AI Reading for Saturday September 13

California passed SB 243 regulating AI companions, requiring periodic disclosures and self-harm protocols, with private enforcement; sent to Governor Newsom’s for signature. - ExtremeTech

For the second year in a row, Gavin Newsom has to decide whether to sign landmark regulation that could strain his ties to the tech industry and help set a national standard; Newsom vetoed a broader version last year, saying, “We dominate this space, and I don’t want to lose that competitiveness.” - Politico

Nvidia’s data center compute revenue fell 0.9% quarter-on-quarter in Q2 fiscal 2026—the first sequential decline in more than a decade—even as the company reported record overall revenues. - TechRadar

It's pedal to the metal as tech executives at a Goldman Sachs conference reported strong AI demand and compute constraints, with emphasis on resource prioritization and identifying sometimes-surprising partnerships. - Business Insider

Microsoft wants to build its own AI chips and models, aiming for more self-sufficiency (like Google) - Windows Central

Bret Taylor’s statement a couple of days ago. - OpenAI

OpenAI launched a five-week mentorship program for startup founders, offering technical guidance, early tool access, and on-site sessions in San Francisco. - CNBC

OpenAI detailed red-teaming partnerships with US CAISI and UK AISI that exposed vulnerabilities and biosecurity issues, prompting engineering, policy, and monitoring fixes. - OpenAI

OpenAI researchers propose confidence-thresholded answering to curb hallucinations, noting token-by-token generation and benchmarks that penalize uncertainty worsen reliability and costs, so the cure might make LLMs unusable - The Conversation

If it can accurately output a probability of hallucination, then you can tune the LLM for more or less hallucination, so you can probably improve it and users can choose based on the use case. But the fundamental problem revealed by the research is people preferred more certainty than is actually available. People demand bullshit economic point forecasts and get anxious when smarter people talk probabilities and decision science terms.

Anyway for academic research and retrieval you want to ground it and combine LLM, BM25, knowledge graphs and GraphRAG, with links to sources.

Apple ships simultaneous translation AirPods, as Google and Meta are also trying to perfect the universal translation Babelfish - CNBC

Former Siri leader departs Apple after failing to deliver stuff that was promised at WWDC 2024 - Apple Insider

Switching from iPhone to Pixel 10 feels like time travel into AI’s future - Android Authority

Meta is expected to unveil 'Hypernova' smart glasses next week, update Ray-Ban models, expand developer access, and showcase AI features and metaverse software. - Engadget

"AI isn’t magic; it’s a pyramid scheme of human labor." Google depends on thousands of contractors to train the AI behind its flagship chatbot Gemini. - Futurism

Edit images using Google's Nano Banana - Google

Elon Musk Appears to Be Firing 1/3 of the People Training Grok - Futurism

CBS found xAI’s Grok and other AI tools spread false claims and altered images after Charlie Kirk’s killing, spotlighting accuracy and moderation gaps. - CBS News

Winning the AI race is more important than a habitable planet - Interior Secretary - The Register

SK Hynix completed HBM4 development, preparing mass production with doubled I/O, higher efficiency, and 10 Gb/s speeds for next-generation GPU memory stacks. - The Register

A third of UK workers polled do not tell bosses about use of AI tools and half think AI threatens the social structure - the Guardian

One of the most common reasons that AI products fail? Data quality - Fortune

What Companies with Successful AI Pilots Do Differently - Harvard Business Review

AI won’t replace managers. But managers who ignore AI will replace themselves - Fast Company

Humanity's cyborg future - Literary Hub

A new study from the University of Pennsylvania reveals inconsistencies in how models like ChatGPT and Claude Sonnet treat slurs and inflammatory language. - Fast Company

How to Use Claude Code Subagents to Parallelize Development. - Zach Wills

Geohot: AI will replace coders the way spreadsheets replaced accountants. - the singularity is nearer

Fair point, and he makes me wonder what a programming language and toolchain truly optimized for AI would look like. 35 is a geezer when you are a prodigy I guess.

Anthropic outlined best practices for MCP tools, urging consolidated, well-specified utilities and programmatic evaluations that improved Claude’s software engineering benchmarks. - Anthropic

MCP-Agent: How to Build Scalable Deep Research Agents - The Alliance AI

Don't ask AI for the seahorse emoji - Futurism

Follow the latest AI headlines via SkynetAndChill.com on Bluesky