We Didn’t Start The Fire #38

🔥Start the year by reading into the latest in AI agents, understanding a basic Flutter concept and some interesting developments in tech.

More into BuildContext

Fireman A came across a very interesting blog about BuildContext, for those of you wondering what this is, this is a very common type that is used across Flutter. If you’ve ever built an app using Flutter, you would have come across this snippet

@override
Widget build(BuildContext context) {}

A very basic understanding is that BuildContext tells us where exactly the widget is in the widget tree, this blog https://chooyan.hashnode.dev/all-i-know-about-buildcontext goes in depth into the close relation of build context with the element tree and internal workings of things like Navigator.of(context) and mounted

Amazon’s Centrally Decentralised Blockchain

Yet another innovation in the centrally decentralised space 😂, Amazon came out with

Amazon Managed Blockchain (AMB), a fully managed AWS service that simplifies building resilient Web3 applications on public and private blockchains. With AMB Access, developers can instantly connect to blockchains like Ethereum, Polygon, Bitcoin, and Hyperledger Fabric without managing infrastructure. AMB Query provides developer-friendly APIs to access real-time and historical blockchain data, eliminating the need for complex ETL pipelines. Seamlessly integrate blockchain data with AWS services and build scalable, secure applications for use cases like token-gated experiences or multichain digital asset wallets.

Hacking the DNS

Kailash Nadh gave a talk on one very intriguing project that he had made as a hobby project.

The premise is that Kailash didn’t seem to happy about the fact that for a simple query like unit coversions or weather data a lot of data was being transmitted (a few KBs), so he built a website that worked with the foundations of DNS where the number of bytes sent are received are very very less. Check out the video below, the talk is quite interesting and unique.

SmolAgents

Hugging Face recently released SmolAgents, a lightweight library designed to ease the creation of intelligent AI agents. It lets devs build agents that are capable of doing tasks like summarization, data retrieval and code execution in about 3 lines of code. It uses Hugging Faces pretrained models to streamline this process and it seems to be pretty intuitive and easy to use.

Read more here.

Overthinking in LLMs

Researchers from the Tencent AI Lab recently introduced the idea of overthinking in o1 and o1 like LLMs. When talking about overthinking they mainly reffer to these models' tendency to expend unnecessary computational resources on simple problems, leading to inefficiencies. For instance, when solving basic arithmetic questions, o1-like models may generate excessively detailed reasoning, using significantly more tokens than traditional LLMs.

To address this, the researchers introduced two metrics: outcome efficiency and process efficiency, which evaluate resource usage by assessing both the correctness of answers and the relevance of intermediate reasoning steps. They proposed a self-training approach that integrates these efficiency metrics directly into the model training process, emphasizing early and accurate responses while preserving reflective capabilities. Strategies such as First-Correct Solutions (FCS) and FCS+Reflection are central to this approach, streamlining computation without sacrificing accuracy. Applying these strategies to the QwQ-32B-Preview model reduced token usage by 48.6% on the MATH500 dataset.

Read more here.

Genesis

Genesis is a comprehensive physics platform designed for general-purpose robotics, embodied AI, and physical AI applications. It integrates several key components:

• Universal Physics Engine: Rebuilt from the ground up, capable of simulating a wide range of materials and physical phenomena.

• Robotics Simulation Platform: Lightweight, ultra-fast, Pythonic, and user-friendly, facilitating efficient robotics simulations.

• Photo-Realistic Rendering System: Provides powerful and fast rendering capabilities for high-quality visual outputs.

• Generative Data Engine: Transforms user-prompted natural language descriptions into various data modalities, automating data generation for robotics and beyond.

Genesis aims to lower the barrier to using physics simulations, unify diverse physics solvers into a single framework, and automate data generation to reduce human effort. It supports cross-platform operation, running on Linux, macOS, and Windows, and is compatible with multiple compute backends, including CPU, Nvidia/AMD GPUs, and Apple Metal. The platform also offers integration of diverse physics solvers such as rigid body, MPM, SPH, FEM, PBD, and stable fluid, enabling the simulation of various materials and physical phenomena.

Take a deeper look at it here.

Mutify

Muteify is a Python script made by one of out firemen to automatically reduce the volume of the Spotify app during advertisements and restore it once music resumes. This tool is intended for Windows users and utilizes the Spotify Web API to detect ads, alongside the PyCaw library to control audio sessions.

Take a look at it here.