Hacker News: tarpitt

New comment by tarpitt in "If Claude Fable stops helping you, you'll never know"

tarpitt — Wed, 10 Jun 2026 01:34:30 +0000

idk I just bought a 7900 XTX for $750 on ebay and it runs gemma and qwen pretty well

New comment by tarpitt in "Cloudflare Turnstile requiring fingerprintable WebGL"

tarpitt — Wed, 03 Jun 2026 02:01:07 +0000

Necroing this, but perhaps you might be interested in some sort of BOINC-like PoW scheme for websites. This was a distributed computing project originially known as Seti@Home. It's not really practical for cryptocurrency PoW applications (despite its use in Gridcoin) due to the centralized nature of the challenge-response, but certainly more useful than captchas or hashes!

New comment by tarpitt in "Cloudflare Turnstile requiring fingerprintable WebGL"

tarpitt — Mon, 01 Jun 2026 02:29:42 +0000

There's a saying that if an idea is stupid, but it works, it's not stupid.

If some computation is "useless" but it serves it's purpose, it's not useless.

The reason why bitcoin network expends so much energy is down to tokenomics, not the system of PoW itself. At equilibrium we expect the power usage to be (blocks/hr) x (BTC/block) x ($/BTC) x (kWh/$), so it's a function of the BTC price and emission rate.

PoW in other context has way different driving factors. In this case, the marginal improvement of fetching the site again for AI bots isn't enough to cover the PoW cost. The PoW cost is outweighed by the net bandwidth cost of all the parties.

New comment by tarpitt in "Atherton spent $145K to delay train electrification. The rest of us paid $400M"

tarpitt — Mon, 01 Jun 2026 00:07:02 +0000

I wouldn't find it so grating if I wasn't constantly reading it. Maybe 5 years ago I would have liked this writing pattern. Same for those diagrams they include.

What it lacks is the effort it takes for a couple minutes to whittle down the writing to the essential parts and bring out the particular style of the author. Even before this wave of LLM stuff started, I knew that it's editing that makes good writing. LLMs are mostly a single-pass process. Maybe it would be improved by having a couple intermediary high-temp thinking stages. the time tradeoff is worth it if it's producing some blog or speech to be consumed by 100s or 1000s of people even if you are only improving the writing by, say, 5%.

To be honest, I don't mind the diagrams when I'm the one prompting it. Other than the fact that it is a waste of tokens, and full of clickable dialogs that waste more of my tokens.