Monday — June 24, 2024
An AI enthusiast rains hats on New Yorkers using object detection, llama.ttf font doubles as LLM enabling local AI text generation, and BrainyAI integrates top LLMs into a free Chrome extension for seamless web assistance.
News
Using AI to drop hats outside my window onto New Yorkers
An enthusiastic dev has implemented an AI-driven "Drop of a Hat" service in midtown NYC, using a Raspberry Pi, an Adafruit stepper motor, and object detection via Roboflow. The system, activated by detecting a person standing in a designated spot for 3 seconds, drops propeller hats onto pedestrians’ heads from their apartment window.
Llama.ttf: A font which is also an LLM
llama.ttf is a unique font file that doubles as a LLM and its inference engine, leveraging the HarfBuzz font shaping engine with Wasm support. This setup allows the font to run inside any Wasm-enabled, HarfBuzz-based application like text editors or email clients, making AI text generation features accessible locally without vendor integration. The implementation hinges on using arbitrary code to handle text shaping, turning symbols into LLM triggers for generating content.
Research
Are LLMs Naturally Good at Synthetic Tabular Data Generation?
LLMs have shown great performance in generating text and images but fall short as synthetic table generators, even with fine-tuning. Their autoregressive nature and random order permutation hinder their ability to model functional dependencies and capture conditional mixtures of distributions. However, by making them permutation-aware, these shortcomings can be mitigated.
Assessing the Emergent Symbolic Reasoning Abilities of LLMs
Open-source LLMs from the Llama 2 family were evaluated on symbolic reasoning tasks involving mathematical formulas. Assessments included a generalist model (Llama 2 Chat) and two fine-tuned versions (MAmmoTH and MetaMath). Results showed that both larger models and task-specific fine-tuning significantly improved performance, especially on simpler formulas. However, even the largest fine-tuned models faced difficulties with these supposedly easier problems.
Code
All-in-one AI web assistant (open-source)
BrainyAI is a free Chrome extension that integrates various LLMs, such as GPT-4, GPT-4o, Claude, Gemini, Moonshot, and LLaMA3, into a convenient sidebar. It offers AI chat aggregation, AI search, web and YouTube summaries, and document conversations—all without leaving your current web page. It's a cost-free alternative to tools like Sider, Monica, and Merlin.
Open-source version of Anthropic's Artifacts UI
E2B’s has released on open-source version of Anthropic’s Artifacts Generative UI feature. It uses their Code Interpreter SDK for AI code execution.
The Abstraction and Reasoning Corpus
The ARC-AGI repository offers a dataset and browser-based interface for solving tasks associated with general artificial intelligence. It features training and evaluation data split into 400 tasks each, aimed at both human and AI test-takers.