DEV Community

Models — smaller, faster, sharper Series' Articles

Back to LYR's Series
Compress a model to a quarter of its size and its smarts barely drop
Cover image for Compress a model to a quarter of its size and its smarts barely drop
lyr profile
LYR

Compress a model to a quarter of its size and its smarts barely drop

#machinelearning #performance #ai
Comments
6 min read
Taking a small AI's accuracy from 42% to 84% — and on some tasks, level with the giants
Cover image for Taking a small AI's accuracy from 42% to 84% — and on some tasks, level with the giants
lyr profile
LYR

Taking a small AI's accuracy from 42% to 84% — and on some tasks, level with the giants

#machinelearning #ai
Comments
7 min read
The AI that was smart on the bench suddenly started making mistakes in production
Cover image for The AI that was smart on the bench suddenly started making mistakes in production
lyr profile
LYR

The AI that was smart on the bench suddenly started making mistakes in production

#machinelearning #testing #devops #ai
Comments
7 min read
The "97% improvement" I threw away
Cover image for The "97% improvement" I threw away
lyr profile
LYR

The "97% improvement" I threw away

#machinelearning #testing #ai
Comments
7 min read
"Add context and the model gets smarter" was half a lie
Cover image for "Add context and the model gets smarter" was half a lie
lyr profile
LYR

"Add context and the model gets smarter" was half a lie

#llm #testing #ai
Comments
9 min read
I cut the bottom-rank rate from 64% to 24%, and held the top-rank loss to 12pt
Cover image for I cut the bottom-rank rate from 64% to 24%, and held the top-rank loss to 12pt
lyr profile
LYR

I cut the bottom-rank rate from 64% to 24%, and held the top-rank loss to 12pt

#machinelearning #testing #performance #ai
Comments
9 min read
Swahili from 0% to 30%. But the value wasn't the score — it was that the failures changed
Cover image for Swahili from 0% to 30%. But the value wasn't the score — it was that the failures changed
lyr profile
LYR

Swahili from 0% to 30%. But the value wasn't the score — it was that the failures changed

#machinelearning #testing #ai
Comments
11 min read
Let the AI pick a translation better than the "right answer"
Cover image for Let the AI pick a translation better than the "right answer"
lyr profile
LYR

Let the AI pick a translation better than the "right answer"

#machinelearning #testing #ai
Comments
10 min read
A small specialist drew level with the strongest model in service
Cover image for A small specialist drew level with the strongest model in service
lyr profile
LYR

A small specialist drew level with the strongest model in service

#machinelearning #ai
Comments
8 min read
Let another AI write your teacher data
Cover image for Let another AI write your teacher data
lyr profile
LYR

Let another AI write your teacher data

#machinelearning #ai
Comments
12 min read
Generation speed is decided by the byte count of the weights, not the parameter count
Cover image for Generation speed is decided by the byte count of the weights, not the parameter count
lyr profile
LYR

Generation speed is decided by the byte count of the weights, not the parameter count

#machinelearning #performance #testing #ai
Comments
7 min read
A general-purpose model is too wasteful a vessel for subtitles
Cover image for A general-purpose model is too wasteful a vessel for subtitles
lyr profile
LYR

A general-purpose model is too wasteful a vessel for subtitles

#machinelearning #devops #llm #ai
Comments
18 min read