Wiola is a fully original small language model architecture built from first principles, unrelated to existing families like GPT, LLaMA, Mistral, or Falcon. It introduces five novel components: Spiral Rotary Positional Encoding (SRPE), Gated Cross-Layer Attention (GCLA), Adaptive Token Merging (ATM), Dual Stream Feed-Forward (DSFF), and WiolaRMSNorm. Released in four sizes (120M to 1.5B parameters), it is fully compatible with HuggingFace Transformers.

Auto-FL-Research (AFR) is a constrained coding-agent workflow for federated learning algorithmic recipe search. It is evaluated on five healthcare FLamby tasks and six LEAF profiles, showing gains on multiple tasks but also revealing seed-sensitive and search-selected failures.

PACE is a modular neuro-symbolic framework that generates feasibility-aware counterfactual explanations by combining neural predictive models with symbolic reasoning. It enforces domain-specific constraints to produce realistic and actionable recommendations, as demonstrated on the Adult Income dataset.

An investigation into NVIDIA's DGX Station reveals its true capabilities for running large AI models locally, including memory architecture details, real-world benchmarks, and community skepticism about its 748GB coherent memory claim.

Imagent is an open-source tool that integrates image, video, and speech generation into AI agent workflows. It provides a unified CLI interface supporting multiple AI providers (e.g., OpenAI, Google, ElevenLabs) and manages generated assets in a local library for reuse.

Interfaze open-sourced diffusion-gemma-asr-small, a multilingual ASR model that uses a diffusion decoder instead of autoregressive decoder. It adds audio to Google's frozen DiffusionGemma using a small adapter of ~42M parameters, covering six languages. Transcription cost scales with denoising steps, not transcript length. Achieves 6.6% WER on LibriSpeech test-clean, leading diffusion peers but trailing autoregressive Whisper.

This article applies the 'Moneyball' concept—using data-driven statistical analysis to find undervalued assets—to the field of Physical AI. It argues that robot data is currently mispriced, with overemphasis on volume and teleoperation hours rather than data novelty and marginal utility. By analyzing scaling laws and the economics of data collection, it proposes that capital efficiency in Physical AI depends on accurately computing and pricing data novelty, not maximizing data volume.

Powered by chipmakers Samsung Electronics and SK Hynix, South Korea is seeing a surge in wealth, but there are questions over who gets to share in the profits.

In an internal town hall, Meta CEO Mark Zuckerberg admitted that the company's agentic AI development is slower than expected—a rare candid acknowledgment. He revealed that AI agent progress hasn't accelerated as hoped and that workforce restructuring has not yielded clean results. Despite significant capex increases, shopping agents remain absent. Zuckerberg expects substantial benefits in 3-6 months, but analysts note success is a matter of time.

A video discussing the reasons behind the perceived collapse of AI and how China is gaining an advantage in the field.

Yann LeCun, former Meta AI chief, argues that current LLMs like ChatGPT lack true intelligence and cannot handle real-world complexity. His new company AMI Labs is developing JEPA, a different AI approach. Meanwhile, Oxford's Ingmar Posner is working on World Models. Both aim for AI that can reason about the physical world.

With the proliferation of AI-generated content in professional settings, the author calls for basic etiquette among LLM users: validate outputs, paraphrase, and attribute sources. A hypothetical workplace story illustrates the pitfalls of blindly copying AI responses.

Shadow arts minister Angie Bell, a former professional musician, demands AI companies ask permission and pay for using Australian creative works. Also: teenager charged with murder, X Corp says no harmful content to restrict, Socceroos face Egypt, toll refunds for Victorian drivers, and Australians view Israel more negatively than China.

A potential deal with the government would allow international tech companies to mine the creative work of Australian musicians. Some of the prime minister’s favourite artists told the Guardian how they feel about it.

A new paper by Yoshua Bengio et al. presents a formal safety argument for the Scientist AI (SAI) Predictor, trained to approximate the Bayesian posterior on epistemically contextualized statements, enabling honest predictions without the predictor itself being an agent. It proves that under plausible assumptions, the probability of training a harmful predictor is small.

Billionaire tech investor Peter Thiel delivered a series of provocative warnings and predictions about the future of artificial intelligence and the West on Tuesday, accusing Pope Leo XIV of inadvertently serving as a “Chinese communist agent” by calling for AI regulation. He also warned of a “democratic-socialist takeover” of the United States’ Democratic Party.

Charity Majors argues that while AI mandates are often unpopular, they can be an effective organizational tool when properly funded and supported. She contrasts mandates that provide resources and cover for learning with those that are punitive and underfunded, emphasizing the importance of consistent leadership and clear tradeoffs.

AI didn’t just make content cheap. It made authenticity expensive. This article explores the shift from trust to suspicion, the tension between valuing information and valuing human effort, and the emergence of a 'proof economy' where verifying authenticity becomes a new cost.

DocuPipe's free AI tool converts any PDF (scanned, digital, or photo) into a clean Excel workbook. Each table becomes its own sheet, key fields go into a summary sheet, and you preview exactly what you download. No signup required.

Vox is an open-source CLI extension that lets you talk to GitHub Copilot out loud, using voice commands to interact with the AI agent hands-free. It works on Windows, macOS, and Linux.

Vercel Chief of Software Andrew Qu explains why agents represent a new software paradigm, the evolution of Vercel's agent framework eve, the importance of skills for up-to-date knowledge, and the need for an agent-readable web.

Rainslice is an AI platform that runs your entire home services business 24/7, handling calls, texts, dispatch, quotes, and customer service automatically.

A $1.3 million cargo theft near Chicago highlights the growing risk to the physical supply chain for AI infrastructure, as thieves target copper wire and data center equipment.

A detailed experiment comparing five Claude models (Opus, Fable, Sonnet, Sonnet 4.6, Haiku) on a full audit of the LangChain Python monorepo. Fable matched Opus in grade (A-) but excelled in generating actionable milestones and quick wins. The article presents findings, strengths/weaknesses, and recommends a multi-model pipeline.

OpenAI Group PBC has reportedly proposed giving a 5% stake to the U.S. government during talks with Trump administration officials. The discussions are at an early stage and may be part of a broader AI policy initiative that could also involve stakes in other AI developers. The move may require congressional approval.

Gist is an AI tool that summarizes arXiv papers into concise slide decks with layered insights, including gist, logic, counter-arguments, and steelman rebuttals. Built with a custom model trained on $130k of AWS Claude credits, it offers a free feed of summaries with TTS audio.

An AI agent that reproduces GitHub issues for you.

Mirrors replays production traces in an isolated environment to test AI agent changes safely before deployment. It creates deterministic replicas, scores tool call accuracy, and catches regressions without touching live systems.

This tutorial walks through building a multimodal retrieval pipeline using RAG-Anything in Google Colab. It covers environment setup, securely entering an OpenAI API key, generating a synthetic multimodal report with a chart and PDF, converting content into RAG-Anything's content_list format, inserting it into the retrieval system, and testing different retrieval modes (naive, local, global, hybrid).

A curve-fitting game where you adjust neural network weights and biases to match a target curve. Time-limited levels challenge you to achieve a 90% fit.