Someone just reminded me of this lecture I gave in 2009 that described the evolution of Google Search from 1999 to 2009. People who are interested in how our search systems work might find this interesting.
It touches on disk-based serving systems, in-memory indices, compression schemes for inverted indices, latency issues due to interference from background processes, queries of death, evolution of hardware, and more..
Video: https://t.co/OnFk1azDk9
Slides: https://t.co/q4WZXRWQg6
Betty Webb MBE
It is with great sadness that today, the Bletchley Park Trust has learnt that Charlotte (Betty) Webb MBE has passed away at the age of 101. Her contributions to Bletchley Park, during World War Two and to ensuring its legacy today, will never be forgotten. https://t.co/K6EeVOZ00E
<think>The more capable these models become, the more people will want to use them for even more tasks.</think>
<answer>Inference alone will devour every GPU or accelerator it can get its hands on.</answer>
@tom_doerr A lot more tokens can go wrong. I recommend making that a 0/1 question and looking at the logprobs. That said, ain’t usually properly calibrated (probs really concentrate near 0 and 1).
@nrehiew_ Works well in some real world use-cases for sure, particularly when the original dimension is high enough. Had success compressing down from 2048 for large-scale semantic image-retrieval applications back about 5 years ago.
I did something similar and noticed the 28x28 patch size. Also noticed they have 3 extra tokens per image (likely modality and such metadata tokens). What I found curious is that it seems they resize to target a certain cap on number of tokens instead of just dimensions. Seem to be this:
max(width, image) <= 1568 and ceil(width/28)*ceil(height/28) <= 1568
First is documented but second is vaguely stated in docs. Image-only token counts seems to equal ceil(width/28)*ceil(height/28)+3.
@WebDevLex Indeed. I was bored that day and wondered what was the minimal command that would get me a fresh pair of cert/key. Had `-days` at first, but then realized I could drop it and the default was a month. 🙃
If like me you keep forgetting how to generate disposable self-signed certs and keys:
openssl req -noenc -keyout key.pem -x509 -subj '/CN=localhost' -out cert.pem
Valid for a month. Replace localhost with whatever. Can test like:
curl --cacert cert.pem https://localhost:$PORT
If like me you keep forgetting how to generate disposable self-signed certs and keys:
openssl req -noenc -keyout key.pem -x509 -subj '/CN=localhost' -out cert.pem
Valid for a month. Replace localhost with whatever. Can test like:
curl --cacert cert.pem https://localhost:$PORT