Daisy, an AI granny wasting scammers' time

O2 has created an AI 'Granny' named Daisy, trained to answer calls from scammers and keep them on the phone for as long as possible, wasting their time and preventing them from targeting real victims. The AI, which is indistinguishable from a real person, is part of O2's "Swerve the Scammers" campaign to combat the UK's growing fraud epidemic.

Read more

AI makes tech debt more expensive

The increasing use of AI in coding has significantly increased the real cost of carrying tech debt, as companies with high-quality codebases can benefit more from generative AI tools, while those with legacy codebases struggle to adopt them. To take full advantage of AI tools, companies should focus on refactoring legacy code and adopting modular architectures that allow for easier integration with AI.

Read more

MomBoard: E-ink display for a parent with amnesia

The author set up an e-ink display in their mom's apartment to help her cope with permanent anterograde amnesia, which prevents her from forming new long-term memories. The display, a BOOX Note Air2 Series device, shows messages from her children to alleviate her anxiety about their well-being, and can be remotely updated with new messages and software changes.

Read more

The barriers to AI engineering are crumbling fast

The barriers to AI engineering are rapidly decreasing as tools become simpler and workflows become standardized, making it possible for anyone familiar with IDEs and basic deployment workflows to build production-ready AI applications. The process can be broken down into six building blocks: models, prompts, knowledge, integrations, tests, and deployment, which can be managed using existing tools and infrastructure.

Read more

AI progress stalls as OpenAI, Google and Anthropic hit roadblocks

Here is a 2-sentence summary of the article:

Leading AI companies OpenAI, Google, and Anthropic are facing unexpected hurdles in developing more sophisticated models, with OpenAI's latest model, Orion, failing to meet performance expectations and Google's Gemini software facing challenges. The setbacks are attributed to difficulties in finding high-quality training data and the high costs associated with developing new models, raising concerns about the validity of the "scaling laws" theory that more computing power and data will lead to significant AI advancements.

Read more

How AI is beating VCs in their own game

GPTree is a novel framework that combines the explainability of decision trees with the advanced reasoning capabilities of large language models (LLMs), allowing for more accurate and interpretable decision-making. The framework achieved a 7.8% precision rate in identifying "unicorn" startups, outperforming both human decision-makers and other AI models.

ArXiv

GPT or BERT: why not both?

Researchers have developed a hybrid training objective that combines masked language modeling and causal language modeling, resulting in a flexible model that outperforms single-paradigm models. The hybrid model, tested on the BabyLM Challenge 2024, has been made publicly available along with its training data and code.

ArXiv

Why is it so hard to find a job now? Enter Ghost Jobs

A study found that up to 21% of job ads may be "ghost jobs" where employers have no intention of filling the position, often in specialized industries and larger firms. This practice can cause job fatigue and distort market signals, and may be contributing to the recent disconnect in the Beveridge Curve, a labor market metric.

ArXiv

BERTs Are Generative In-Context Learners

Researchers found that masked language models, like DeBERTa, can also perform generative tasks through a simple inference technique, without additional training or changes. This discovery reveals that masked and causal language models have complementary strengths, suggesting that a hybrid approach combining both could be beneficial.

ArXiv

Language agents achieve superhuman synthesis of scientific knowledge

Researchers developed a language model called PaperQA2, optimized for accuracy, and found it matched or exceeded human performance on literature research tasks, including writing more accurate Wikipedia-style summaries. PaperQA2 also identified contradictions in scientific papers, with 70% of its findings validated by human experts, demonstrating its potential to exceed human capabilities in scientific research.

ArXiv

Show HN: Memoripy – AI Memory Layer with Short- & Long-Term Memory, Clustering

Memoripy is a Python library that manages and retrieves context-aware memory interactions using both short-term and long-term storage, supporting AI-driven applications with compatibility for OpenAI and Ollama APIs. It features contextual memory retrieval, memory decay and reinforcement, hierarchical clustering, and graph-based associations to aid in contextually relevant retrieval.

Repo

Machine Learning Algorithms in Depth

This text describes the contents of the book "Machine Learning Algorithms in Depth" by Vadim Smolyakov, covering various machine learning algorithms, including Bayesian inference, deep learning, and unsupervised learning. The book is divided into chapters, each focusing on a specific topic, and includes code examples in Python to illustrate the concepts.

Repo

Simple Premade Local AI Assistant

SPLAA is an AI assistant framework that utilizes voice recognition, text-to-speech, and tool-calling capabilities to provide a conversational and interactive experience. It uses LLMs available through Ollama and has capabilities for extending functionalities through a modular tool system.

Repo

ErisForge: A Python library to customize and control LLM by Abliteration

ErisForge is a Python library that allows users to modify Large Language Models (LLMs) by applying transformations to their internal layers, creating ablated or augmented versions that respond differently to specific inputs. The library provides features such as modifying internal layers, measuring refusal expressions, and supporting custom behavior directions, and can be used to induce different response behaviors in LLMs.

Repo

AG2 (formerly AutoGen) is a programming framework for agentic AI

AG2 (formerly AutoGen) is an open-source programming framework for building AI agents and facilitating cooperation among multiple agents to solve tasks. It offers features such as agents capable of interacting with each other, facilitates the use of various large language models (LLMs) and tool use support, autonomous and human-in-the-loop workflows, and multi-agent conversation patterns.

Repo