HNHacker News
TopNewBestAskShowJobs

BearOso

1,429 karma · joined August 20, 2016

submissionscomments
BearOso··on Memory has grown to nearly two-thirds of AI chip component costs
What novel data hasn't already been used in training? What new algorithms are there? Can you post some links so we can read about them?
BearOso··on Memory has grown to nearly two-thirds of AI chip component costs
Unless there's a new paradigm, scaling up is all they can do to improve performance. They've shrunk down all the way to 1-bit models and all the low-hanging fruit is gone. There's no way for them to get much smaller, so they have to get bigger and faster to meet expectations.
BearOso··on Anthropic's "Profitability" Swindle
Meta also took on an additional $100B in debt recently to push AI and fired 4000 employees. The economics just don't add up at Meta, so they're a bad example.
BearOso··on Steve Wozniak cheered after telling students they have AI – actual intelligence
> Weren’t Schmidt’s comments on AI the harsh “truth” from the perspective of someone who directly benefits from the wealth extraction capabilities of AI?

There are no wealth extraction capabilities yet. It's a money pit. They're certainly hoping it'll surpass some breakpoint and become profitable by brute-forcing compute power, but that's very optimistic. The propaganda Schmidt is pushing envisions that future in hopes of raising current stock prices so they can afford the brute-forcing that's very unlikely to succeed.

My prediction is that we'll keep the tools we've acquired, probably refined a bit, but the LLM path is eventually a dead-end. After this, if they still try to monetize, remote models will be extremely expensive.

BearOso··on Everything in C is undefined behavior
We also have a very limited number of compilers and a small number of prevalent architectures today. As long as you know the behavior of the target compiler and architecture, the behavior is defined, it's just not specified.
BearOso··on Bun's experimental Rust rewrite hits 99.8% test compatibility on Linux x64 glibc
From the leaked internal prompts, Opus 4.7 vs 4.6 recomputes several times over before returning the result. For heavy use like this, it costs Anthropic far more than you're paying as a consumer. They rely on the light users to offset the whales, and they're still at a significant net loss. If you tried this as an end-user, they might cut you off (though I understand their data centers are underutilized, so that wouldn't be for logistic reasons). Being part of the company and directly sanctioned, the author has unlimited access.

~7x overcompute * ~7x real cost to Anthropic * your 10-20k estimate for consumer use is my thought for actual total cost. If the honeymoon period runs out and they're still in business, this is what everyone will pay.

BearOso··on Bun's experimental Rust rewrite hits 99.8% test compatibility on Linux x64 glibc
It costs several times what it would cost a small team of engineers, even assuming you gave the engineers more time to do it. I'm guessing (wildly) this was around 0.5M USD in compute time. You do get the result quicker, though.
BearOso··on Super ZSNES – GPU Powered SNES Emulator
Not competing, I just develop it occasionally for fun. I'm the first to suggest bsnes or another of Near's projects as a first choice. Mesen2 is pretty good, too, but the Avalonia UI kind of irks me. Definitely run that through Retroarch.

I don't see anything worthwhile in this Super ZSNES project yet. We were all ZSNES users back in the days.

BearOso··on Super ZSNES – GPU Powered SNES Emulator
> This is from the original authors of ZSNES. I think they know what they're doing.

ZSNES is popular because of legacy and nostalgia. It was very fast and came at just the right time, but the developers aren't quite coding gods. Their expertise on the SNES is no match for late/later developers like Near and Sour.

With Super ZSNES, being a Unity project, you can get a fairly clear decompile of the IR, and the code doesn't seem all that impressive. It's alpha-quality, but generally coded like the original ZSNES was. Optimization is completely missing and accuracy is still out of whack, but synchronization is improved and it's doing a little better job of counting cycles. "GPU-powered" is a big stretch. ~~They're only taking advantage of it for fixed-function perspective transform on mode 7~~ Scratch that. It borrows the line-based algorithm from bsnes-hd, including the trick to interpolate the transform variables between mins and maxes. So the only GPU feature it uses is blending for bump-mapping.

BearOso··on Super ZSNES – GPU Powered SNES Emulator
I came up with something similar to your idea, a GPU compute PPU for future Snes9x. What they're doing is using legacy fixed-function API to draw quads, then blend a bump map on with the final image. It's weird. We have the tools to do some really cool things with GPUs, but they chose this. I'm more impressed by all the post-processing shaders people have come up with for all the other emulators.
BearOso··on GitHub Copilot is moving to usage-based billing
If you paid attention to the power requirements and amount of hardware being put into data centers, you should have realized that it cost them an order of magnitude more than you were being charged. To rework your analogy: they hooked you, now they're gonna see if they can reel you in.
BearOso··on New 10 GbE USB adapters are cooler, smaller, cheaper
Probably because with USB 3.2 2x2 they were reviewing too many longer cables that didn't meet the requirements, so they lowered the length so companies didn't submit them only to fail to get certified. It's worth noting that 1.2m is now in the USB4 spec.
BearOso··on New 10 GbE USB adapters are cooler, smaller, cheaper
The cable length is only for the spec. You can get longer cables that achieve the higher bandwidth, they're just not certified for that.
BearOso··on Windows 9x Subsystem for Linux
It's mostly explained if you go to the project page. For me, the I would say the hardest thing about something like this is gleaning the Microsoft driver APIs. In the 9x days, Microsoft documentation was not quite thorough and difficult to access. It's still not pleasant.
BearOso··on Claude Opus wrote a Chrome exploit for $2,283
This was using an exploit already fixed in a recent version and publicly known. It's worthless on the black market or as a bug bounty.
BearOso··on They're Made Out of Meat (1991)
> To reduce us to anything less is to deny the awesomeness of the cosmos itself.

Teacher: "Photosynthesis makes energy from water, CO2 and light. The mitochondria are the power centers of the cell."

Grade-schooler: "How do they work?"

Teacher: "Um. Um..."

Modern scientist: "Quantum entanglement and tunneling. We don't really understand any of it."

BearOso··on The cult of vibe coding is insane
It's also crazy more expensive to run than we thought. That doesn't bode well when their loss-leader period is over and they need to start making money.
BearOso··on We sped up bun by 100x
The commit I linked shows that it didn't even read the user name and email from git's config file, but used a test name, which means it's woefully incomplete.

Then there's stuff like this: https://github.com/hdresearch/ziggit/blob/master/src/cmd_bra...

It's just one giant function. Sometimes big functions are necessary. This one is clearly AI generated and not very readable for a human. This is just from a quick glance.

BearOso··on We sped up bun by 100x
You have to maintain a completely separate implementation of AI generated code that's translated from C, so not even idiomatic zig.

Edit And then I go their repository and read commits like this https://github.com/hdresearch/ziggit/commit/31adc1da1693e402... which confirms it wasn't even looked over by a human.

BearOso··on We sped up bun by 100x
They mention Anthropic, so I assumed something similar. At $5 per 5 million tokens, 13 billion would cost $65,000. However, the image in the article shows over 17 billion used, which is $85,000. That's an entry-level programmer's yearly salary. It doesn't quite pass their code tests, and it's automatic code translation, so it's going to be a pretty direct transcription. There's still probably a lot of messy code to clean up. I'm not sure it's worth it.
BearOso··on Nvidia rolls out its fix for PC gaming's "compiling shaders" wait times
This is effectively what fossilize for steam can do. It records the shaders and structures at an intermediate level, then rebuilds them for the specific hardware. It also does distribution, so you always have ridiculously low shader compile times. I like it because it makes proton better than running under Windows because it eliminates shader stutter.
BearOso··on Windows 95 defenses against installers that overwrite a file with an older one
It was SimCity Classic.
BearOso··on MacBook M5 Pro and Qwen3.5 = Local AI Security System
That first bullet is a bit sketchy. Benchmarks, particularly geekbench, may have increased 6x, but that's being manipulated.

The GPUs have become much larger, so 6.8x is believable there, as is the inclusion of a matmul unit boosting AI.

The 2.x numbers are the most realistic, especially because they represent actual workloads.

BearOso··on Review of Microsoft's ClearType Font Collection (2005)
At the traditional 96 dpi, you have to be 3 ft away to exceed the retinal density. Personally, I sit at half that distance. Something around 200 would be more ideal. Laptops you might sit even closer.

Mobile devices, unless you get really close to the screen, have matched the retinal density for a while. Most people hold the device at about 8 inches, so 450 dpi is the value to hit.

Edit These measurements assume 20:20 vision, which is the average. Many people exceed that. So you'd need slightly higher values if you're feeling pedantic.

BearOso··on Faster C software with Dynamic Feature Detection
For function-multiversioning, the intrinsic headers in both gcc and clang have logic to take care of selecting targets. You also don't need to do dispatch manually when writing manual optimizations--the same function name with different targets is supported and dispatches automatically.
BearOso··on Making Video Games in 2025 (without an engine)
> Things like the famous fast inverse square root are short, but I would hesitate to describe it as simple.

Not the best example. That snippet was in use at SGI for years and actually written by Gary Tarolli. Quake's optimization was mostly done by Michael Abrash.

The original id engines were also famously inflexible. They fit the mold of "developing an engine, not a game" to a T. What you saw them do was all they could do. Look at how much Half-Life needed to add to be viable. idtech3 also only broke out of its niche because Ritual and Infinity Ward heavily modified it and passed it around. There's a good reason the engine-based ecosystem is so prominent now.

BearOso··on How far back in time can you understand English?
It's fairly easy when it's in print like this. When it's handwritten I have trouble just going back to 1900.
BearOso··on 15 years later, Microsoft morged my diagram
From the beginning, one of the advertising tricks they have used for AI is FOMO. I presume that is so they can sell you as much of it as they can before you realize its flaws.

Everybody's so worried about getting in on the ground floor of something that they don't even imagine it could be a massive flop.

BearOso··on Thanks a lot, AI: Hard drives are sold out for the year, says WD
Only if you instantiate it once.

If you use it like an agent and stick it in a loop and run it until it achieves a specific outcome it's not.

BearOso··on Thanks a lot, AI: Hard drives are sold out for the year, says WD
Did you pay attention in computer science classes? There are problems you can't simply brute-force. You can throw all the computing power you want at them, but they won't terminate before the heat-death of the universe. An LLM can only output a convolution of its data set. That's its plateau. It can't solve problems, it can only output an existing solution. Compute power can make it faster to narrow down to that existing solution, but it can't make the LLM smarter.
← PreviousPage 3 of 20Next →