Avatar
John
bcbf9644d3f475d00eb9c6e467385ce16d4546c1a24222ccfa542bf776eaba95
ML Engineer (Former) attorney

"There can be no such thing as conditional property"

- Ragnar Danneskjöld

Chatbots gonna accelerate language learning like crazy

Just changed the system prompt to solo hablas en espanol and its working great

Speech to speech would be crazy but readings good too

Transformers one has some charm

But its like the origin of optimus vs Megatron

Megatron starts out good and they make a big deal out of him changing with his willingness to fight / kill transformers

Meanwhile the "good" guys are gleefully killing transformers too.

But its only bad when Megatron does it

Nostr has 1,461,501,637,330,902,918,203,684,832,716,283,019,655,932,542,976 users

Idk why I enjoy loading up my 3090 with vLLM requests so much

That's hilarious

"We're afriad that results showing that [thing] is bad may be used by people who don't support [thing]"

Telling on yourself that the main goal isn't truth/helpfulness

Or their in so deep their worldview forces them to ignore contrary evidence

Dang I just lost like all enthusiasm for this project

8b on akimbo 3090s can "read" ~3.8 pages/sec

OK so srsly limiting the response output gave giant speed boost. Token generation is like 10x slower than prompt reading.

Outlines let's you constrain model output but I don't see how to have a conditional requirement. Like result:boolean quote:required if true

So gonna break it into 2 calls, one for the result and one for the quote

Lmao so the speedup was BC half the requests bounced BC the prompt was over the context limit

Cutting the output length by 60% gave me a 2x speedup.

A huge LLM application is turning unstructured data into structured data

Fading the nostr consensus that trump would be better. I think Harris is long term better by being worse

Adversity creates strength or something.

extends the cheap SATs period and/or hastens the collapse

"O(n^3) won't be that bad, my input is small"

420 LLM calls and 7 minutes later for a 1/7th scale test

Man who's business is effectively being a bitcoin custodian says using a custodian is good

I think folks are overreacting

The audio quality is better and they tend to speak slower/clearer