.
@IEEESpectrum has just published @RobustAI
's Director of Robotics Benjie Holson's (
@robobenjie
) piece on why programming robots is hard and the pitfalls of building APIs to make is possible for "non-roboticists" to do it.
CODE
-
Programming Robots: Challenges and API Design Pitfalls
By
–
-
Apple’s Context Management and Tool Use with App Intents
By
–
i’m very very curious on how Apple did context management and tool use over ~unbounded set of app intents! if you come across anyone willing to share please send them our way, i’d love to get the engineering stories out of them
-
Local SLM Routing with Server-Side RAG and Function Calling
By
–
Routing w/ a local MM SLM (with maybe lora hot swaps?) and server side larger param with local RAG and function calling basically?
-
Semantic Indexing and Vector Databases in AI Systems
By
–
semantic index sounds a whole lot like a vector database tbh
-
Replicate introduces secret input type for secure sensitive data handling
By
–
We now support a new **secret** input type for securely passing sensitive values to models like passwords and API tokens. Now you can: – Download and upload weights to Hugging Face
– Upload metrics and artifacts to Weights and Biases
– Store file outputs on S3/R2/GCS -
AI Backend Demonstrated as Future Technology Direction
By
–
AI backend is the future. Great demo Ashpreet!
-
LoRA Training 4-8x Faster Than Full Fine-Tuning
By
–
4/ Faster to train! Training with LoRA can easily be 4-8x faster than full fine-tuning.
-
LoRA Fine-Tuning with Predibase Speeds Up Inference 3x
By
–
6/ They are only getting better! LoRA fine-tuning with Predibase makes inference 3x faster compared to the base model.
-
LoRA Training Parameters: Minimal Memory Footprint at 0.2%
By
–
3/ Extremely low memory footprint. LoRA trainable parameters are only 0.2% of total amount of model parameters.