calmdownkarm retweetou
calmdownkarm
3.9K posts

calmdownkarm
@karmanya
I ship probabilistic things. ML Engineer with 3 years of industry experience.
Philadelphia, PA Entrou em Ağustos 2008
2.8K Seguindo493 Seguidores

I have been curious about this as well.
Would be a real shame if it is just some PEFT adapters being tuned.
Abhishek Maiti@o_v_shake
how does openai host so many finetuned models, that too serverless on-demand?
English

@ankurb @tejesh_1 is building readpods.com - not sure if it lets you chat yet but might be in the future
English

@ramachandranesk @FiftyTwoDotIn Don't think I understand the article - the implication seems to be that a lot of researchers are spending time looking at sanskrit/rule based systems instead of following the deep learning approach, but doesn't concretely make that claim or substantiate it?
English

if you want to know why it's "hopeless" for India to build foundational AI like ChatGPT, it's because for decades, India has tried to use Sanskrit in its approach to language processing. 7mos of reporting & research explain, in my story for @FiftyTwoDotIn
fiftytwo.in/story/restrict…
CP Gurnani@C_P_Gurnani
OpenAI founder Sam Altman said it’s pretty hopeless for Indian companies to try and compete with them. Dear @sama, From one CEO to another.. CHALLENGE ACCEPTED.
English
calmdownkarm retweetou

Starting a series on what I've been doing for past 4 years. Feedback is always appreciated.
link.medium.com/naFnuCeIaAb
English
calmdownkarm retweetou

Live footage of PIs checking their cloud computing and API bills after #NeurIPS2023 deadline.
GIF
English

@Saurav_Varma @karmanya @angrykabootar This table is a comparison for serverless gpu options for inference though!
English

@Saurav_Varma @karmanya @angrykabootar @karmanya use the oldest billing account(they generally track your account health based on billing history) that you can use. That worked for us when working with GCP(I assume aws has a similar concept)
English
calmdownkarm retweetou

Transformer Puzzles
- (github.com/srush/Transfor…)
7 short visual puzzles for learning how Transformers can compute basic algorithms. Based on Thinking Like Transformers.
(Presented at ICLR Blog today. Doing these has been one of the most useful ways to grok how LLMs compute)


English

Really fun work I did last semester got accepted to ACL. arxiv.org/pdf/2305.01528… in collaboration with zhu.codes /Alex Feng/@LangTechLara and Chris Callison Burch.
English

I recreated OpenAI's Todo Chatgpt plugin tutorial using @FastAPI and wrote some documentation for fellow ML engineers that haven't touched web stuff in a while. github.com/CalmDownKarm/a…
English

@RajaswaPatil @guardrails_ai @inspiredco_ai Look at openai evals prs probably a good place to start looking
English

What are some good works for LLM output testing?
I know some deterministic validators being implemented by @guardrails_ai and a few text quality checks implemented in Critique by @inspiredco_ai.
Any other semantically open-ended checks for something like Abstractive/Open QA?
English

something I found kinda cool from the GPT4 whitepaper is that RLHF doesn't seem to be as powerful as I thought it was, it definitely makes a difference, but I was overestimating its utility. Puts LLama in a much better light.
cdn.openai.com/papers/gpt-4.p… (page 27)
English
calmdownkarm retweetou





