I have no idea i made that up
youre like 2 years out of date
2nd tier is Sagemaker which is also pretty much entirely up to you / minimally-managed from AWS side
Point being that in these use cases I believe there's enough use-case specific optimization that you end up getting a worse product out of a fully-managed service
The benefit of fully-managed services is ability to spin something up fast with minimal understanding of the underlying technology
For example: Deep lens, aws transcribe, aws polly. Series of services which allow you to do CV, transcription, voice synthesis/tts with minimal knowledge of the technology in the field. Quickly spin up an app that transcribes all your meetings or something.
For these you pay extra for the convenience of knowing less and have less configuration options. Basically - do I need to do exactly what they already built for me.
I think in modern LLM use cases you need enough specific customization, domain-specific knowledge (RAG), business-specific agentic workflows, blah blah blah. You have to build it yourself
Like pretty much every space they get into
Yeah this aws shit just sounds gay
I tried to use sagemaker and gave up and just rented GPUs with SSH access.
This sounds like what god intended
AWS is one of the less woke big techs - they are mostly known for slave driving their H1Bs and firing people these days
Woke in general is probably fading from the tech market. Companies that went too deep (like Google) will have to somehow claw their way back or just fade into obscurity
I dont know why jones lies about this shit
SSH no gay web terminal shit you just open up terminal and log into the fucking machine
I know you can do that with aws too but they have all those fucking options and shit on the website it’s just super gay
I currently try to get ChatGPT to validate aerospace design documents.
You guys have no fucking clue how over it is. They're going to use ChatGPT to verify planes.
Would you rather ChatGPT or black women
You can probably use Sagemaker to streamline/standardize some things but the cost is learning Sagemaker
Both options will be expensive
In your case you are learning some GPU server management whatever. Typical path from there is reinventing the wheel and realizing 2 years later Sagemaker does it for you, but by this time you are customized enough or have a specific reason you need to keep doing it yourself so you just keep doing it all yourself
Jones is 100% on medication
We need to replace all recent black hires at Boeing with ChatGPT
The black women are way better than the black men, I can tell you that much.
The convenience of sagemaker isn't cost afaik it's that they've already figured out some common mlops problems and save you from having to do all the ops stuff yourself
You're completely fine to just do it on ECS that's probably what 90% of AWS customers are doing and most likely cheaper
We need to represent marginalized groups in machine learning and train ChatGPT on black people Facebook posts
1 Like