Hacker News: WithinReason

New comment by WithinReason in "Haunting Photos Show the Aftermath of the Kursk Submarine Disaster in 2000"

WithinReason — Tue, 07 Apr 2026 16:23:34 +0000

Also:

"No search was launched for more than six hours."

"It ultimately took over 16 hours to locate the stricken vessel, which lay on the seabed at a depth of 108 meters (354 feet)."

New comment by WithinReason in "AGI Is Here?"

WithinReason — Sun, 05 Apr 2026 13:36:36 +0000

because the I in AGI stands for intelligence, not sentience.

New comment by WithinReason in "AGI Is Here?"

WithinReason — Sun, 05 Apr 2026 13:35:43 +0000

besides, even humans don't have general intelligence

New comment by WithinReason in "H.264 Streaming Fees: What Changed, Who's Affected, and What It Means"

WithinReason — Fri, 03 Apr 2026 12:19:31 +0000

Dolby just sued Snapchat over patents for using AV1:

https://www.techspot.com/news/111865-dolby-sues-snap-over-vi...

New comment by WithinReason in "Solar Balconies Take Europe by Storm"

WithinReason — Thu, 02 Apr 2026 06:21:49 +0000

"In 1839, the ability of some materials to create an electrical charge from light exposure was first observed by the French physicist Edmond Becquerel"

So yes

New comment by WithinReason in "Ask HN: Distributed data centers in our basements"

WithinReason — Tue, 31 Mar 2026 15:51:05 +0000

> 1. And data centers also exist in cold places. But if you put 8kw of extra heat in someone's home that previously didn't need cooling, it might need it now.

That's the entire point of being in a cold place that you don't need active cooling. Just open the window.

> 2. But running more servers with worse uptime is less efficient and requires more capital expense than running fewer servers with better uptime.

Even if the cooling is free? Not even free, the cost is negative since it saves heating cost.

> 3. and LLMs are a bad choice to run residentially for other reasons, particularly power density

Can you explain the connection of LLMs to power density? This point makes no sense.

> 4. Absolutely not -- basically all industry data protection standards have physical security standards. At least, any of the ones that matter.

You can lock a box physically

> 5. That is true, there are data centers without 24/7 access. They tend to struggle to compete, though.

Why though if redundancy exists, like you said? Would they still struggle to compete if the cooling cost was effectively negative?

> 6. Is it? Residential power and cooling costs more -- and that's the majority of the cost to colocate servers

You can make cooling cost negative, if that's the majority of the cost, then that's great! And you can also place your servers in residential areas with the cheapest power.

New comment by WithinReason in "Ask HN: Distributed data centers in our basements"

WithinReason — Tue, 31 Mar 2026 15:19:18 +0000

1. Many places in the world don't ever need cooling

2. If servers are distributed then downtime is distributed, you can virtually guarantee that some of the servers over the world will be online so you can get effectively 100% uptime, something that is not possible in a data center

3. To serve tokens you need very little bandwidth, it's just text in and out

4. All of this is down to the HW and the SW itself, not the building. That is, the box that's being deployed.

5. Just switch to a different server until the problem is resolved, in this model there is no urgency. You just need redundancy which you can afford with how much cheaper this would be.

New comment by WithinReason in "Vulnerability research is cooked"

WithinReason — Tue, 31 Mar 2026 07:34:05 +0000

If it can be trained with reinforcement learning then it will happen

New comment by WithinReason in "Run a 1T parameter model on a 32gb Mac by streaming tensors from NVMe"

WithinReason — Tue, 24 Mar 2026 16:55:31 +0000

Have you ever generated access frequency statistics for the experts in these models, something like a histogram?

New comment by WithinReason in "Microsoft's "fix" for Windows 11"

WithinReason — Tue, 24 Mar 2026 10:33:21 +0000

Heartbreaking: The Worst Company You Know Just Made A Great Point

LLM written FlashAttention: 1.7x PyTorch

WithinReason — Mon, 23 Mar 2026 14:28:13 +0000

Article URL: https://xcancel.com/__tinygrad__/status/2036003426933551586

Comments URL: https://news.ycombinator.com/item?id=47490045

Points: 1

# Comments: 0

New comment by WithinReason in "Tinybox – A powerful computer for deep learning"

WithinReason — Sat, 21 Mar 2026 21:34:19 +0000

Only gamers understand that reference

-- Jensen Huang

New comment by WithinReason in "Oregon school cell phone ban: 'Engaged students, joyful teachers'"

WithinReason — Fri, 20 Mar 2026 16:02:45 +0000

For the walk home

New comment by WithinReason in "Chuck Norris has died"

WithinReason — Fri, 20 Mar 2026 14:20:34 +0000

I'm sure he'll get better soon

New comment by WithinReason in "Show HN: Duplicate 3 layers in a 24B LLM, logical deduction .22→.76. No training"

WithinReason — Thu, 19 Mar 2026 14:52:51 +0000

Are you Turing testing?

New comment by WithinReason in "Show HN: Duplicate 3 layers in a 24B LLM, logical deduction .22→.76. No training"

WithinReason — Thu, 19 Mar 2026 14:47:57 +0000

These layers are residual layers, so what a layer does is:

x = x + layer(x)

so it's not too surprising that they can be used recurrently

New comment by WithinReason in "Show HN: Duplicate 3 layers in a 24B LLM, logical deduction .22→.76. No training"

WithinReason — Thu, 19 Mar 2026 14:27:01 +0000

Some transformers have a block recurrent structure, here is a paper that made a similar observation recently:

https://www.alphaxiv.org/abs/2512.19941

Invisalign Became the Biggest User of 3D Printers

WithinReason — Tue, 17 Mar 2026 12:19:38 +0000

Article URL: https://www.wired.com/story/how-invisalign-became-the-worlds-biggest-3d-printing-company/

Comments URL: https://news.ycombinator.com/item?id=47411658

Points: 2

# Comments: 0

New comment by WithinReason in "Executing programs inside transformers with exponentially faster inference"

WithinReason — Fri, 13 Mar 2026 14:55:57 +0000

My guess would be in the ballpark of about 10000 times less efficient

New comment by WithinReason in "Are LLMs not getting better?"

WithinReason — Thu, 12 Mar 2026 12:48:26 +0000

If you look at a separate trend for the smaller Sonnet models, you can see a rapid trend