "Effective Altruist" ideologues.

https://archive.is/20260812011323/https://www.telegraph.co.uk/gift/23731f653ebf2657

>However, the AISI had made some very strange decisions in this regard.
<It had disabled safeguards on purpose, left the systems attached to the internet and then let the AI software agents do their worst, while pretending not to notice.
>The AISI also allowed access to the Tor anonymous network, which let the AI access the dark web.
<“No enterprise network I have ever been on has allowed Tor. It is always disabled. That’s a big clue that this is a setup,” one expert told me.
>In response to the “rogue” event, the AISI produced its first ever “incident report” – formatted up to look and sound like a formal industry document.
<“It’s an ethically challenged publicity stunt,” says one highly experienced senior security expert who has led the development of global technical security standards.
>Nor is this the first time the quango has arguably tried to hoodwink us. When Anthropic sought help to promote Mythos as a dangerous cyberweapon earlier this year, AISI obligingly conducted a test on targets with their defences disabled.
<When Rishi Sunak hosted a summit on AI in 2023, the EAs quickly captured it and turned it into a vehicle for their hobby horse. “Three of the four partner organisations on the lineup are bankrolled by EA donors,” Politico UK reported at the time.
>One of their recommendations was the establishment of a new quango called the AI Safety Institute – the “Security” is a recent rebrand.
<Now a question: how do you think Jade Leung, the Prime Minister’s official AI advisor, reacted when she heard about the AISI’s latest amateurish stunt? Did she spell out the ways in which the testing fell short?
>That’s not likely, because Leung is not only No10’s official AI advisor but also the AISI’s chief technology officer – a clear conflict of interest.
<She is also a senior and long-standing member of the Effective Altruist movement. Leung received $180,000 in 2020 from the Centre for Effective Altruism, set up by William MacAskill, the EA founder, and then worked at his Future of Humanity Institute in Oxford.

>>34106
>next token prediction but marxistly
Waow. New frontiers in ML (no pun intended) theory.

i saw a video on titkok of some guy trying to cut small restaurants some slack because they use AI for their ads because "they don't have art departments" like chipotle or whatever, and the comments are 99% "small business suck", and small business owners desperately fending for their life. 5 years ago this would have been unthinkable, everyone would be like "yeah but nestle is worse" or whatever. chat, is luddite sentiment bad? leftypol led me to believe that it's reactionary but it seems that AI hatred has been extremely didactic for some people.

>>34157
Next time try get more than n=1 data points to make a statement on anything

>>34185
never implied to be a rigorous study, and in fact it's fairly obviously an anecdote, my dumb little retard, but it's more like at least 300 comments, plus the likes in the top comment are in the thousands, so it's not really n=1

It's a bit worrisome that a small army of journalists and tech influencers are preemptively trying to justify the OpenAI hack by saying that the AI is uncontrollably rogue,and adding to the mudpile a bunch of ultimately useless trivialities about AI agents leaving breadcrumbs in a makeshit "message board" (which by the looks of it was effectively being spammed with useless shit), even though it's clear that the breach could have been easily prevented through standard cybersecurity mitigations (see attached video). By the looks of it, they're trying to bury the lede: that OpenAI is, in practical terms, liable of whatever their bot does, because they are running hacking testing suites within an insecure sandbox. And it's still up in the air if they were merely incompetent, or if they were deliberately negligent to gather real-field penetration test data for training. And they're doing this because if there's precedent that a negligent actor is liable with what the bot does, then suddenly every fear of "AI replacing human workers" vanishes into thin air, because it would eventually lead to the most obvious outcome: that only people already in the field, or with basic knowledge of computers, can run agents connected to the internet in a safe way, and laymen cannot. Case in point: thus dumb trivial "hack", an AI booked tapped into an open cancel appointment endpoint: https://www.abc.net.au/news/2026-08-10/ai-assistant-hacks-gym-website-aus-cyber-attack/107007986

>>34157
I swear three decades ago it was completely normal for people running restaurants to design their own brochures (maybe with delegating some tasks to their teen kids). And ten years ago it was normal to use the internet to hire someone from a country with lower cost of living to get professional illustrations for 100 bucks or less. As far as I can tell you can still do that. People are becoming retarded.

>>34218
yeah there's a trend of generating things with AI that just don't need to be generated because they already exist and are free to use, like a map of US states with names. and i dunno, i guess the way it's being advertised as an incredibly transformative technology has led to some people understanding it as a quasi-magical technology, not because it generates stuff on demand, no siree, but like just the act itself of routinely using AI is being associated with being successful, like waking up at 5am to do nothing, or meditating. like if you keep using AI enough your brain will grow big or whatever.

<How accurate have Ed Zitron's AI skeptic predictions been?
https://danluu.com/zitron/
This is a big essay containing among other things a long list of false predictions (one or two of these I would let slide because they were too vague IMHO) and shows him grossly misrepresenting financial data.

>>34256
Despite the "disclaimer" in the intro, this guy is a heavy AI user which obviously means that he needs AI companies not to fail.

> Feb 2024: "I believe we're reaching the upper limits about what generative AI can do and how accurate its outputs can be."
< Wrong4
The footnote says that this is wrong because you can put the AI in a loop where it checks its output and tries again if it is garbage. This is not an improvement to AI, but a crude hack that might work well when generating code, but cannot be universally applied to ever field where AI is used.

Most of the other predictions do not have a justification for the "wrong" answer, he just claims that they are wrong, therefore I can't check them. There's no reason to trust this guy any more than that Zitron fella.

>>34256
Don't care about Ed specifically nor do i read him tbh, what annoys me is that discussions about the AI bubble always ends up being about his blog and specific claims.

What matters at the end of the day is that data center bonds & debt are currently garbage and keep getting worse.

>>34257
Yes, this is the same thing with "reasoning". It's a combination of backtracking + making the LLM retry the thing over and over again. There's no true "one-shotting" anymore and LLM labs now don't even have an incentive to improve the actual quality of inference.

(Btw this is why LLMs for code consume tokens like ravenous beasts compared to using them for RPing or something, since it's going through two layers of mechanisms that makes them retry shit constantly).

>>34257
>but a crude hack that might work well when generating code, but cannot be universally applied to ever field where AI is used.
it's also a thing that people "discovered" you were able to do since chatgpt3 and autogpt, it was just too expensive for frontier companies to offer as a product but now they can swallow the loss. it's so annoying that AI hypebeasts need to resort to manufacturing literally year's old history, because it's not that hard to build a case against ed zitron. nathan tankus already has a piece, albeit it holds its punches. ed zitron is not an economist or any sort of authority, his financial predictions are literally sourced from pseudonymous "quants" in substack, and he is only popular because he's going against the grain at a moment where every financial and tech space journalist is literally going fucking insane.

Both ChatGPT and Grok are down right now!

>>34263
Happens every other week during South Asian morning hours.

>>34262
One thing the Tankus piece forgot to mention in arguing against the characterisation of the bubble as similar to the GFC is that the data centre bonds & other financial instruments tied to the centres have had their yields increased and their ratings lowered gradually. This is in contrast to the GFC where the incriminating instruments were rated as top of the line until the very start of the crisis.

>>34263
heh whatever

>>34264 (samefag)
Ok this specific outage episode was funnier than the others because it cascaded. It started at Anthropic which lead to Grok being overwhelmed and going down by the sudden usage spike from Claude refugees, then the same thing happened to OAI from the combined Claude + Grok refugee wave and finally Gemini got raped.

Horde of addicts destroying everything in their path lol. Extremely funny how a lot of them can't survive without LLMs.

It's looking increasingly likely that OAI's newest hyped-up model is a dud, especially when compared to Anthropic's stuff.

It's not publicly out for regular people to get a practical feel for it though.

>>34268
>It's looking increasingly likely that OAI's newest hyped-up model is a dud
this has been the case for every new release since chatgpt 4.5

seit 3 tagen schaue ich mir wie ein besessener an, was die leute so mit gpt astra basteln. das ist alles so irre. https://x.com/bunkaich/status/2096126334628139221

>>34272
lol labs and mathematicians are fighting each other for who produced the solution first, amazing. this bodes well.

>>34274
ja fantastische, enthschuldigung. ich bin ein berliner.

>>34275
Yeah, it sucks if the original discoverers got screwed but it's undeniable now that we are getting new maths by cranking the AI machine.

https://vibemathed.com

All this has happened in the last 2-3 months and normies haven't caught up to the new reality yet.

>>34277
>by cranking the AI machine.
it's hard to tell if it's because we're cranking the AI machine super hard or if it's because there's untold amounts of money being poured into discovering new proofs to serve as milestones for literal AI ads. but i guess as long as the AI simpletons are happy I guess it doesn't matter

>>34279
>I guess it doesn't matter
It doesn't. New maths advances science and humanity even if it's marketing for OpenAI's IPO.

Not but seriously, this is the insane part: if you are any sort of research and use ChatGPT to help you validate your work, or really do any sort of brainstorming with ChatGPT, your work is not yours. That's the takeaway from this millennium prize kerfuffle. How is your boss not jumping from the rooftop upon discovering that, yes, OpenAI can lay claim to anything you do, or come up with, by stealing your own prompts, therefore rendering your business obsolete tomorrow is unknown to me, but I suppose the coin has not landed yet.

>>34281
es ist die nächste stufe der ausbeutung.

>>34277
Tbf a lot of stuff on that website is unreviewed garbage. It even has amateurish people who just spam their own shit on there.

I don't think these results are as impressive as they seem. It's kind of like the 4 colour theorem. They verified a result that everyone already assumed, so what did we learn? When something like this is solved by humans it is usually by some novel trick or combination and people learn from that, but in this case it seems they just power through the big and complex stuff? The limits of human intelligence is actually desirable here I tthink.

>>34286
There's probably at least one instance of an LLM discovereing something novel wrt maths out there, but this specifically isn't. The actual novel method that the LLM applied to Euler & N-S had only recently cropped up in the mathematical literature in 2023, used by Spanish researchers on a bunch of related problems.

>>34277
anything to not make the bubble pop lol

also Suno just wrecked their last model because muh licenses, any good chinese alternative?

https://infosec.exchange/@0xabad1dea/117239484057326836
Apparently most of the security bugs found by these tools are nothingburgers.

>>34296
I downloaded a local one from the comfy repo but it's absolute shite lol

>>34268
i don't know if it's a dud, but it's kind of weird how every model has a gimmick it wishes to excel at, and then they silently drop the ball on subsequent models, like chatgpt4 was programming, 4.5 was math, sol was cybersecurity, and now astra is 3d modelling. it's almost like they aren't generalizing and making them super competent at a field makes them less competent at other fields, but i dunno

>>34297
The counterargument is that it still found some nasty real vulnerabilities.
Yes, it was an extremely small number. Countable, even. But that is still better than nothing.

>>34302
If you're the type of person to obsessively check various benchmarking (I know they're mostly bullshit but whatever) websites you'd eventually notice a degradation in some random benchmarks after a new release from a vendor. For example, Qwen3.8 scores noticeably worse than Qwen3.6 on some less-important benchmarks. The recent Astra release by OAI famously revealed small degradations on two benchmarks that prevented it from completely topping the AA index. Fable 5.1 has a slightly higher hallucination rate than Fable 5. Etc.

It's like a mild version of catastrophic forgetting, where if a model doesn't constantly get RLed for something it would slowly start to blink on it.

>>34303 (samefag)
That said, it's still obvious that Project Glasswing was just pure grift from the start.