JPEG XL

AFAIK it's a VR optional title with singleplayer or multiplayer cross/'platform' (VR and flatscreen)

I'm mostly curious if it'll have JXL support like the deck

Since it runs SteamOS I guess it should

Since it runs KDE and IIRC KDE defaults to wanting JXL, it probably has JXL.

Now if only they'd use it instead of Lossless AVIF for screenshots

what were the conclusions? i still can't find any of that

some wins some lossless, but box86/64 is generally faster and easier to setup

Has anyone tried using the new affinity software by canva

Everytime I try to use the object selection tool on a JPX photograph it just go wack and freezes up

anyway the softwares cool im so glad its free

Interesting. Just got this on the Android app

sus

congratulations, you have been a u g m e n t e d

Is this the meme channel

no rules prohibit the proliferation of graphical or literary memetic devices in this here chatroom

Oh god, I updated my desktop client too and now <#1019989424232222791> threads have *momentum* while scrolling the channel list, even though the others don't so it just looks awful

which flag is this

it's a parody of this image

I'm not sure microsoft approves of that flag

they've been annexing gaming companies all over

you commonly see it when people talk about the "big three" gaming companies

and now, since Xbox the hardware platform is failing fast even as Xbox the publisher is doing OK, if the new Steam Machine does even a little well it might just kill the Xbox.

guess who got to take 40 TB of LTO tape cartridges AND an LTO-5 tape drive home from work <:Hypers:808826266060193874>

on the condition that I wiped all of the tapes which took...months

now I need something to plug the PCIE host bus adapter into, otherwise I won't be able to interface with the drive

nice

or Framework needs to release a Serial Attached SCSI expansion card <:KekDog:805390049033191445>

it's old tech, LTO-5 came out in 2010 but most of these tapes are LTO-4 which came out in 2007. but after writing these tapes last up to 30 years in good conditions, in terms of data integrity, so they're really nice for archival <:BlobYay:806132268186861619>

could be done over a thunderbolt to pcie lmao

hmm

for some reason I expect those to be more expensive than a second hand PC but that would be a lot more convenient for sure

Perfect for storing at friends and families houses for offline offsite backup

yeah that's a good point

I've offered to store some of their backups too ^^

sounds good

most people don't have backups

I've never messed with tape drives before. is there any special filesystem or file formats you need to use with them or do you just interface with them as like a very slow storage drive?

they can work normally IIRC, but are very slow for random I/O

can be very fast for sequential though

some get to 300MB/s, faster than many HDDs

`tar` was often used to prepare archives for tape storage

``` GNU tar is an archiving program designed to store multiple files in a single file (an archive), and to manipulate such archives. The archive can be either a regular file or a device (e.g. a tape drive, hence the name of the program, which stands for tape archiver), which can be located either on the local or on a remote machine. ```

speaking of, this seems pretty nice: https://www.nongnu.org/lzip/tarlz.html

in fact that’s where it got its name from (**t**ape **ar**chive)

up to LTO-4 they only have a concept of 'files'. you go to the *end of data* then write a new file and an *end of file* marker. you need to remember which file represents which data, externally, there's no index on the tape itself. you can then ask it "give me the fifth file on this tape" and it'll start reading the whole tape until it gets to the fourth end of file marker, and it'll give you the data from there until the end of file marker LTO-5 and newer do have a file system called LTFS which allows you to actually mount the tape as a file system available to your computer. it's still just a linear tape so deleting files doesn't actually delete them (just remove them from the index) and new files are always appended to the end of data

<@&807636211489177661>

Can we check what invite link they're using to join? They might be using the same one

at least it's fully opt in

https://www.threads.com/@dblasphemy/post/DRJAcGNjLYq

Global Cloudflare outage

Finally fixed

https://www.tomshardware.com/tech-industry/artificial-intelligence/elon-musk-wants-foundry-partners-to-build-100-200-billion-ai-chips-per-year-musk-says-chipmaking-industry-cant-deliver-on-his-goals These clowns will be waking up to an extremely harsh reality lol. Peak delulu shit.

ah, yes, 12 to 25 "AI" chips per year per person, because surely all that has a market

> Our first goal is to prove that 2 + 2 = 4. Next we'll prove that x + y = y + x. And at the end we'll see if we can prove Fermat's Last Theorem. What a gentle introduction this tutorial is.

is that that Lean tutorial? I think I’ve done it (minus the Fermat part, which they don’t actually expect you to do)

a few days ago I heard on the news that some bank company here in the netherlands thought it was "worrying" that the netherlands as a country is not investing enough in AI technology and businesses, fearing we'll be left behind in the market I thought, good, we'll have an economy left when the bubble bursts <:KekDog:805390049033191445>

yeah

https://arstechnica.com/ai/2025/11/tech-giants-pour-billions-into-anthropic-as-circular-ai-investments-roll-on/ People have gone insane.

Blender 5.0 released, now with HDR color spaces https://www.blender.org/download/releases/5-0/

So recently we've got GPT-5.1, Grok 4.1 and Gemini 3.0

banger

what is that image in the bottom right??

https://tenor.com/view/shark-cable-fiber-bite-gif-16468772

Shark biting one of the undersea cables

oh lol

Blender has already had some HDR for a while in Kronos+AgX+Filmic mapping and P3+Rec.1886+Rec.2020 displays for a while now (Filmic for as long as I have used the program), but this update brings Rec.2100+ACES1.3/2.0, more mapping on HDR displays, and variable working spaces

https://cdn.discordapp.com/attachments/803196804651941959/1441034599043170355/RDT_20251120_1850557836450280439316956.jpg?ex=6920538d&is=691f020d&hm=53ec1301abbad48cb01a1a49006b1699ceef64ebc756b42949d39ee90d7b8598&

funnily enough it feels like the unpaid open source devs are more reliable than aws and cloudflare

most of modern digital infrastructure is held together with duct tape and spit, the proprietary stuff even more so than the open source stuff

https://tenor.com/view/skinny-homer-skinny-homer-homer-skinny-gif-26715248

propietary software is just better at hiding the kinks

Someone talk if US gov can control TOR of course yes they can, look at cloudflare

it doesn't really work like that

the tor network is much more decentralized than cloudflare's network, and consensus over the state of the network is sort of spread out between a handful of consensus nodes run by various digital rights organizations in different countries

then even if they forced them to report a consensus that for example had 500 volunteer-run relays removed, we would all see that since the consensus information is public, it's the same information tor clients use to create a circuit

Cloudflare is all networking traffic routed through a single company TOR is personal traffic routed through at least 3 volunteer run nodes, changing to others with the press of a button They're about as far opposites as you can get

I used to run a non-exit tor relay at one point. unfortunately many services incorrectly block all relays instead of just exist relays so I couldn't access some sites from my home network anymore

ideally they don't block tor anyway, but you know

https://arstechnica.com/ai/2025/11/were-in-an-llm-bubble-hugging-face-ceo-says-but-not-an-ai-one/

this would be a somewhat agreeable position if there was something more popular done with not-LLMs than media generation, which I feel has basically the same issues as LLMs

There have long been a whole bunch of uses for machine learning stuff and I don't see any increased uptake in those (and feel like sometimes those are being discarded for LLMs and generative stuff)

STT and TTS say hello

the frontpage of the English Wikipedia is shining some light on queer people today <:BlobYay:806132268186861619>

https://github.com/ocaml/ocaml/pull/14369#issuecomment-3556593972 bit of OCaml drama

(LLM-induced)

https://cdn.discordapp.com/attachments/530853682678595594/1441297293797228666/image.png?ex=69214834&is=691ff6b4&hm=7fc18f34a0f735c79d0841a462476532fcf1f75359a278b1ede748eb13c9f1e9&

Ah yes, the JXL reference display https://youtu.be/vDWQHZ4Fy6I

https://cdn.discordapp.com/attachments/481168746787110913/1441461940596506725/image.webp?ex=6921e18b&is=6920900b&hm=9c0f7cfa9f6ff8f8347cd7e039bf8b78aff0f5b14751e4c344ecb974d8464f2a

it keeps evolving

look at the lil ferris

left-pad ?

what's that

a Node package that everyone depended on until its developer pulled it from npm and npm had to put it back to unbreak everything

it would pad a string up to the desired width

https://en.wikipedia.org/wiki/Npm_left-pad_incident

thanks

it's for inserting the right amount of spaces before a string so the end of the string ends up at a certain position

which is what, four lines of code?

the Wikipedia article contains this screenshot of “Contents of `left-pad` at the time of its npm removal”:

(I personally think the developer’s response to kik’s initial e-mail was rather unreasonable)

oh god that reminds me, iirc on crates.io there are the `is_equal` and `is_odd` crates, which work by pulling in the other and taking the opposite of it. so if you try to include either it won't work because of a circular dependency

might be an ironic nod to left-pad

it is yeah

or rather the equivalent packages on npmjs

there’s also `right-pad`, but apparently both are deprecated

it seems plains JS has `.padStart` and `.padEnd` now

this is probably what it's referencing

interestingly, going by download counts, people seem a lot more interested in odd numbers than even numbers

which is odd

no, bad pun

this is not the case anymore

I think it was at one point but I could be misremembering. it might have been a different pair of crates

https://youtu.be/zr-eZ_pLTNI

This probably won't interest many people, but I found an (old) animated Google logo GIF that's only accessible over IPv6: https://ipv6.google.com/images/ipv6_logo.gif

https://www.reuters.com/world/china/china-takes-spat-with-japan-over-taiwan-un-vows-defend-itself-2025-11-22/ God this does not end. Now I'm afraid it will affect every anime-style gacha game from China.

> Sanae Takaichi, unlike all previous prime ministers of Japan, really treats Taiwan seriously and respects President Donald J. Trump. This triggers the EVIL true face of Communist China that wants to conquers Taiwan regardless bloods and tears, and the government has claimed that someone should behead Takaichi by any cost. Crazy dogs (so-called "Wolf Warriors") that grew in mainland China and brainwashed by the party led by President Winnie Xi the Pooh can't NEVER be tamed!

Also the same PM that is a US puppet.

What was the need for the remark even?

https://tenor.com/view/nothing-ever-happens-chud-chudjak-soyjak-90-seconds-to-nothing-gif-9277709574191520604

I mean there's already leaks of Honkai Star Rail altering their next destination because of this bullshit.

That's China's political correctness

Yeah, but there's rumours about not letting Ningning from girl group Aespa perform in NHK's new year event from Japan's side.

It's going both ways.

And there's no one to handle this.

China don't even admit there is an UN member is invaded Now want to ask "justice" at UN?😂

https://www.bloomberg.com/news/articles/2025-11-22/japan-blasts-china-s-un-letter-claim-as-baseless-urges-dialogue

About that...

Seems like they're still not willing to at least have the PM retract the remarks lol.

Tone deaf.

https://en.wikipedia.org/wiki/China%27s_final_warning

Well they cancelled Japanese events in China left and right so not much faith they will stop there.

"welcome to introduction to real analysis. this course is still under development. if you notice any mistakes, extra points if you send a PR. if you don't know what a PR is, we'll go over it later." Basic math courses at universities will soon become programming courses

https://bsky.app/profile/vuc.bsky.social/post/3m636ghnab22k

Does anyone know of a good ffmpeg wrapper (like shutter encoder) for android?

the DJ is actually just a super cool guy

from his YT

> Father Guilherme believes that electronic music is a privileged way to build a better world. On the dance-floor, inclusion, tolerance and respect are promoted. In electronic music, those who have faith and those who don't, dance together. People of different ethnicities and cultures dance together. Music brings humanity together, in its beautiful diversity.

So China is upset if other countries say they would intervene and help Taiwan if Beijing were to invade and start murdering Taiwanese

The Communist Party has killed more of its own subjects than any government in the history of planet earth and everyone acts as if they are a normal and legitimate nation

It's really disgusts me that everyone treats with them and trades with them rather than putting them in the same category as other failed states like Nazi Germany or North Korea

Meanwhile it was unwise for the perpetrator of the Nanjing Massacre to like openly say that they can defend Taiwan. And now cultural exchanges between China and Japan are at risk of collapsing completely.

https://www.straitstimes.com/asia/east-asia/china-says-japan-sent-shocking-wrong-signal-on-taiwan

Before Japan massacre China, there was China (同盟會) betray Japan (黑龍會) After USA nuke Japan, then [USA bio-attack China](https://web.archive.org/web/20240417052006/https://www.ecohealthalliance.org/wp-content/uploads/2024/04/Emails-02.pdf) (or [China bio-attack USA](https://www.justice.gov/usao-edca/pr/operator-central-california-bio-lab-indicted-distributing-adulterated-and-misbranded), depend on your position)

Seems not many people knowing the definition of "China" throughout history

Wasn't Taiwan known as Formosa island before the rebels fled to it?

Which rebels

The Chinese rebels?

This is too ambiguous

what's the use case you need, there are a few but they are generally specialized uses.

In general, I would say the majority of the time you're better off just using your own native bindings.

I mean there was a time before Taiwan was called Taiwan. It wasn't always known by that name. Something gave it the current name.

The name has been used since Qing Dynasty

Ah makes sense.

It's just that Formosa was used more frequently internationally

Defending Taiwan from Chinese aggression is always morally correct, though saying so may be dangerous in the world of realpolitik.

Mostly want a gui to convert and compress videos because my phone always saves them in ridiculously high quality I found this https://github.com/fj-gruenewald/eazyffmpeg-android-app?tab=readme-ov-file but it doesn't seem to work on modern android, and there's one app on the play store that just was unusable due to bugs. Wondering if you guys know of any options

Which android fork has a bypass charging at OS level (not per app)? if smartphone supports bypass charging

~~termux-gui~~ I would probably just bundle ffmpeg and pass commands right to it tbh

did you know? If you have an amateur radio license, you can send text messages to the ISS. They have an auto-responder that'll tell you when the next time the ISS will fly over you is, if you include your GPS coordinates in the message.

that's neat

https://www.gmanetwork.com/news/topstories/world/967248/china-criticizes-japan-s-plan-to-deploy-missiles-on-island-near-taiwan/story/ Man they did not think through this twice before settling on Yonaguni coast just 110km of Taiwan.

I wonder who's missiles they will use, not sure if Japan has significant domestic production of them

https://www.tomshardware.com/tech-industry/artificial-intelligence/insurers-move-to-limit-ai-liability-as-multi-billion-dollar-risks-emerge Insurance companies are bailing out from AI nonsense as well.

> Known as the IBM Quantum System One, it is the first circuit-based cornmeal quantum computer dyslexia makes for incredibly funny misreads sometimes

Lol

Wonder when they're planning to build a quantum computer that can factor the number 21

https://en.wikipedia.org/wiki/PLUR

I think I broke grok

surely JXL has lapped transforms

we considered it at some point 😛

abandoned in favor of gaborish I imagine :D

I read that gaborish achieves a similar effect while being much cheaper to decode

Finally got a new VPS set up to run an IRC bouncer on

My more than a decade of backed up IRC history is taking a while to copy to it

JPEG XR uses lapped transforms

JXL doesn't

<https://scp-wiki.wikidot.com/scp-1313> > Description: SCP-1313 is an anomalous series of logical processes, capable of being defined as a mathematical equation to which the answer is a single female specimen of Ursus arctos. The equation itself does not appear to be inherently anomalous, but rather a quirk of mathematics — rather than producing any number in R\A (the set of all real numbers that are not animals), SCP-1313 resolves to produce a tangible, adult, and frequently enraged grizzly bear. The resolution does not have to be physically represented, and simply considering the problem long enough to arrive at the solution has resulted in a bear's manifestation in or around the subject solving it.

Huh, Caddy only supports Gzip and Zstd, not Brotli. Wonder why that is.

Ah, apparently they used to include it, but they thought it was way too CPU heavy for on the fly compression, so they recommend pre-compression for Brotli. And Caddy does support redirecting to Brotli files when the request header indicates support.

Brotli is too CPU heavy for real-time compression? Unless I'm missing something, with the right settings it can achieve similar speeds to DEFLATE while being smaller.

there’s a plugin for it https://github.com/ueffel/caddy-brotli

but yeah, static assets can use precompression

https://caddy.community/t/brotli-on-the-fly-v2/19796

https://x.com/punilive_holo/status/1993593896518234483?t=TXDttDepblg8uxfYvH6qfg&s=19

The letters look diagonal but they aren't.

My kindle touch got JXL support before chrome 😤

It's jailbroken, I decided to try displaying to the framebuffer manually, wanted to author images on my laptop so I needed an image format for the transfer

https://qr.ae/pCMcwu floating-point no-op :D

I'm 90% sure it's a NOP for 8087, so as it's running on a separate chip it might have some logic in existing (though x86 has 9 different "normal" no-ops ranging from 1 to 9 bytes in size)

Neural network LLM architectures displaying all layers they contain, with respect to actual sizes: llama llama3 mistral mistral-small-3 mixtral qwen3 qwen3-moe The difference between models of the same architecture, 4b and 12b, is only in the number of blocks; 12b has three times as many than 4b, identical blocks, that's all.

https://x.com/jeremyphoward/status/1994532591685570942

Unfortunately I'm not surprised, but I've sent that to my friend at GitHub anyway

Amazing stuff

Finally found 🥹

I decided to play with error diffusion dithering for a bit. Processing pixels in the order of a low-discrepancy quasirandom sequence, distributing the error to surrounding pixels still left to process in all directions, using a 7x7 Gaussian kernel. I think it turned out nice, especially sigma 1.5 for the kernel.

Small amount of hysteresis between frames and I'm liking how it handles animation as well, despite the error diffusion

I know I was surprised that no one's done blue noise GIF dithering, at least not in a publicly usable tool yet. Seems like a simple swap from 8x8 bayer like we did in libjxl decoding

I was wondering if error diffusion might preserve detail better than just applying blue noise. Haven't compared yet, just wanted to get it looking nice first, especially without excessive flickering when flipping between slightly different images.

Yeah, it looks pretty good so far

this is fascinating

I'm amazed at the fidelity, I can make the details out a lot better than I feel like I should for a black and white dither

I'm cooking up a silly little tool to monitor and control PCs via MQTT

AY

<@&807636211489177661> reported this for spam but idk if it goes to y'all or only to discord

<:BanHammer:805396864639565834>

https://www.theverge.com/news/835453/openai-ownership-thrive-holdings https://arstechnica.com/tech-policy/2025/12/openai-desperate-to-avoid-explaining-why-it-deleted-pirated-book-datasets/ Self-annihilation continues.

https://www.theverge.com/news/836212/openai-code-red-chatgpt

code red to more distill, more censor, more overfit benchmark, more geoip ban

Financial Times: in 800 million user, only 5% user pay <:FeelsAmazingMan:808826295768449054>

Google's another possible monopoly would be an even bigger disaster

A monopoly that makes no money, sounds like a great deal for them

https://www.instagram.com/p/DRuZ-xXlkT9/

that's actually hilarious

my lord the memes

if true, this confirms that they are, presumably unintentionally, training their models on the output of themselves or other models, feeding back any discrepancies and reinforcing them

generation loss get new meaning<:kekw:808717074305122316>

you know those memes where its lewd content and it changes to something else semi-shaming you? this will be hilarious when a user generates pornography and its fucking charlie kirks face <:kekw:808717074305122316>

charlie also caused the loss of a generation in yet another sense

Seems like valve will be using a fork of waydroid's internals for proton, a shame they never reached out

will be interesting considering waydroids licencing

oh yeah this very advanced control and monitoring system is coming along nicely... ```sh $ mosquitto_sub -t 'computer/cirno/ctl/bluetooth' | while read -l new_status if test $new_status = true rfkill unblock bluetooth else rfkill block bluetooth end end $ rfkill event | while read -l null rfkill -J | jaq -r '.rfkilldevices | map_values(select(.type == "bluetooth"))[0].soft == "unblocked"' | mosquitto_pub -t 'computer/cirno/status/bluetooth' -l -r end ```

DeepSeek 3.2 supposedly beats GPT 5 in many ways and is open weights, MIT-licensed. That doesn't sound great for OpenAI either.

Everyone prefers OpenAI because it's faster and more available, because they're burning more of investors' cash per query, but the tech they have is no longer anything special apparently.

But who's gonna be responsible for popping the bubble?

At least one major investment firm has to pull out to pop the bubble.

It's such a weird bubble since LLMs are clearly useful and the open ones can't be taken away, while the datacenter spending is clearly a ridiculous waste of money 🤔

LLMs still hallucinate.

Same kind of bubble as always, overestimating short term effects causing a bubble, underestimating long term effects

And dangerous these days after insurers pulled out of AI claims so there's no cushion to fall into.

Not surprising that insurers pulled out from that mess.

They do not want to be in trouble for lawsuits related to AI at all.

OpenAI singlehandedly caused the memory shortage by the way.

We wouldn't be facing a shortage till 2028 if not for them.

I doubt there are may groups of creatives that haven't been targets of copyright infringement by AI companies hungry for training data. IIRC some book authors have sued those companies, and I would also expect that record labels and movie studios would be eager to sue if they had proof of infringement.

Plus some of the older training data is known to have included illegal content.

Was it just OpenAI? I know that they snapped up ~40% of the global RAM production capacity, but I would expect other AI companies to have made similar deals.

RAM prices started going to the Moon after the infamous October 1st deal by OpenAI.

TPU can save the bubble<:galaxybrain:821831336372338729>

Gamers Nexus did a few videos on the RAM shortage, including one on the [circular economy](<https://www.youtube.com/watch?v=h3JfOxx6Hh4>) that some of the involved tech companies have created for themselves.

Google will run into more trouble inside if they try to hoard more RAM away from their consumer division. Samsung's mobile division was not happy about their RAM division snubbing them.

FWIW lawsuits related to AI can be more than just copyright related, if there's a LLM in the process and it returns shit results then end up in bad consequences, that's a LLM caused problem, and insurers want nothing to do with that

Fair point.

https://arstechnica.com/ai/2025/12/microsoft-slashes-ai-sales-growth-targets-as-customers-resist-unproven-agents/ It's gonna come down to earth.

Yikes https://investors.micron.com/news-releases/news-release-details/micron-announces-exit-crucial-consumer-business

They've killed Crucial in favour of AI

bruhhhhhhhhhhhhhh

ouch

crucial made good memory too

As if memory prices weren't bad enough, now we've lost one of the main suppliers

because users don't need memory in the computers that view the slop from generative AI

Lots of other companies will presumably keep buying Micron DRAM chips to sell DDR to consumers.

This affects the DIY PC customers only

aka the least profitable and the most annoying group for the manufacturers

Google excuse "OpenAI are competitor" to counterpoint Department of Justice, and win<:ugly:805106754668068868>

https://www.reuters.com/sustainability/boards-policy-regulation/google-contemplated-exclusive-gemini-ai-deals-with-android-makers-2025-04-22/

<@184373105588699137> since we were talking about why Valve used FEX instead of Box64/32 Turns out Valve basically created FEX and has in fact funded full time salaries for its development since day one, in 2018: https://www.theverge.com/report/820656/valve-interview-arm-gaming-steamos-pierre-loup-griffais

Also, someone just pointed out that Alyssa Rosenzweig lists her work on the Asahi Linux drivers as being done under her contract with Valve: <https://rosenzweig.io/resume-en.pdf>

So maybe they actually do have plans to get Steam working well on macOS again???

interesting if so

https://www.nature.com/articles/d41586-025-03506-6 Even research papers will be slop at this rate.

https://www.youtube.com/watch?v=pFK35DrsFY8 video from 12 years ago, I wanted to watch that to get a general idea of how WebP compression worked it started off with "OK Glass, present slide." gosh, I completely forgot that was a thing

https://www.pcworld.com/article/2998935/ram-is-so-expensive-samsung-wont-even-sell-it-to-samsung.html

<:kekw:808717074305122316>

not sure if this is the best source but I just looked at the slides from https://www.slideshare.net/slideshow/the-vp8-video-codec/8468382 and ignored the ones that related to compression between frames

Apple looks generous nowadays

oh thanks, this is useful

I was using the spec earlier, but I much prefer when there's some diagrams to look at to back up what I'm reading

anyone knows what's the best free offline pdf to jpg conversion tool?

like any ghostscript GUI frontend

https://substack.com/inbox/post/179453867

And people still say it's not a bubble.

I use GIMP to open PDF and export as image.

Yikes, a perfect 10 https://arstechnica.com/security/2025/12/admins-and-defenders-gird-themselves-against-maximum-severity-server-vulnerability/

I use XnView MP - You can set dpi as You wish.

Is Ars Technica broken for anyone else?

I just insulted Claude Sonnet by calling the stuff it was adding to an implementation of config files "syntactic diabetes". I'm probably not the first one to coin that term but I did independently come up with it 🙂

Classy af.

https://www.tomshardware.com/tech-industry/artificial-intelligence/logitech-ceo-says-ai-gadget-makers-are-chasing-problems-that-dont-exist This IS a bubble, no one can convince me otherwise.

Microsoft can't even convince enterprise customers to buy AI stuff

I mean, I recently started using Cursor with Claude Sonnet 4.5 and it actually a powerful tool. Of course it can be dumb at times, but it can also do stuff way faster than me, and with the right nudging, it can do it quite well. So I don't think this is just a bubble, it is a major paradigm shift in how to do coding and research effectively, I think comparable with the difference between coding directly in machine code or asm and coding in some higher-level language, or the difference between doing research in pre-internet times (with physical libraries and physical books/journals being the main information/communication technology) and doing research using internet as a tool.

But of course there is also a lot of stupid use of AI, a lot of hype, and a lot of solutions in search of problems

yeah

I mean Microsoft is trying to sell Copilot in Notepad and stuff lol

AI is a tool but it's not being used as a tool, but outright replacement for white-collar workers.

Recent versions of Microsoft Word autogenerates alt text for inserted images, and IME the alt text is always incorrect.

The error diffusion dithering video I posted earlier... I coded maybe 50 lines and Gemini did the rest (of the coding and ffmpeg commands)

This is a symptom of a more general problem in capitalism, where there's a constant pressure to replace well-paid jobs with decent working conditions with some other solution that is considered "good enough" even if it means fewer jobs, worse jobs, worse working conditions. For example, call centers for tech support used to actually be staffed with people who knew the product and who could escalate issues to product managers who had actual decision power to improve things. They got replaced by cheaper people with fewer skills and with less ways to escalate issues. Then they got replaced by even cheaper people abroad who often cannot even speak the language of the customer very well, along with automated menus etc. And now they are getting replaced by AI chatbots. I don't think the problem here is AI itself. That's just the technology enabling companies to go even further in the wrong direction.

Like I don't mind the usage of AI as strictly an assistant for white-collar jobs but they're going way far beyond that, which is the reason it's fueling speculations of a AI bubble.

LLMs can't be made perfect enough to be usable without stealing the entire internet.

And everyone in this year ignores the Chinese Room problem.

They want to brute force that one so badly.

I do think that in many white-collar sectors, AI can provide a productivity boost comparable to what the industrial revolution did (machine-powered factories vs artisanal, human-powered production), and I think fundamentally there are only two real ways to handle that: 1) mass layoffs and a lot of unemployment 2) drastic reduction in working time, say from a 40h week to a 20h week, or whatever is needed to keep the same headcount

Whether or not it can provide that productivity boost is irrelevant, what's more important is if the current AI bulidup will be sustainable or not.

IBM and Google CEOs warned about irrationality in the AI market which is already bad.

And then there's rumours of 100+ % loss per 200USD/month user which is also bad.

and it's always a mystery about Copilot being which Copilot

https://www.tomshardware.com/tech-industry/dell-and-lenovo-set-to-increase-server-and-pc-costs-by-as-much-as-15-percent-as-soon-as-this-month-according-to-industry-sources-dram-and-ai-demand-create-tight-market-for-businesses-and-consumers >sell AI to businesses >but businesses can't afford PCs in the first place

> widsom AI would have caught that

(yep: )

Intentional to not trigger anti-AI people.

https://www.theverge.com/ai-artificial-intelligence/838776/maga-ai-executive-order-david-sacks MAGA is eating itself again.

https://www.tomshardware.com/tech-industry/artificial-intelligence/lisa-su-rejects-talk-of-an-ai-bubble-at-wired-event

Who the fuck will buy ROCm compute?

If writing typos and ungrammatical sentences becomes a sign/'proof' of human intelligence, then soon AI will be producing output with intentional typos and it will become a race to the bottom to produce the worst possible unreadable garbage

OpenAI bought a big amount of AMD GPUs recently

They got a deal with AMD where OpenAI would buy lots of AMD stock at an inflated price and in return got a big discount on AMD GPUs.

The bubble-bursting event 'bout to be this. 🔥

Right now it's being used for a lot of things it shouldn't. Machine learning and neural networks have a lot of amazing use cases, but we're using a rocket engine to boil a pot noodle and wondering why everything is so expensive

NNs also have a lot of amazing use cases where the NN in question is super tiny and can run in real time on a microcontroller, but there seems to be little interest in that.

There's definitely a bubble, and I hope it pops real soon, but AI certainly isn't disappearing, just restructuring to what it's actually good for

Like the Internet did.

I hear 85% of data center construction is still funded by real cash flow so it's gonna burst later.

😔

But good news! Lots of organisations are deferring AI spending till 2027.

https://www.theregister.com/2025/12/05/microsoft_365_prices_up_2026/ "Sir our AI sales quota are down. What should we do?" "Increase the prices."

https://www.theregister.com/2025/12/05/ai_is_not_a_bubble/ Investment already 2.6% of US debt. This is super super fucked.

https://techcrunch.com/2025/12/05/chatgpts-user-growth-has-slowed-report-finds/

it's insane how good phones have gotten, playing steam metro 2033 redux and getting ps3 like gaming experience xD

The latest chatbot by a Chinese https://anuneko.com

https://anuneko.com/#/share/019af217-0428-72c5-b842-2302f0f5cb2c

Quite gentle in English

wow, <#848189884614705192> slow-mode is slow

https://youtu.be/lHiMg-DLick

I ordered an Arm SBC to put inside of my old Eee PC, but now I can't find it, so when the SBC arrives, I'm gonna have to build a custom case and keyboard for it 🙂

Looking at this wondering why nothing is progressive loading, then remembering everything is WebP or AVIF https://youtu.be/T-qyNFjZaQs

I don't think progressive matters here YouTube is not designed to be viewed with 56k network and won't optimize for it It won't use progressive thumbnails even if WebP or AVIF supports it (like they don't do progressive JPEGs)

I wasn't talking about YouTube

They demo Wikipedia and their own store page

The image on the Linus Wikipedia page is also (baseline) JPEG And the images on lttstore.com have alpha, so progressive JPEG won't work there either

I really wish internet was more optimized for slow connections

a lot of my current customers are people who use cellular or sattalite internet and boy it sucks

https://www.theregister.com/2025/12/08/gartner_recommends_ai_browser_ban/

https://www.theverge.com/ai-artificial-intelligence/840034/oh-look-coreweave-is-issuing-2-billion-more-debt

https://www.tomshardware.com/tech-industry/artificial-intelligence/research-commissioned-by-openai-and-anthropic-claims-that-workers-are-more-efficient-when-using-ai-up-to-one-hour-saved-on-average-as-companies-make-bid-to-maintain-enterprise-ai-spending

They're desperate to sell snake oil.

https://www.androidcentral.com/apps-software/ai/google-seeks-to-pile-on-the-ads-in-gemini-as-reports-say-advertisers-have-been-notified

OpenAI to bring ads to ChatGPT: WE UNSUBSCRIBE!!! Google to bring ads to Gemini: OK

Because Gemini are always free for personal [so far] Never hype itself is AGI™

Google already told Gemini app users that there's no plans for ads lol.

The AI stuff is on shaky ground rn.

ChatGPT is always free too

So did OpenAI but everyone trusts Google only

People enjoy forgiving Google

Did you use GPT at 2023?<:FeelsAmazingMan:808826295768449054>

I mean Google of all things shouldn't have to tell people they don't plan ads in Gemini but they had to anyway.

for now

Suddenly my name is so bright

I mean they know how unprofitable Gemini likely is lol.

They don't need to worry about if it's unprofitable at all

Prices serve more like traffic control

Many people won't exactly pay for a chatbot either.

Almost nobody was paying for ChatGPT.

"gpt-image-1.5" is hinted

Image generation right?

Hard to beat Nano Banana Pro.

I really don't like this name

https://mistral.ai/news/devstral-2-vibe-cli

I'm using their le Chat too

https://www.wheresyoured.at/nvidia-isnt-enron-so-what-is-it/

> and the US venture capital system is set to run out of cash in the next year and a half. WHY TF THIS ISN'T BEING TALKED ABOUT?! This is extremely serious.

https://vxtwitter.com/rtwlz/status/1998187533528903777

https://vxtwitter.com/IntCyberDigest/status/1998332052907671884

<:This:805404376658739230>

https://lwn.net/Articles/1049831/

the front page of Hacker News in 2035 as imagined by Gemini: https://dosaygo-studio.github.io/hn-front-page-2035/news

someone had it imagine the articles and comments as well: https://sw.vtom.net/hn35/news.html

the comments are quite funny (for example for sudo in Zig)

https://sw.vtom.net/hn35/pages/90098999.html 😂

Apparently the click bait is unintentional lol

it's not clickbait anymore though, since he added (successful). The original title was just "The end of the kernel Rust experiment" 😄

this is great news!!

https://www.phoronix.com/news/Linux-6.19-Drops-Genocide The comment section is a trainwreck.

Happy to contribute

Which content (themes/channels) from Bilibili (chinese "YouTube") you can recommend?

I'm surprised how long it took for rust to get brought up (5 posts lol)

I remember that

He had two opened bottles of overpriced low smoke point olive oil he used to fry the food with, instead of the high smoking point olive oil from the same company (which I think he also had, but just didn't use)

Nobody care the mysticism hype

So you will see the no-evil company win again

https://x.com/xbox002000/status/1998982826281284000

https://youtu.be/nS5rj80L-pk

I wonder what the endstate will look like

oh that's a nice project

I really like Mario Galaxy on DS too

https://arstechnica.com/information-technology/2025/12/oracle-shares-slide-on-15b-increase-in-data-center-spending/

https://arstechnica.com/ai/2025/12/disney-invests-1-billion-in-openai-licenses-200-characters-for-ai-video-app-sora/

mac and cheese

lol

People in Taiwan are the BIGGEST fans of Google, buying Android phones and accepting their RIDICULOUS features, using Google Chrome whenever and wherever to see FAKE ads on the entire webpage, willing to pay for the OVERPRICED YouTube Premium, promoting FAKE ratings on Google Maps to get USELESS gifts, and boosting Google Gemini like the ultimate AGI candidate by spreading FAKE NEWS of OpenAI and ChatGPT. What people in Taiwan invented for AI is only something so small called FLAI that is crying in the corner and reaching bankruptcy!

Remember when Discord's typing indicator was eating CPU because it's animated UI elements instead of an SVG? They just did it again with voice channels, but this time tabbing out doesn't even stop it, so I've permanently lost 6% of my CPU

Incredible

ah, the shifting blue gradient?

The lines do the Spotify thing when someone is in the chanel

oof. I enabled reduced motion in accessibility settings, I hope that helps

https://arstechnica.com/gadgets/2025/12/chatbot-powered-toys-rebuked-for-discussing-sexual-dangerous-topics-with-kids/

https://ariadne.space/2025/12/12/rethinking-sudo-with-object-capabilities.html

Gemini 2.5 Pro defeats Claude Opus 4.5!

looks like the "human baseline" model has some potential

jokes aside, I'm surprised the human baseline is only 83.7%

you can try it yourself https://simple-bench.com/try-yourself

I suppose it might be less trivial if you're not aware they're all trick questions

the only one I got wrong was question 6, and I'm not convinced in wrong no matter how many times I read it

did you miss the part about the fast-approaching nuclear war?

we get news like that all the time, it'd hardly actually surprising to someone carefree. actually knowing people like them, them being without internet would be the more surprising part for them lol

fair. I assumed John assumed it was true, rather than dismissing it

Yes for question 6 the actual human answer is F, not A. Some people have been predicting a fast-approaching nuclear war for decades now and Jen could very well be one of them.

I'd answer A too, but there might be a hidden cultural variable at play. The statement says that news are relayed with certainty and seriousness, which increases the chances of the war actually having a tangible devastating effect to society. People still don't want to die or ponder about others potentially dying, as carefree as they might be. People have casual sex all the time (at least in Western society -- the cultural variable), so telling their escapades to others isn't as shocking (especially to a carefree person) in comparison as a war that could kill or maim swaths of people.

like, I've been told by my close ones about their hookups and I'm like "nice!" LOL

I'm assuming disclosing a sex escapade doesn't imply a confession of cheating in Jen's case

it depends on the person saying it. for some people, if they told me that in all seriousness I would be very concerned, for other people I'd assume they're joking, for yet other people I'd assume they heard it from a questionable source and I would not be concerned

though more than anything I suppose I understood the intent of the question, being to see if LLMs would miss the world-ending scenario because they were tripped up by all the talk about relationships

hmm, I guess it also depends on the interpretation of "certainty"

I took it as "the nuclear war isn't a bluff, it's actually coming"

Why would opinions about Google suddenly become extremely positive this year? The truth is that Google fans want to make Google become the AI monopoly so the Google's bad records can be hidden forever. They're spreading false accusations even on Wikipedia. https://w.wiki/GdHb

Textbook vandalism tbh.

The multilingual ability should be the fundament of LLM

And, GPT-5 can't correct Chinese and Japanese

Are you even a Chinese or Japanese user?

OK I don't really know much (I don't dare to say I know)

People nowadays are very unfriendly and hostile when talking about anything negative towards Google

What

People hate google more than ever

And anyway most AI companies, google and openai included, are completely evil and aligned with fascism

With Gemini 3.0 Pro people suddenly forgave Google

Sounds like some niche thing most people don't care about.

and the loudest

I've thought about it more and I agree that F is the correct answer

mainly because it says "devastated." heartbreak would likely be more devastating to me, even though I would be a lot more terrified and worried about nuclear war

lol

HP and DELL disable HEVC decoding at the laptop firmware level, even if the processor supports it, interesting <:Stonks:806137886726553651>

I love patent cartels

is it costs too much?

The fees increased

Wish Rust had a way to carry around the things the compiler can prove about a type with it. Like, if I have some newtype around integers that only lets you create even ones, the compiler knows this when the type is created, but I want that to be available to the compiler when the type is destructured

that's more or less what safety invariants do 🙂 (although of course it's not compiler-enforced)

Having it available to the compiler would be the appeal for me 😉 give me my fully safe "I already checked that this index is in bounds when I created it, please optimize accordingly" wrapper type 😛

yeah you can't really do that in Rust

I think some people are thinking about adding refinement types to the language though

I saw, Flux

and I am trying to convince some people to cook up automatic proving of the safety conditions of unsafe code

but we'll see how that goes

I don't know if that only verifies tho, or if actually also feeds the compiler the propositions. Assuming the former

That sounds cool! Could only be some of them, right? Since they can be arbitrary

there's no way it feeds the compiler the propostions, if nothing else that'd require way too much nonlocal reasoning

ah nvm, I misread

> SAFETY: the Turing machine passed to this function must terminate 😛

it might insert something like `if !(condition) { unreachable_unchecked!() };`

yeah no idea how well that'd work, but I think AI automated proving is at a level where it can provide proofs for safety conditions that are easily verifiable by humans

That's precisely what I had in mind

Hmm what sort of AI do you have in mind? There's some pretty decent proof search in Lean and probably other proof assistants too

(which I mean to contrast with a "thrown an LLM at it" sort of approach)

alphaproof or an open project I learned about a few hours ago from <@795684063032901642> but already forgot the name of

(but stuff that produces lean proofs, anyway)

https://projectnumina.ai/

Could convert Rust code into Lean with Aeneas and then use DeepSeek-Prover-V2 to hopefully automate as much of the proving as possible. Though Aeneas doesn't have much support for unsafe right yet.

I suppose proving things about unsafe code will be easier once we actually have formal semantics for it.

perhaps https://github.com/rust-lang/rust/issues/128044 will be useful eventually

Ah that's cool. Idk much about how the properties established by / needed for contracts propagate, tho

in principle this is possible, whether the compiler currently takes note of it is a different question. you could create a type that takes in an integer and divides it by 2 (and probably should panic if the number is odd) and have it multiply by 2 in its `get()` method. alternatively, just panicking on odd numbers during creation could be enough, but putting the multiply by 2 right there where it's used is an easier way for the compiler to immediately see it can never be odd (again, if it even looks at that)

I prefer `Option` to panicking, but tbh the evenness was just in order to pick a property that wouldn't be solved by ranged integers

if you're interested, looking at invariants of values function can or can't see or values types can or can't have is akin to [symbolic execution](https://en.wikipedia.org/wiki/Symbolic_execution)

one of the things I learned a lot about when I was hyperfixated on malware detection

yes, I was thinking Aeneas + numina 😛

time to rewrite jxl-rs in the Aeneas-compatible subset of Rust /j

I'd rather expand Aeneas's subset

😛

so I was looking at the official active cooler for the raspberry pi 5, right

now, there's an obvious flaw, these fins on the heatsink are blocking the airflow, they're oriented the worst possible way

no problem, I thought, I'll look for third party coolers

because surely at least one person was able to design a cooler that could be appreciated, and isn't a random heatsink with a random fan slapped on it. right?? ## [foreshadowing]

two axial fans blowing directly into the plate they're sitting on, one radial fan ignoring most of the heatsink, one axial fan blowing directly into the plate it's sitting on...

I was not able to find a *single* cooler that seemed intelligently designed

I'll give an argument that the way they're oriented is maybe fine - it's because that way creates more turbulance - you don't want laminar flow, because that leaves a layer of still air attached to the surface

idk how real that is but I doubt that doesn't affect things

I think you may be right, but in the balance between creating turbulence and making sure the air actually gets to many of the fins, I don't think enough of the latter is happening with this one (granted, I don't know how much turbulence would be the right amount)

what laptop coolers do with this little room to work with, is a duct with many thin fins. I wonder if they have like a bump in those to create turbulence or if they're mostly straight

big fan of how this one *floats above the chip it's supposed to be cooling*

okay story time over, I was slowly losing my mind as I browsed coolers and could not find a single one that didn't have some glaring oversight

I mean it's 10W max, any chonk of aluminum is likely overkill

I'm pretty sure there's laptops with 15W TDPs that don't even have fans

they do make large passive cooling cases like that and reportedly they do heat up quite a bit

you can use it without any heatsink, but it'll throttle. active cooling is recommended if you want to overclock it

I was looking into getting a cooler because I like active cooling...conceptually? (I can't articulate why, but probably autism), but I don't like any of the ones I've seen...

oh dang I found one

something that actually looks decent

<https://argon40.com/products/argon-thrml-30mm-active-cooler>