Business, Deals & Funding
Claude Code Changelog
v2.1.153
Version 2.1.153 of Claude Code adds a `skipLfs` option for plugin marketplace sources to skip Git LFS downloads, displays a one-time notice when npm global installs can't auto-update (with `/doctor` providing fixes), passes `COLUMNS` and `LINES` environment variables to status line commands for terminal-aware output sizing, and improves `claude agents` autocomplete to suggest native slash commands and bundled skills in addition to project skills.
Why it matters
This is a modest but practical maintenance release. The `skipLfs` option is a nice quality-of-life addition for users who don't need large files from plugin repos. The auto-update notice with `/doctor` integration is thoughtful UX design that helps users self-diagnose installation issues. Passing terminal dimensions to status line commands is a small but useful detail that enables better-formatted output. The agents autocomplete improvement makes the dispatch experience more discoverable. Overa…
DATAVERSITY Smart Data
Your AI Governance Isn’t Broken, Your Checklist Is
The article argues that AI governance frameworks fail not because they are poorly designed, but because they don't translate into actionable, day-to-day decisions. The author describes designing a practical responsible AI impact assessment process—a gate-based checklist tested on two live use cases. Key insights include: starting governance with privacy and data protection rather than abstract principles, which revealed hidden risks early in one pilot; distinguishing between 'blockers' (non-negotiable requirements) and 'conditions' (requirements that can be temporarily accepted with remediation plans), which prevented governance from stalling delivery or being bypassed; and ensuring human oversight is always a mandatory blocker. The core message is that governance must be a design input integrated into workflows, not a late-stage compliance exercise or one-size-fits-all policy artifact.
Why it matters
This is a refreshingly practical take on AI governance that addresses a real and widespread problem. The distinction between blockers and conditions is particularly valuable—treating every governance requirement as equally critical is a common failure mode that leads to either paralysis or quiet non-compliance. Starting with privacy and data mapping rather than abstract principles is also smart, as the sentiment analysis example powerfully illustrates how early data scrutiny can fundamentally r…
TechCrunch AI
Vertu wants CEOs to run companies from an AI foldable starting at $6,880
Luxury smartphone brand Vertu has unveiled the Alphafold, an AI-powered foldable phone starting at $6,880 (up to $46,800 for premium models) targeting CEOs and executives. The device features an AI agent called Hermes Agent, built on the open-source Hermes project by Nous Research, which connects to enterprise systems like ERP and CRM to handle approvals, scheduling, sales tracking, and operational reporting through natural language. It can route requests across multiple AI models including GPT, Claude, and Gemini, and integrates with over 80 apps. The phone includes a proprietary A5 security chip for isolating sensitive data, with on-device processing and data redaction capabilities, though it has not yet undergone third-party security audits. Vertu, the Hong Kong-based company that has changed ownership multiple times, is attempting to reinvent itself for the AI era by combining luxur…
Why it matters
This feels like a classic Vertu play: wrap existing technology in luxury materials and charge a massive premium, but this time with an AI narrative bolted on. The enterprise AI agent concept is genuinely interesting — connecting a mobile device directly to ERP/CRM systems through natural language could be valuable for executives — but there's little reason this couldn't run on any flagship smartphone. The lack of third-party security audits is a significant red flag for a device marketed to CEO…
Guardian AI
Are robots nearing their ChatGPT moment? – podcast
The podcast discusses whether robots are approaching a breakthrough moment similar to ChatGPT's impact on mainstream AI adoption. It highlights a recent milestone where a robot named Lightning beat the human world record at Beijing's half marathon by nearly seven minutes. China is leading the robotics charge, with the government pledging to invest more than £100 billion in robotics over the next 20 years. The episode explores how robots are already entering the workforce and what further developments are needed for them to become part of everyday life, drawing parallels to how chatbots have already become commonplace.
Why it matters
This article reflects the growing excitement and hype around humanoid and AI-powered robotics, particularly driven by China's massive state investment. The comparison to ChatGPT's breakthrough moment is apt — just as large language models suddenly became accessible and useful to the general public, physical robots may be approaching a similar inflection point where cost, capability, and reliability converge to enable widespread adoption. However, the challenges for physical robots are fundament…
Guardian AI
AI ‘art’ is boring, soulless theft – and when I see it as an artist I see red | Jess Harwood
Visual artist Jess Harwood argues that AI-generated art is uncreative, soulless, and constitutes theft from human artists. She values the traditional, hands-on process of creating art and expresses frustration at the rise of AI in creative fields. She contrasts the authentic human experience behind real art—illustrated by her enjoyment of a Split Enz concert—with the hollow output of AI, asserting that using AI would drain meaning from her creative existence.
Why it matters
The author is strongly opposed to AI-generated art, viewing it as boring, soulless theft that undermines genuine human creativity and the livelihoods of real artists.
TechCrunch AI
Why Google’s AI can’t spell Google (or anything else)
Google's AI Overviews feature is again producing embarrassing errors, this time with basic spelling and letter-counting tasks. The AI claims there are two P's in 'Google,' misidentifies letters in common words, and misspells names. These failures stem from how large language models (LLMs) process text using tokens rather than individual letters, meaning they don't truly 'read' words the way humans do. Google acknowledged the issue, calling letter-counting a 'known challenge for LLMs.' Researchers like Matthew Guzdial explain that transformer-based models encode whole words or syllables as tokens without understanding their constituent letters, making spelling tasks fundamentally difficult. The article also notes a previously patched issue where searching 'disregard' produced a prompt-injection response instead of a definition. Despite these persistent problems, Google continues to doubl…
Why it matters
This article highlights a genuinely important and underappreciated limitation of LLMs that has real consequences when deployed at scale in a product used by billions. The fact that Google is rebuilding its flagship search product around technology that cannot reliably count letters in a word is both absurd and telling about the current AI hype cycle. The piece does a good job explaining the technical reasons behind the failure — tokenization means LLMs literally don't see individual letters — a…
NY Times
Pope Leo’s A.I. Vision Might Not Be Strange Enough
The article discusses Pope Leo's encyclical or formal statement on artificial intelligence, arguing that while the Pope offers criticism of A.I., his vision may simultaneously normalize the technology. The author suggests that the Pope's approach to A.I. doesn't go far enough in its strangeness or radicalism, occupying a middle ground as both critic and normalizer of artificial intelligence.
Why it matters
The author appears to believe that Pope Leo's stance on A.I. is insufficiently bold or imaginative. While acknowledging the Pope's critical perspective on artificial intelligence, the writer argues that the papal vision paradoxically serves to normalize A.I. by engaging with it on conventional terms rather than offering a more radically different or theologically provocative framework for understanding the technology. The opinion suggests that a religious leader's response to something as trans…
Ars Technica AI
Websites have a new way to spy on visitors: analyzing their SSD activity
Researchers have developed a new browser-based surveillance technique called FROST (fingerprinting remotely using OPFS-based SSD timing) that allows websites to spy on visitors by measuring subtle timing differences in their solid-state drive activity. The attack uses simple JavaScript interacting with the Origin Private File System (OPFS) to create a contention side channel, measuring I/O latency variations caused by other processes competing for SSD resources. By running these timing measurements through a pretrained convolutional neural network, attackers can determine which websites are open in other tabs (even across different browsers) and which applications are running on the visitor's device. The technique requires no user interaction beyond visiting the attacking site. However, it has limitations: it requires creating an extremely large OPFS file (likely a gigabyte or more), wh…
Why it matters
This is a genuinely concerning development in the ongoing cat-and-mouse game of web privacy. The fact that a website can infer what other sites and applications a user has open—purely through JavaScript running in a browser sandbox—represents a meaningful escalation in fingerprinting capabilities. The contention side-channel approach is clever because it bypasses traditional browser isolation mechanisms by exploiting shared physical hardware rather than software-level boundaries. That said, the…
From X/Twitter
- Halligan makes the case for the five-tool operator: the person who can spec, build, ship, market, and measure — and argues compensation needs to reward it.
- Florian Roth outlines four stages of AI coding hype — from amazement to expansion to the phase most people don't talk about yet.
- Peter Yang got tired of making slides, so he built an AI skill that turns a rough outline into a polished HTML deck in minutes.
- Claire Vo says Codex's /goal command solved all her problems — and put together a 30-minute breakdown of how it works.
- Brian Halligan says a senior engineer at a frontier AI lab now ships in a week what a five-person team shipped in a month — and comp systems haven't caught up.
- John Yeo walks through the agent his team built to automatically query logs and investigate support tickets — one of their biggest internal levers.
From Reddit/HN/YC
- [Hacker News] Taste Skill is an open-source front-end framework designed to keep AI agent outputs from drowning in slop.
- [Hacker News] New research maps the GitHub Actions language — its usage patterns, how workflows evolve, and where reliability breaks down.
- [Hacker News] Microsoft admits it cannot guarantee data sovereignty for cloud customers — a quiet, significant concession.
- [Hacker News] A paper makes the case for putting AI data centres in space, arguing orbital infrastructure solves power and cooling at scale.
- [Hacker News] An open question making the rounds: can you actually train a 1T-parameter LLM with just 8 GPUs?
- [Hacker News] A classic revisited — benchmarking energy efficiency across programming languages reveals C and Rust at the top, Python near the bottom.