Inspired by the latest AI Demo Days in London, I built this:
Run any AWS Lambda function as a Large Language Model (LLM) tool without code changes using Anthropic's Model Control Protocol (MCP)
Danilo Poccia
Chief Evangelist (EMEA) @ AWS – Serverless, IoT, AI/ML. Few pics, some music. My opinions. Complexity is a science. He/him.
Here's all the info to use Anthropic's Claude 3.7 Sonnet with Amazon Bedrock: how to use “extended thinking”, tool use with reasoning, updates on computer use, thinking blocks, and more.
https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-anthropic-claude-37.html
I updated the multimodal chat to support Anthropic's Claude 3.7 Sonnet with extended thinking mode. Note that by default it uses 16K of budget tokens for reasoning. Edit config.ini to disable/enable model reasoning and set a different budget.
Inspired by the latest AI Demo Days in London, I built this:
Run any AWS Lambda function as a Large Language Model (LLM) tool without code changes using Anthropic's Model Control Protocol (MCP)
Evaluating AI Agents – Nice short course to learn how to set up evaluations and add observability to an AI agent for insights and debugging
https://www.deeplearning.ai/short-courses/evaluating-ai-agents/
Also, have a look at Phoenix, an open-source AI observability platform designed for experimentation, evaluation, and troubleshooting
I went down the rabbit hole of high dimensional orthogonal vectors analysis:
https://github.com/danilop/high-dim-orthogonal-vectors-analysis
With higher dimensions, most random vectors are almost (for example, within 1 degree of tolerance) orthogonal between each other.
Because vectors are used as machine learning embeddings, and cross product is often used to compare the similarity of vector embeddings, this results show that higher dimensions provide exponentially more expressiveness for the concepts linked to those embeddings.
It seems like the curse of dimensionality that negatively affects many algorithms that use vector distances can, for vector embeddings, be of help by providing more space for meanings.
New building trick to blend in with existing architecture #London
Lightweight LLM for converting (unstructured) text to structured data – A novel training procedure and decoding mechanism outperform much larger foundation model prompted to perform the same task:
https://www.amazon.science/blog/lightweight-llm-for-converting-text-to-structured-data
Here's Matheus' AWS Weekly Roundup 👉 AWS Step Functions, AWS CloudFormation, Amazon Q Developer, and more (February 10, 2024)
Introducing JSONL support with AWS Step Functions Distributed Map 👉 Process large collection of items stored in JSONL and optionally exports the output as a JSONL file. This launch also includes support for additional delimited file formats and more control over result formatting.
https://aws.amazon.com/blogs/compute/introducing-jsonl-support-with-step-functions-distributed-map/
How Transformer LLMs Work 👉 Super interesting and visual course to understand the key components of transformers, including tokenization, embeddings, self-attention with a look at recent improvements to the attention mechanism
https://www.deeplearning.ai/short-courses/how-transformer-llms-work/
Here's Donnie's AWS Weekly Roundup 👉 DeepSeek-R1, S3 Metadata, Elastic Beanstalk updates, and more (February 3, 2024)
Amazon EventBridge enhances event source discovery in the console
When you create a rule, The EventBridge console now displays the source and detail type of all available AWS service events. Additionally, the EventBridge documentation now includes an automatically updated list of all AWS service events, facilitating access to the most current information.
This makes it easier for customers to discover and utilize the full range of AWS service events when building event-driven architectures.
Client Info
Server: https://mastodon.social
Version: 2025.04
Repository: https://github.com/cyevgeniy/lmst