Qualcomm to acquire Arduino

Qualcomm Technologies has agreed to acquire Arduino, a leading open-source hardware and software company, to accelerate developers' access to its edge computing and AI technologies. The acquisition will combine Qualcomm's leading-edge products with Arduino's vast ecosystem and community, enabling millions of developers to create intelligent solutions faster and more efficiently, with new tools such as the Arduino UNO Q board and Arduino App Lab development environment.

Read more

Deloitte to refund the Australian government after using AI in $440k report

Deloitte will provide a partial refund to the Australian government for a $440,000 report that contained several errors, after admitting it used generative artificial intelligence to help produce it. The report, which reviewed the targeted compliance framework and IT system used to automate penalties in the welfare system, was found to contain "hallucinations" where AI models may have filled in gaps, misinterpreted data, or tried to guess answers, leading to incorrect references and citations.

Read more

Robin Williams' daughter pleads for people to stop sending AI videos of her dad

Zelda Williams, the daughter of late actor Robin Williams, has pleaded with people to stop sending her AI-generated videos of her father, calling the practice "dumb" and "a waste of time and energy". She criticized the use of AI to recreate her father's voice and likeness, saying it's "personally disturbing" and not what he would have wanted, and also spoke out against the broader trend of using AI to animate images of deceased people.

Read more

America is now one big bet on AI

The provided text appears to be a webpage from the Financial Times, with various sections and articles listed, but there is no specific article or text to summarize. However, one headline reads "America is now one big bet on AI", which suggests that the US is heavily investing in artificial intelligence, but the full article is not provided due to a paywall.

Read more

GPT-5-Codex is a better AI researcher than me

The author attempted to train the strongest AI model possible on a laptop in five minutes and was able to train a 1.8M param transformer with the help of GPT-5, producing simple stories. With the assistance of Codex, a more advanced AI tool, the author was able to train a better model, producing more coherent and longer stories, demonstrating the potential of "vibe research" where humans collaborate with AI models to perform complex technical tasks.

Read more

Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning

Researchers have successfully scaled up evolution strategies (ES) for fine-tuning large language models (LLMs), finding that it can efficiently search over billions of parameters and outperform existing reinforcement learning (RL) methods in several areas. This breakthrough unlocks a new direction in LLM fine-tuning, offering advantages such as improved sample efficiency, robustness, and stability, and providing an alternative to traditional RL techniques.

ArXiv

The Missing Link Between the Transformer and Models of the Brain

The Dragon Hatchling (BDH) is a new Large Language Model architecture inspired by the brain's scale-free biological networks, offering strong theoretical foundations, interpretability, and performance comparable to Transformer models like GPT2. BDH's biologically plausible design, which relies on synaptic plasticity and Hebbian learning, allows for sparse and positive activation vectors, enabling interpretability of state and demonstrating monosemanticity in language tasks.

ArXiv

Pretraining with hierarchical memories separating long-tail and common knowledge

Modern language models' performance relies on scaling parameters, but this approach is impractical for edge devices with limited resources, as it stores unnecessary world knowledge. A proposed memory-augmented architecture uses small language models that access large hierarchical parametric memory banks, achieving comparable performance to larger models while reducing parameter requirements.

ArXiv

HSGM: Hierarchical Segment-Graph Memory for Scalable Long-Text Semantics

The Hierarchical Segment-Graph Memory (HSGM) framework is introduced to tackle the challenge of semantic parsing of long documents by decomposing the input into meaningful segments and constructing local semantic graphs, reducing complexity from $O(N^2)$ to $O(Nk + (N/k)^2)$. HSGM achieves significant improvements in inference speed, memory reduction, and maintains high accuracy, enabling scalable and accurate semantic modeling for ultra-long texts and real-time NLP applications.

ArXiv

Anti-establishment sentiment on TikTok influencers and expertise on social media

Distrust of institutions is increasing, particularly in the US, and social media may be contributing to this trend as content creators often position themselves as experts and dismiss institutional authority. A study of TikTok found that anti-establishment sentiment is most prevalent in conspiracy theory content, but also discovered that platform incentives may encourage users to post such content, even in areas like finance and wellness where it is relatively rare.

ArXiv

Launch HN: LlamaFarm (YC W22) – Open-source framework for distributed AI

LlamaFarm is an open-source framework for building retrieval-augmented and agentic AI applications, allowing developers to build powerful AI models locally and extend them anywhere. It provides a flexible and extensible architecture, with a simple CLI and production-ready endpoints, enabling developers to easily swap out components, such as runtimes, embedders, and databases, without rewriting their application.

Repo

Show HN: LLM-Use – An LLM router that chooses the right model for each prompt

LLM-Use is a production-ready intelligent routing system that automatically selects the optimal Large Language Model (LLM) for each task, featuring real-time streaming, A/B testing, quality scoring, and comprehensive analytics. The system supports multiple LLM providers, including OpenAI, Anthropic, and Google, and allows for customizable configuration, cost optimization, and advanced quality scoring using natural language processing techniques.

Repo

Show HN: Self-hosted gateway for video generation APIs (Sora/Runway/Kling)

MediaRouter is an open-source video generation gateway that provides a unified API for multiple AI video generation providers, including OpenAI's Sora, Runway, and Kling, allowing users to switch between providers with a single API and maintain control over their API keys and data. The platform offers features such as cost transparency, customizable workflows, and a modern React UI, and can be easily deployed using Docker, with a setup time of approximately 30 seconds.

Repo

AI ML Jargon

This text is a glossary of AI and machine learning terms, explained in plain English with the help of Rick and Morty asides, covering topics such as supervised, unsupervised, and reinforcement learning. The glossary aims to provide minimal fluff and maximum signal, with code snippets and gotchas, to help practitioners apply ideas immediately in real-world work, including code reviews, design docs, and debugging.

Repo

Open Agent Specification (Agent Spec): A Unified Representation for AI Agents

Agent Spec is a portable, platform-agnostic configuration language that allows Agents and Agentic Systems to be described with sufficient fidelity, and it is supported by SDKs in various languages, including Python, through PyAgentSpec. PyAgentSpec is a Python SDK that enables users to build Agent Spec-compliant agents in Python, define assistants by composing components, and export them to YAML format, with the goal of facilitating the process of building framework-agnostic agents programmatically.

Repo