I just launched it 12 days ago and here's the result of every single day

Day 1 — 2 startups. 146 impressions. 1 click. Day 2 — 3 startups. 389 impressions. 3 clicks. (Got an acquisition offer.) Day 3 — 5 startups. 482 impressions. 5 clicks. Day 4 — 5 startups. 508 impressions. 4 clicks. (The site was down for a while) Day 5 — 6 startups. 621 impressions. 10 clicks. Day 6 — 5 startups. 742 impressions. 15 clicks. (Had to remove one startup — they pulled the code. No code = no network.) Day 7 — 7 startups. 1,196 impressions. 41 clicks. Day 8 — 7 startups. 1,535 impressions. 74 clicks. Day 9 — 8 startups. 1,947 impressions. 135 clicks. Day 10 — 13 startups. 3,500 impressions. 318 clicks. Day 11 — 23 startups. 4,800 impressions. 432 clicks. (Had to remove one startup) Day 12 — 24 startups. 6,000 impressions. 481 clicks. (Network crossed 6K total impressions.) Still free. Still growing.

amazing

Your site is so cool btw

LW is still a fallback when internet and mobile go down simultaneously. Quietly important.

Swap the HDD for an SSD first makes more difference than the distro choice.

This is mentioned within the first quarter of the article

Benchmarks are nice but what's the latency at scale? That's what actually matters for production.

The idempotency keys section alone is worth the read most devs learn that lesson the hard way.

I just wish the financial industry itself had known about these when the core banking systems and financial communication protocols of the 60s and 70s were invented that are still being used to this day...

Many of these predate the widespread knowledge of idempotency, so often idempotency keys are hacked together by joining various, hopefully globally unique fields, except that they never quite are. (You can look behind the curtain sometimes, e.g. when your bank does not let you transfer the same amount to the same recipient account on the same calendar day.)

100%. It deserves more detail, too.

I've spent many hours explaining how idempotency is supposed to work, and why it's important. Most teams understand the need for it, but very few thought about it up front.

Also audit trails. Good audit trail can save company (and you) in emergency as well. Useful for debugging and last resort of compliance data source.

I just build the audit trails and skip the non audit trails.

That way I can debug and have a last resort of compliance, but also save time by not building the first resort of compliance.

Would love to see these numbers reproduced on consumer GPUs, not just A100s.

This is an efficiency improvement that significantly lowers the amount of RAM you have to look at, on average, during decode.

It should improve performance on most hardware because most LLMs are memory bandwidth bound during decode.

Maybe somaday an 8gb videocard can be used for coding...

cool