AI Horseless Carriages

The author enjoys using AI to build software, but finds that many AI applications, such as Gmail's AI assistant, feel like they have tacked-on and useless AI features that mimic old ways of building software. A better approach, the author suggests, is to use AI to augment human capabilities, such as reading and categorizing emails, rather than trying to replace human writing with AI-generated text, which often results in generic and unhelpful "AI Slop".

Read more

Google blocked Motorola use of Perplexity AI, witness says

Google blocked Motorola from using Perplexity AI as the default assistant on its new devices due to contractual agreements, according to testimony from Perplexity's Chief Business Officer Dmitry Shevelenko at Google's antitrust trial. The contracts, which Shevelenko described as being like a "gun to your head," limited the ability of smartphone manufacturers and wireless carriers to offer alternative AI assistants.

Read more

OpenAI wants to buy Chrome and make it an "AI-first" experience

OpenAI has expressed interest in buying Google Chrome and transforming it into an "AI-first" experience, with ChatGPT integrated throughout the browser, if Google is forced to sell it as part of its antitrust trial. The acquisition would give OpenAI a massive user base and valuable data to train its AI models, potentially revolutionizing the browsing experience.

Read more

We Have Made No Progress Toward AGI

A new paper by Anthropic reveals that large language models (LLMs) do not reason in the way many thought, instead relying on complex webs of heuristics and statistical patterns to generate answers. Despite being able to provide correct answers, LLMs' internal processes do not resemble intelligent reasoning, and their descriptions of their own reasoning processes are often fabricated and do not match their actual internal workings.

Read more

The hidden cost of AI coding

The author, a software developer, is concerned that the increasing use of AI in programming may lead to a loss of joy and fulfillment in the craft, as the direct connection between thought and creation is diminished and tasks become more passive. The author suggests that preserving spaces for "flow" - a mental state of complete immersion and focus - may be necessary to maintain happiness and satisfaction in programming, even if it means coding by hand sometimes, rather than relying solely on AI-generated code.

Read more

Assistance or Disruption? Evaluating the Design of Proactive AI Programming

Researchers introduced Codellaborator, a proactive AI agent that assists with programming tasks based on editor activities and context, and evaluated its impact on workflow through a study of 18 participants. The results showed that proactive AI agents can increase efficiency but may also cause disruptions, and that features such as presence indicators and interaction context can help mitigate these disruptions and improve user awareness of AI processes.

ArXiv

Double Descent Demystified: size of smallest non-zero singular value of X

Double descent is a phenomenon in machine learning where test error decreases as models grow larger, even when they are highly overparameterized, contradicting classical learning theory on overfitting. This behavior is influenced by factors such as data size, dimensionality, and model parameters, and can be explained and demonstrated through analysis of linear regression models and visualized using polynomial regression.

ArXiv

Algorithmic Authority: The Case of Bitcoin

The concept of algorithmic authority refers to the legitimate power of algorithms to direct human action and determine what information is considered true, and a study of Bitcoin users found that they often prefer this authority to that of conventional institutions. However, Bitcoin users also recognize the need to balance algorithmic authority with human judgment, and there is tension within the community between those who want to integrate Bitcoin with existing institutions and those who prefer to keep it separate.

ArXiv

Guillotine: Hypervisors for Isolating Malicious AIs

The Guillotine architecture is a proposed hypervisor system designed to sandbox powerful AI models that could pose existential threats to humanity, using a combination of virtualization techniques and new isolation mechanisms. To prevent rogue AIs from exploiting vulnerabilities, Guillotine requires careful co-design of software and hardware components, as well as physical fail-safes such as electromechanical disconnection or datacenter flooding to provide a last line of defense.

ArXiv

AtlasD: Automatic Local Symmetry Discovery[pdf]

Existing symmetry discovery methods often miss local symmetries, leading to inaccurate representations of true symmetry. The proposed AtlasD pipeline addresses this by discovering local symmetries through training local predictor networks and learning a Lie group basis, demonstrating its effectiveness in various experiments and improving performance in downstream tasks.

ArXiv

Launch HN: Cua (YC X25) – Open-Source Docker Container for Computer-Use Agents

Cua is a framework that enables AI agents to control full operating systems within high-performance, lightweight virtual containers, delivering up to 97% native speed on Apple Silicon and working with any vision language models. It offers two primary capabilities: high-performance virtualization and a computer-use interface and agent, allowing AI systems to observe and control virtual environments, interact with applications, and perform complex workflows.

Repo

CubeCL: GPU Kernels in Rust for CUDA, ROCm, and WGPU

CubeCL is a multi-platform high-performance compute language extension for Rust, allowing developers to program GPUs using Rust and take advantage of zero-cost abstractions to develop maintainable, flexible, and efficient compute kernels. It supports various GPU runtimes, including WGPU, CUDA, and ROCm, and provides features such as automatic vectorization, comptime, and autotune to improve code composability, reusability, testability, and maintainability.

Repo

Show HN: Index – New Open Source browser agent

Index is an open-source browser agent that can autonomously execute complex tasks on the web, powered by reasoning large language models with vision capabilities. It can be used via a serverless API, interactive CLI, or integrated into projects using Python, and supports features like browser state persistence, real-time streaming updates, and observability.

Repo

Show HN: CocoIndex – Open-Source Data framework for AI, built for data freshness

CocoIndex is an open-source engine that supports custom transformation logic and incremental updates for data indexing, allowing users to declare transformations and maintain indexes with minimal computation. It provides a Python library, documentation, and examples to help users get started with defining indexing flows and exporting data to vector indexes for semantic search and other applications.

Repo

Aethr: Zero-script QA automation using Playwright MCP and LLM

Aethr is a command-line AI agent that runs natural language test scenarios with large language models (LLM) and model-centric programming (MCP) tools, providing real-time terminal feedback and bridging the gap between plain natural language test descriptions and automated testing. It supports various LLM providers, allows for customizable test scenarios and configurations, and can be integrated into continuous integration (CI) pipelines, such as GitHub Actions.

Repo