5 ms·

Amazon Bedrock Is Now Generally Available

33 points by epberry 3y ago

paulddraper 3y ago
Tangential, but I kinda miss the days when this would be SFM, Simple Foundation Models.
bguberfain 3y ago
It says it has support for Llama 2, but in a deep page you can read that it is "coming soon". Anyway, good to see support for serverless inference and painless train of Llama 2 models!
- willtemperley 3y ago
  Interesting timing given CloudFlare just released Workers AI, which actually has Llama 2 [1], not just coming soon. Great to see some competition.
  - willtemperley 3y ago
    [1] https://blog.cloudflare.com/writing-poems-using-llama-2-on-workers-ai/ https://blog.cloudflare.com/writing-poems-using-llama-2-on-w...
leoqa 3y ago
Asking because I’m lazy: if I need to transcribe audio in real-time, is there a state of the art model I can plug into?
- CrimsonCape 3y ago
  https://github.com/ggerganov/whisper.cpp https://github.com/ggerganov/whisper.cpp https://github.com/Const-me/Whisper https://github.com/Const-me/Whisper I had fun with both of these. They will both do realtime transcription. Bit you will have to download the training data sets…
- theossuary 3y ago
  I saw whisper recommended, but I was curious how it compares to the other robust ASR systems (like nvidia's Nemo + Riva). I found this Twitter thread that seemed relevant: https://nitter.net/lunixbochs/status/1574848899897884672 https://nitter.net/lunixbochs/status/1574848899897884672 Long story short, it depends on what you want to use it for. Different models and different training sets can help optimize for different things. Also, if you're in a domain with very uncommon speech patterns (think doctor shorthand or radio lingo), you'll need to understand how difficult it will be to customize generated models to do better in your space. I think Nemo + Riva does well at this; but I'm not as familiar with other options.
  - leoqa 3y ago
    Yeah I fine tuned a model 2 years ago but it was a big pain and performance didn’t get better than 85%
- crakenzak 3y ago
  whisper from OpenAI works great for me.
ranman 3y ago
If you want to use this but are tired of waiting for the SDKs to get updated you can do this: ``` curl -sS https://cdn.caylent.com/bedrock_new.zip https://cdn.caylent.com/bedrock_new.zip > bedrock.zip mkdir -p ~/.aws/models/ unzip bedrock.zip -d ~/.aws/models ```

33 points by epberry 3y ago

Tangential, but I kinda miss the days when this would be SFM, Simple Foundation Models.

It says it has support for Llama 2, but in a deep page you can read that it is "coming soon". Anyway, good to see support for serverless inference and painless train of Llama 2 models!

willtemperley 3y ago
Interesting timing given CloudFlare just released Workers AI, which actually has Llama 2 [1], not just coming soon. Great to see some competition.
- willtemperley 3y ago
  [1] https://blog.cloudflare.com/writing-poems-using-llama-2-on-workers-ai/ https://blog.cloudflare.com/writing-poems-using-llama-2-on-w...

leoqa 3y ago

Asking because I’m lazy: if I need to transcribe audio in real-time, is there a state of the art model I can plug into?

CrimsonCape 3y ago
https://github.com/ggerganov/whisper.cpp https://github.com/ggerganov/whisper.cpp https://github.com/Const-me/Whisper https://github.com/Const-me/Whisper I had fun with both of these. They will both do realtime transcription. Bit you will have to download the training data sets…
theossuary 3y ago
I saw whisper recommended, but I was curious how it compares to the other robust ASR systems (like nvidia's Nemo + Riva). I found this Twitter thread that seemed relevant: https://nitter.net/lunixbochs/status/1574848899897884672 https://nitter.net/lunixbochs/status/1574848899897884672 Long story short, it depends on what you want to use it for. Different models and different training sets can help optimize for different things. Also, if you're in a domain with very uncommon speech patterns (think doctor shorthand or radio lingo), you'll need to understand how difficult it will be to customize generated models to do better in your space. I think Nemo + Riva does well at this; but I'm not as familiar with other options.
- leoqa 3y ago
  Yeah I fine tuned a model 2 years ago but it was a big pain and performance didn’t get better than 85%
crakenzak 3y ago
whisper from OpenAI works great for me.

ranman 3y ago

If you want to use this but are tired of waiting for the SDKs to get updated you can do this: ``` curl -sS https://cdn.caylent.com/bedrock_new.zip https://cdn.caylent.com/bedrock_new.zip > bedrock.zip mkdir -p ~/.aws/models/ unzip bedrock.zip -d ~/.aws/models ```