Norway imposes near ban on AI in elementary school

Norway is implementing a near-total ban on generative AI for elementary students (ages 6-13) starting August 2026 to mitigate negative impacts on foundational learning in core subjects. While younger pupils are restricted, lower secondary students may use AI tools under strict supervision, and upper secondary students will be taught appropriate usage for professional preparation. This policy shift accompanies a broader return to physical textbooks and a ban on social media for children under 16.

Read more

AI Engineer Claims to Have Cracked Linear A

AI engineer Tom Di Mino claims to have deciphered the Minoan Linear A script, identifying it as an extinct Semitic language related to biblical Hebrew. Utilizing Claude Code, Di Mino developed Python scripts to automate the cross-referencing and analysis of the digitized Linear A corpus, enabling systematic hypothesis testing at scale. The research produced a lexicon of 408 terms and proposed phonetic values for 40 signs, including 13 previously unknown Linear A symbols.

Read more

Is AI ruining our skills? Early results are in – and they're not good

Recent studies indicate that reliance on AI tools is causing "deskilling" among highly trained professionals, including physicians and software engineers. Clinical data shows that physicians' manual diagnostic accuracy significantly declined when AI assistance was removed after a period of use, suggesting a loss of cognitive focus and independent proficiency. Researchers are now investigating how to maintain human expertise as AI integration becomes standard across technical and medical workflows.

Read more

Companies rein in AI usage as costs strain budgets

Enterprises are scaling back AI initiatives as high operational and implementation costs begin to strain corporate budgets. This shift reflects a transition from experimental adoption to a more disciplined focus on ROI and cost management. Market concerns regarding the financial sustainability of large-scale AI projects are further evidenced by significant stock fluctuations in the professional services sector.

Read more

Generative AI Is Having Its Herbalife Moment

Replit and Cursor are leveraging influencer marketing to promote "vibe-coding" to non-technical audiences, a strategy the author likens to predatory MLM and crypto schemes. This marketing exploits economic anxiety by promising easy software entrepreneurship while ignoring critical risks such as security vulnerabilities, GDPR liability, and unpredictable LLM token costs. The author argues that selling the dream of successful app development to non-coders is unethical, as it obscures the technical and financial realities of building and scaling production-grade software.

Read more

From AGI to ASI

The report examines the transition from AGI to ASI, defining the latter as intelligence exceeding that of large human organizations. It identifies four development pathways—scaling, paradigm shifts, recursive self-improvement, and multi-agent collectives—while addressing potential technical bottlenecks. Rather than a single step change, the authors suggest a series of AI-driven breakthroughs across science and technology, necessitating global interdisciplinary preparation.

ArXiv

Thermodynamic Measure of Intelligence

This paper defines intelligence as the "lawful amplification of rare but valid futures," quantified through a thermodynamic measure of lift. It demonstrates that recursive self-simulation is necessary and nearly sufficient for high intelligence, as systems must model their own actions within the world to identify and actuate unlikely trajectories. This framework establishes a universal metric applicable to entities ranging from LLMs to information engines.

ArXiv

GitHub Copilot and Dev Productivity: An Observational Dose-Response Analysis

A study of 16,223 Microsoft engineers over 43 weeks utilized a Poisson Pseudo-Maximum Likelihood model with two-way fixed effects to isolate the impact of GHCP on productivity. Controlling for individual skill and active coding time, the research found that high GHCP usage leads to a 40.5% increase in PR completion efficiency. The results demonstrate a monotonic gradient with diminishing returns and remained robust across seven falsification tests targeting potential confounds like task reallocation and PR slicing.

ArXiv

ISAC Privacy: Challenges and Solutions for 6G

ISAC in 6G networks leverages mmWave and sub-THz frequencies for high-resolution spatial sensing, introducing privacy risks across location, behavioral, and physiological data levels. This work categorizes these sensing capabilities, identifies challenges such as bystander exposure and profiling, and outlines research directions for developing privacy-preserving ISAC frameworks.

ArXiv

NanoEuler – GPT-2 scale model in pure C/CUDA from scratch

nanoeuler is a GPT-2-class LLM implemented from scratch in C and CUDA, featuring hand-written forward and backward passes without external ML libraries or autograd. The architecture incorporates modern components such as RoPE, SwiGLU, GQA, and a custom FlashAttention kernel, all validated via double-precision gradient checks. The project provides a complete pipeline for pretraining and SFT on consumer hardware, serving as a transparent educational artifact for end-to-end model engineering.

Repo

I built an 11-LLM consensus engine to detect AI hallucination

The Multi-LLM SaaS Starter Kit is a production-ready boilerplate featuring 14 provider integrations and a semantic consensus engine that uses embedding similarity to detect hallucinations. It includes 13 self-evolution loops for MoE routing, RLHF, and genetic prompt evolution, alongside EU AI Act audit-grade compliance tools and fail-closed gates for high-risk actions. The kit provides a full infrastructure stack—including auth, Stripe billing, and Firestore—designed to accelerate the deployment of regulated or high-reliability AI SaaS products.

Repo

Pipeline-parallel LLM inference across GPUs on separate machines

Shard is an LLM inference engine that enables pipeline-parallel execution of models too large for a single GPU across geographically dispersed machines over the open internet. It achieves ~30 tok/s for GLM-5.2 744B and ~40 tok/s for gpt-oss-120B over WAN by leveraging pipelined speculative decoding with a CUDA-graphed draft model. The core innovation lies in treating WAN round-trip as the scarce resource, using async pipelining to hide latency and make the system throughput-bound, alongside secure, authenticated transport of activation tensors. Shard aims for decentralized, uncensored inference while addressing privacy concerns for intermediate activations.

Repo

AI Agent on Android

RikkaHub Agent is an Android fork of RikkaHub that transforms a native LLM chat client into a versatile on-device agent with over 80 system tools. It enables AI-driven automation through workflows, scheduled tasks, SSH, Termux integration, and a remote Telegram bot interface. The platform supports local inference via LiteRT and Gemini Nano, integrates with MCP servers, and features a multi-layered safety framework for granular tool permissions.

Repo

Wyolet Relay – high throughput, open source LLM router

Wyolet Relay is a self-hosted, scalable gateway that provides a unified OpenAI- and Anthropic-compatible endpoint for multiple LLM providers. It enables key pooling for automatic failover and load balancing, granular rate limiting, and comprehensive cost tracking across more than 400 models. Designed for high-throughput production environments, it offers sub-2ms added latency and integrates with existing SDKs as a drop-in proxy.

Repo