ourworldindata.org/grapher/test...

FediOnFire was a hobby project designed to showcase a firehose of public statuses across the Fediverse. Developed by Roni Laukkarinen (@rolle), it aimed to highlight the diversity of the network. Unfortunately, some community members misinterpreted and discredited the project, which led to ugly […]

This JupyterLab extension utilizes GitHub Copilot to boost productivity with code generation, error fixing, and interactive notebook creation. Users can enjoy inline code completions, chat interface for code queries, and extensibility for custom tools […]

DeepSeek R1: All you need to know 🐳

The article covers various aspects of the model, from its architecture to training methodologies and practical applications. The explanations are mostly clear and detailed, making complex concepts like Mixture of Experts (#MoE) and reinforcement learning […]

Apparently AI reasoning models like Deepseek-R1 and OpenAI o1 suffer from "underthinking", where they abandon promising solutions too quickly, leading to inefficient resource use. To address this, a "thought switching penalty" (TIP) was developed, which improved accuracy across math and science […]

Key Points:
➡️ AI errors running amok like a robot revolution, featuring Google search gaffes and Comcast calendar calamities.
➡️ Flubstitution: when AIs confuse template placeholders with Mad Libs.
➡️ Case study in how open-source session recording gets downright mysterious.
➡️ Dell's […]

Maintaining consistent code style is challenging in software development, especially in large or distributed teams. Using .editorconfig in .NET projects helps enforce uniform coding standards automatically, resulting in a professional and uniform codebase with reduced friction in code reviews. […]

Thomas Cherickal asked Google Gemini to produce 30 prompts for OpenAI's 'Deep Research' reasoning model that could change the world as we know it. Deep Research had to break quite a sweat to answer these prompts (up to 30 minutes), and the result is fascinating […]

An interesting take on the "prompt-and-pray" approach, where business logic lives entirely in prompts. This is inherently unreliable and inefficient.

Structured automation, which separates conversational ability from business logic execution, may be more suitable for enterprise-grade […]

A guide for AI engineering.

The article covers various aspects of AI engineering, from foundational skills in mathematics, statistics, and programming to advanced topics like deep learning, generative AI, and LLMs. It is relevant to anyone interested in AI engineering, providing actionable […]

The story of the Agentic Copilot for Obsidian!

Born in 2023, this AI plugin for the Obsidian PKM has already amassed over 370,000 downloads by February 2024. With a range of features from integrating AI models to local offline capabilities, Copilot for Obsidian is paving the way for advanced […]

Brief analysis of DeepSeek R1 and its implications for Generative AI:
➡️ DeepSeek R1 exhibits powerful reasoning behaviors, achieved through scalable Group Relative Policy Optimization (GRPO).
➡️Emergent self-reflection and Chain-of-Thought (CoT) patterns improve reasoning performance […]

An image of a stack of the three "Feynman Lectures on Physics" books. Volume I is on top of the stack

The Feynman Lectures on Physics

✨ Now, anyone with internet access and a web browser can enjoy reading a high quality up-to-date copy of Feynman's legendary lectures.

This edition has been designed for ease of reading on devices of any size or shape; text […]

[Original post on fosstodon.org]

Key Points:
➡️ DeepSeek AI exposed a database containing over 1 million log lines, secret keys, and sensitive information.
➡️ The database allowed full control over operations without authentication.
➡️ DeepSeek has since fixed the security hole.
➡️ The company is now facing scrutiny […]

Key Points:
➡️ Artificial Analysis provides independent evaluations of AI models and API providers.
➡️ The Quality Index ranks models based on various dimensions of intelligence.
➡️ Output Speed measures tokens per second generated by the model.
➡️ Price is calculated as USD per million […]

In response to the #DeepSeek disaster #OpenAI has released the o3-mini model, which has improved reasoning capabilities at a lower cost than its predecessor. It offers faster response times, reduced computational requirements, and three levels of reasoning (low, medium, high). This model is […]

An screenshot of a user's collection of Mastodon lists.

Sharing Mastodon Lists:
➡️ Mastodon lists are private and not directly sharable like BlueSky or Twitter lists.
➡️ Users can authorize an app for READ ONLY access to their Mastodon lists.
➡️ The app lets you select all or individual lists for download as […]

[Original post on fosstodon.org]

DeepSeek R1 is revolutionizing AI with its cost-effective approach. By using specialist models and innovative training methods, it achieves high performance at a fraction of the cost. This challenges the traditional AI economics. More about this technology here […]

A logo for Python decorators.

PythonDecoratorLibrary - Python Wiki

This page is meant to be a central repository of decorator code pieces, whether useful or not 😉. It is NOT a page to discuss decorator syntax!

wiki.python.org/moin/PythonDecoratorLibr...

#Python #Programming #PythonDecorators

Key Points:
➡️ JavaScript does not natively provide an API to cancel a regular Promise.
➡️ Promise.withResolvers() returns an object with a new Promise and functions to resolve or reject it.
➡️ Cancelable tasks can be created using Promise.withResolvers() by exposing a cancel method.
➡️ […]

A Logo for Node.js with TypeScript support

Starting with v23.6.0, Node.js supports TypeScript without any flags. This means you can run TypeScript files directly in Node.js without transpiling. However, there are some limitations to be aware of:
➡️ no support for non-JavaScript language features, JSX […]

[Original post on fosstodon.org]

The DeepSeek Logo

Key Points:
➡️ The full R1 model requires 400GB of disk space and significant CPU, GPU, and RAM resources.
➡️ Distilled models are available, ranging from 1.5-70B parameters, requiring 1-40GB of disk space.
➡️ The tutorial explains how to install and run […]

[Original post on fosstodon.org]

Meta's Community Notes system is effective in addressing misinformation? Yes, until it's not when:
➡️ There is not enough user engagement.
➡️The community response is too late.
➡️The community is poisoned by bad faith actors.
➡️The 'algorithm' works against it.
➡️Facts are NOT a […]

Key Points:
➡️ Meta is a controversial partner in the Social Web Foundation, raising concerns about its influence on the Fediverse.
➡️ The Fediverse has two models: growth Fediverse and movement Fediverse, each with different goals and values.
➡️ The growth Fediverse focuses on technical […]

#Ollama models:
➡️ #CodeLlama offers versatile development assistance with advanced pattern recognition and infilling capabilities.
➡️ Qwen 2.5 Coder transforms debugging by providing intelligent solutions to root causes.
➡️ WizardLM2 excels in rapid prototyping and multilingual […]

Google researchers have developed a new Transformer model called "Titans" that gives language models long-term memory. Titans uses "surprise" as a metric to decide what to remember and when to forget. It outperforms traditional models in handling long texts and contextual tasks. The team plans […]

Scientists have discovered "sunken worlds" deep within Earth's mantle using a new seismographic imaging method. These blobs resemble ancient crust but are found in unexpected locations, posing a mystery to researchers. The new method offers a clearer image of Earth's interior, revealing […]

Key Points:
➡️ ESM3 is a generative language model that reasons over the sequence, structure, and function of proteins.
➡️ It can generate functional proteins far from known proteins, simulating 500 million years of evolution.
➡️ It can generate fluorescent proteins with a chain of […]

Key Points:
➡️ Malicious PyPi package 'pycord-self' targets Discord developers, stealing authentication tokens and creating a backdoor for remote control.
➡️ Introduced in June 2024, downloaded 885 times.
➡️ Token theft and backdoor installation are the primary malicious functions.
➡️ […]

The Unreasonable Effectiveness of Eccentric Automatic Prompts

The paper shows that trivial changes to prompts can significantly impact LLM performance. Automated prompt optimization is a more reliable method for improving performance than manually adding "positive thinking" to prompts.

The […]