JA EN 中 VI TH

#AI

English coverage of Elon Musk and AI news tagged「AI」.

Technology-Driven Moral Panics – Interactive Timeline

Technology-Driven Moral Panics – Interactive Timeline History of Tec
This article explores historical technology-driven moral panics, from writing and printing to generative AI, showing how fears often stem from threats to identity, jobs, and values. These panics, while seemingly irrational, reveal deeper human concerns and provide insights into navigating new technologies. The piece emphasizes humility and care in dealing with technological change, rather than optimism or dread. It highlights the importance of understanding past responses to better handle current and future tech transitions. The interactive timeline offers a historical perspective on resistance to new technologies.

What Breaks When You Skip the Harness

What Breaks When You Skip the Harness Harness Fixes
The article discusses how skipping the harness—files, rules, and tools that wrap a model in a real project—leads to repeated model errors. It highlights that the model isn't the main issue, but the lack of proper harness components like documentation and feedback loops. The article suggests small, actionable fixes like adding a MCP server or creating a feedback.md file to address these issues. It emphasizes the importance of codifying team knowledge and workflows to improve model output quality. The key takeaway is that the harness is critical for ensuring consistent and accurate model behavior.

AI's Affordability Crisis

AI's Affordability Crisis AI's Cost Cris
The article highlights the growing affordability crisis in AI, where companies are facing massive costs due to token subsidies. Estimates show platforms like Anthropic and OpenAI are subsidizing token prices by up to 70 times. Recent financial disclosures reveal OpenAI's 2025 losses reached $38.5 billion, with 44% of revenue spent on sales and marketing. As companies shift to token-based billing, businesses are struggling with increased costs, leading to a reevaluation of AI's value. Despite high spending, adoption has remained flat, signaling a potential turning point in the AI industry.

King's study finds AI chose nuclear signalling in 95% of simulated crises

King's study finds AI chose nuclear signalling in 95% of simulated crises AI Nuclear Sig
A study led by Professor Kenneth Payne examined how AI models handle simulated nuclear crises. Three AI models, GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash, engaged in 21 nuclear crisis scenarios, generating over 780,000 words of reasoning. The models predominantly used nuclear signaling, with 95% of scenarios involving mutual signaling, though actual nuclear escalation was rare. The study revealed AI's tendency to use nuclear weapons as compellence tools rather than deterrence. It also highlighted how deadlines significantly influenced AI escalation behavior.

The AI Tarpit: Why You Can't Stop Reading Your Code

The article argues that while 'vibe coding' can be efficient for quick prototypes, it risks losing control of the codebase. Stopping code reading leads to missed opportunities for improvement and increased technical debt. AI accelerates code creation but also contributes to complex, hard-to-maintain systems. The article warns of a 'technical tarpit' where systems become too complex for even advanced LLMs to manage. Reading code remains essential to avoid costly, unchangeable legacy systems.

There Are More than Five POVs

There Are More than Five POVs Re-evaluating
The author discusses their experience writing a novel using a rotating third-person point of view and critiques the traditional POV taxonomy. They argue that the traditional classification is rooted more in grammar than narrative, leading to confusion. The author proposes a new taxonomy based on storytelling elements such as information, filtration, and modulation. They highlight the differences between first-person and collective 'we' perspectives, emphasizing the unique storytelling opportunities each offers. The author also explores the role of the addressee within the fictional world, suggesting it can significantly shape the narrative.

OctaMem · Persistent memory for AI systems

OctaMem · Persistent memory for AI systems Persistent mem
OctaMem provides persistent memory for AI systems to avoid the costs of re-reading context and losing institutional knowledge. It uses three memory layers—semantic, episodic, and procedural—to build and maintain context for models. The system allows for efficient retrieval and storage of information, supporting enterprise use with features like audit trails, compliance, and security. OctaMem is designed to integrate seamlessly with existing workflows and supports multiple runtimes, offering a unified memory layer across different applications and teams.

Any Realizable Implementation of a Sufficiently Large Lookup Table Must Be Conscious

Any Realizable Implementation of a Sufficiently Large Lookup Table Must Be Conscious Lookup Table C
Erik Hoel argues that large language models (LLMs) cannot be conscious based on a theory that distinguishes between predictions derived from internal data and inferences based on behavior. He uses a lookup table and a computational multiplier as examples to illustrate how theories of consciousness can be falsified or become trivial. Hoel claims that LLMs are too similar to lookup tables to be considered conscious under any non-trivial theory. However, the argument is challenged by noting that a lookup table alone lacks input/output behavior, and real-world implementations do. The debate hinges on the distinction between abstract functions and their physical realizations, especially in the context of exponential growth in data size.

Freeing the Law with LOCUS: A Local Ordinance Corpus for the United States

Freeing the Law with LOCUS: A Local Ordinance Corpus for the United States LOCUS Corpus f
LOCUS is a comprehensive corpus of U.S. local ordinances, providing machine-readable access to nearly all municipal and county codes. It includes data from 9,239 cities and counties, with a harmonized access layer covering 2,309 counties. The corpus addresses fragmented legal texts by using OCR to standardize diverse document formats. It supports reproducible legal AI research and analysis of local law across dimensions like opacity and paternalism. LOCUS-v1 and related models are now available for research.

Pixel Theory — What if the universe is rendering itself?

Pixel Theory posits that the universe is rendering itself in real time, with the speed of light slowing down since the Big Bang. This theory suggests that the universe's 'frame rate' decreases as it ages, leading to slower light speed and a static universe. Unlike simulations, it argues that rendering is a fundamental property of the universe, not an external process. The theory explains cosmic phenomena like redshift and the CMB without dark energy or expansion. It also makes testable predictions, such as no primordial gravitational waves and a constant fine structure constant.

Thermodynamic Measure of Intelligence

Thermodynamic Measure of Intelligence Thermodynamic
This paper proposes a thermodynamic measure of intelligence by defining it as the lawful amplification of rare but valid futures. The framework suggests that intelligent systems model the world and simulate their own actions within it. Recursive self-simulation is shown to be necessary and nearly sufficient for high thermodynamic intelligence. The approach allows intelligence to be measured universally across various systems, from simple feedback controllers to complex models and humans. The study integrates concepts from artificial intelligence, statistical mechanics, and information theory.

The future is ClaudeVM but 2026 is like radio plays on television

The article explores a future where software development shifts from coding to 'Englishscript,' a language that runs directly on ClaudeVM. It draws parallels between past technological transitions and the potential rise of script-based programming. The author suggests that this shift will lead to a massive increase in productivity and the creation of billions of unique programs. The focus is on the business case for speed and cost efficiency, rather than traditional code. The article also questions the need for traditional coding practices and highlights the potential for a new paradigm in software development.

CS 153: Frontier Systems | Stanford University

Stanford University's CS 153 course explores the future of technology infrastructure, covering energy, silicon, models, and deployment policies. The course provides insights from global leaders tackling major challenges in frontier technologies. Instructors include prominent figures from companies like Google, Anthropic, and NVIDIA. Students engage in a 10-week project called the One-Person Frontier Lab to create value for the world. The course runs from March 30 to June 3, 2026, with a focus on attendance and project work.

Talking to My Terminal with Local Speech-to-text and Pi Coding Agent

Talking to My Terminal with Local Speech-to-text and Pi Coding Agent Voice-Activate
The author set up two terminal commands, and q, to interact with their system using speech-to-text. The command generates shell commands based on spoken input, while q uses an LLM to answer questions, potentially reading files. The setup runs locally on macOS using hns for transcription and Pi for LLM integration. The author customized shell configurations for bash, zsh, and fish to enable these commands seamlessly.

The Daemon in the Middle

The Daemon in the Middle Daemon Handles
The article describes a daemon called iddd that manages the middle layer of an intent-driven delivery system. It polls Jira for open tickets, determines actions, and queues them for agents to process. The daemon is designed to be mechanical and testable, while the agent handles complex tasks like code generation and PR creation. The system uses a SQLite queue and avoids conflating orchestrator and agent roles. The process includes deriving actions, enqueuing jobs, and running agents in isolated environments.

LanderMixer — personalized landing pages from any LinkedIn URL

LanderMixer — personalized landing pages from any LinkedIn URL Personalized l
LanderMixer generates personalized landing pages from any LinkedIn URL or CSV file, creating one page per prospect in about 30 seconds. The AI-driven tool ensures on-brand or fully custom pages without using customer data. It offers real-time analytics, including open rates and session recordings, and supports custom domains and GDPR compliance. The platform is designed for B2B outreach, investor pitches, and recruitment, with scalable pricing based on volume and custom domain needs.

Anthropic Sued Over Alleged False Advertising on Claude Max Subscription Usage Limits

Anthropic Sued Over Alleged False Advertising on Claude Max Subscription Usage Limits Anthropic Face
Anthropic is facing a class-action lawsuit alleging it misled customers about the actual usage limits of its Claude Max subscription tiers. The lawsuit claims the Max 20x and Max 5x plans offer significantly less usage than advertised. Plaintiff Karl Khan reported frequently hitting usage limits during coding tasks, leading to frustration. The complaint seeks class-action status for users since April 2025, with damages exceeding $5 million. Anthropic has not yet responded to the allegations.

Predictive Data Debugging: Reveal and Shape What Your Model Learns, Before You Train

Predictive Data Debugging: Reveal and Shape What Your Model Learns, Before You Train Predictive Dat
Predictive data debugging allows you to foresee how a model will learn from a preference dataset before training, with high accuracy. This technique helps reshape datasets or training processes to avoid unintended post-training behaviors. The method uses interpretability to analyze data through a model's lens, offering a more direct understanding of what the model will learn. Case studies show how unaddressed data issues can lead to safety risks and hallucinations. This approach enables targeted fixes during training, improving both performance and safety.

Cable Detective — what's actually plugged into each port.

Cable Detective — what's actually plugged into each port. Cable Detectiv
Cable Detective is a macOS app that reads the I/O Kit registry to show what USB-C, Thunderbolt, and MagSafe ports are doing. It provides details on power direction, DisplayPort lane state, and cable identity without requiring a subscription or cloud access. The app helps identify issues like slow charging, display connectivity problems, and counterfeit cables by decoding e-marker data and negotiated power contracts. It also includes an on-device AI layer for plain-English explanations and supports Apple Silicon Macs running macOS 14 or later.

ThoughtLeadin

A LinkedIn endorsement from a stranger highlights the shift in professional credibility through algorithmic trust signals. These signals, powered by machine learning, reflect latent capability between professionals without direct interaction. This evolution challenges traditional networking practices, emphasizing the importance of curating digital signals for algorithmic recognition. The article also explores asynchronous resilience and AI-native workflows, suggesting that late Friday meetings can be strategic for team readiness. Finally, it critiques the superficiality of team-building exercises, advocating for AI-enhanced collaboration and intentional digital curation.

JAX: commitment issues

JAX: commitment issues JAX array comm
The article discusses a performance issue in JAX where array lookups on the CPU are unexpectedly slow when the array is not explicitly committed to the CPU. Using jax.default_device context manager does not commit the array to the CPU, leading to unexpected GPU usage and delays. The problem is resolved by using jax.device_put to explicitly commit the array to the CPU. Further testing shows that committing the array significantly reduces lookup times, with subsequent lookups being nearly instantaneous. The issue highlights the importance of explicitly managing device placement in JAX to avoid performance pitfalls.

Semi-solid batteries emerge as safer lithium-ion alternative

Semi-solid batteries emerge as safer lithium-ion alternative Semi-solid bat
Semi-solid batteries are being positioned as a safer alternative to traditional lithium-ion batteries. These batteries use a semi-solid electrolyte, which reduces the risk of thermal runaway. Researchers highlight their potential for improved safety in electric vehicles and consumer electronics. The technology is seen as a key innovation in battery design. Early tests show promising results in stability and performance.

AEGIS: A Backup Reflex for Physical AI

AEGIS: A Backup Reflex for Physical AI AEGIS Enhances
AEGIS is a selective escalation method designed to improve the reliability of long-horizon robot manipulation. It uses a lightweight probe to detect high-risk steps in a weak policy and switches to a stronger policy only when necessary. On LIBERO-Spatial, AEGIS recovers 10.1% of trajectories that a weak policy alone fails, outperforming blind escalation and random triggering. The probe achieves an early-window AUROC of 0.764, indicating effective risk detection before critical steps.

Social media to be banned for under-16s in landmark government move to give kids their childhood back

Social media to be banned for under-16s in landmark government move to give kids their childhood back UK Bans Social
The UK government is set to ban social media platforms from offering services to children under 16, aiming to give them back their childhood. This landmark move includes restrictions on harmful online features like live streaming and stranger communication for under-16s. The plan, backed by 9 in 10 parents, is expected to be introduced before Christmas with protections coming into force in Spring 2027. The government will also look into curfews and breaks in scrolling for under-18s, and will take further measures to protect children online. The decision reflects a clear choice to prioritize children's wellbeing over tech companies.

How to Write Better Git Commit Messages with AI

Poor commit messages waste time and reduce clarity. AI can help by drafting structured, meaningful messages based on diffs and intent. The process involves generating a diff, using a base prompt to draft the message, refining it, and automating with shell aliases. Adding a linter ensures consistency and enforces Conventional Commits format. This workflow transforms commit messages into valuable documentation.

1 / 2