systems over syntax | building https://xitter.cf/t.co/9Gf40mLqq6
Ahmedabad
Joined March 2024
- Tweets8.9K
- Following737
- Followers2.1K
- Likes50.3K
deep retweeted
Introducing Orcus.
No LLM API key required to make AI model requests.
Orcus uses your web credentials to access AI models available through supported provider websites and deliver responses to your applications through a familiar, API-style interface.
From model invocation and streaming to structured outputs, tool calling, and batch processing, Orcus brings essential AI integration capabilities into one developer-friendly SDK.
Explore Orcus: orcus.site/docs/gemini
Have you ever paid for LinkedIn Premium?
Was it actually worth it, or did you cancel it after the free trial?
馃檹馃槶 The state of media in this country is genuinely embarrassing. Giving clowns all the attention while real issues get swept under the rug
after watching a movie there are only two options:
>forget everything 3 seconds later
>make it your whole personality trait for months until another movie takes over
First in my bloodline to see a billionaire getting trolled on his own platform.
(Maza aa raha hai, juice kahan hai mera)
Today I went deeper into LLM inference and the systems behind it
Focused on attention and KV-cache optimization, compute vs. memory bottlenecks, and the trade-offs that shape inference performance, throughput, and batch size.
Getting a clearer picture of what it takes to make model serving efficient at scale.
Spent today going deeper into the data and inference side of building AI systems
Focused on how data quality and curation affect model behavior, how training data can be improved or synthesized, and how inference can be optimized for better latency, throughput, and resource efficiency
Wrapped up a deep dive into fine-tuning
>memory bottlenecks in training
>full fine-tuning vs PEFT
>quantization
>adapters & soft prompts
>LoRA / QLoRA
>backpropagation & trainable parameters
>model merging & multi-task fine-tuning
getting a much clearer picture of how foundation models can be adapted efficiently for different tasks without retraining everything from scratch
deep retweeted
As most you ask me so... wrote down what I learned in a simple, beginner friendly way.
If you're also learning monitoring/observability, this might help.
Wrapped up a deep dive into fine-tuning
>memory bottlenecks in training
>full fine-tuning vs PEFT
>quantization
>adapters & soft prompts
>LoRA / QLoRA
>backpropagation & trainable parameters
>model merging & multi-task fine-tuning
getting a much clearer picture of how foundation models can be adapted efficiently for different tasks without retraining everything from scratch