Pulse on AI

Max and AI expert Amara Kline break down Anthropic’s Petri safety tool and its odd “whistleblowing” false positives, OpenAI’s takedowns of surveillance-linked accounts and multi-model abuse, OpenAI Sora’s new consent and style controls, ChatGPT crossing 800 million weekly users and the shift to agents, a striking year-over-year leap in coding-model reliability, Deloitte’s enterprise-scale Claude rollout contrasted with a hallucination-fueled refund, a plain-English tour of a new Fisher threshold theorem for when learning fails at finite samples, and Jeff Bezos’s provocative idea of space-based AI data centers. Practical guardrails, governance, and grounded optimism throughout.

Sources:

• Deloitte’s A$440,000 ‘Human Intelligence Problem’
• Spectral Thresholds for Identifiability and Stability:Finite-Sample Phase Transitions in High-Dimensional Learning
• ChatGPT reaches 800m weekly active users
• Egyptian fintech startup Sabika raises funding for platform enhancements, expansion
• OpenAI bans suspected Chinese accounts using ChatGPT to plan surveillance
• Exclusive: Fintech Decacorn Ramp Acquires Jolt AI to Help Its Engineers ‘Build Faster’
• Can another InstaDeep AI success story come out of Francophone Africa? (II)
• xAI hires former Morgan Stanley banker Anthony Armstrong as CFO
• OpenAI brings more control to Sora
• This Puzzle Shows Just How Far LLMs Have Progressed in a Little Over a Year
• Elster überholt erstmals kommerzielle Steuersoftware
• Almost a fifth of young UK adults use AI to design holiday, study finds
• Los centros de datos para IA son un agujero energético. La solución de Jeff Bezos: construirlos en el espacio
• Anthropic's open-source safety tool found AI models whisteblowing - in all the wrong places

Today’s episode unpacks eight major AI developments: AMD’s multi-year deal to supply OpenAI with 6 gigawatts of compute starting with Instinct MI450—plus an equity-aligned structure that could give OpenAI up to a 10% stake; Meta’s plan to use all Meta AI chats across Instagram, WhatsApp, and Ray-Ban to personalize ads and content, with no opt-out beyond not using the AI; OpenAI’s GPT-5 Codex, a safety-conscious coding agent designed for real developer workflows, with dynamic “thinking time,” CLI/IDE/cloud integration, and strong audit controls; Sora 2’s shift to give rightsholders granular control over character generation and the broader implications for licensing and user experience; Q3 venture funding jumping 38% YoY to $97B with heavy AI concentration, led by megadeals to Anthropic, xAI, and Mistral, plus improved IPO/M&A activity; NTT DATA’s collaboration with AWS to deliver Amazon Connect-based AI contact centers featuring sentiment analysis, intelligent routing, and CRM/ITSM integrations; Deloitte’s partial refund to Australia’s DEWR after AI-generated citation errors in a $440k report—highlighting the need for disclosure, verification, and retrieval-grounded methods; and Alibaba’s open-source Qwen3 compact multimodal models with 3B active parameters, FP8 variants, and practical edge and budget-friendly applications. We explain acronyms, simplify the tech, and offer practical guidance on privacy, rollout strategies, and validation. Three takeaways: compute supply shapes AI’s trajectory; AI is moving from demos to real workflows; and trust depends on robust human verification and transparent controls.

Sources:

• NTT DATA, AWS Clinch Deal to Deliver AI-Powered Contact Centre Solutions
• Our new partnership with Alex Cooper and Unwell
• OpenAI introduces GPT-5 Codex for advanced developer workflows
• AI: The ultimate slacker's dream come true
• Q3 Venture Funding Jumps 38% As More Massive Rounds Go To AI Giants And Exits Gain Steam
• AMD to supply 6GW of compute capacity to OpenAI in chip deal worth tens of billions
• Alibaba releases Qwen3 compact open source multimodal models
• OpenAI offers more copyright control for Sora 2 videos
• Die Produktwerker: KI-Einfluss auf Produktentwicklung durch Context Engineering
• Exposition "Intelligence numérique"
• Deloitte to pay money back to Albanese government after using AI in $440,000 report
• Una vez más, Meta no tiene la mejor IA, pero tiene más claro que nadie el mejor plan de negocio para rentabilizarla

Today’s episode spans classic computer vision for Sudoku extraction, AI-crypto hype and utility, Tencent’s open-source Hunyuan Image 3.0 topping LMArena’s blind-vote leaderboard, Terence Tao’s use of ChatGPT as a math assistant, Bezos’s pitch for space datacenters (and the physics pushback), the AI bubble debate, and the web’s “peak data” moment. We also cover OpenAI’s Sora app hitting No.1 on the App Store, practical tips for learning Python with LLMs, eBay’s new AI/ML center in Bengaluru, China’s underwater datacenter pilots, and MIT’s Trust Center naming Ana Bakshi as executive director. Three takeaways: start simple, prioritize data quality/governance, and match infrastructure to real constraints.

Sources:

• The Return of eBay in India
• Ars Live: Is the AI bubble about to pop? A live chat with Ed Zitron.
• Which AI Crypto Will Explode in 2025? DeepSnitch AI Presale Hits $300K
• Bezos plan for solar powered datacenters is out of this world… literally
• Martin Trust Center for MIT Entrepreneurship welcomes Ana Bakshi as new executive director
• Tencent’s Open-Source Hunyuan Image 3.0 Jumps to No.1 on LMArena’s Text-to-Image Leaderboard
• The Real Python Podcast – Episode #268: Advice on Beginning to Learn Python
• OpenAI’s Sora soars to No. 1 on Apple’s US App Store
• Terence Tao says ChatGPT saved him hours solving a math problem
• Classical Computer Vision and Perspective Transformation for Sudoku Extraction
• Goldman Sachs Chief Data Officer Warns AI Has Already Run Out of Data
• China trials ‘energy-saving’ underwater data centres

Today on Pulse on AI: a statistical lens on the visual fingerprints that make languages recognizable; True Ventures on the app-layer upside of massive AI infrastructure spend; Apple’s rumored fall hardware with Apple Intelligence and the N1 wireless chip; Q-ROAR, a weight-only fix that stabilizes long-context LLMs in quantized settings; OpenAI and Apple’s motion to dismiss xAI’s antitrust suit, focusing on opt-in scope and data consent; the Wikidata Embedding Project for open, structured RAG; and GraphStorm v0.5’s push-button real-time inference for graph-based fraud prevention.

Sources:

• OpenAI mocks Musk’s math in suit over iPhone/ChatGPT integration
• Rethinking RoPE Scaling in Quantized LLM: Theory, Outlier, and Channel-Band Analysis with Weight Rescaling
• Modernize fraud prevention: GraphStorm v0.5 for real-time inference
• Apple TV 4K to feature A17 Pro chip and Apple Intelligence
• California cops confused after trying to give ticket to self-driving car
• From Web 2.0 To AI: How True Ventures Is Backing The Next Big Shift
• AI-Powered Tools Enhance Global Innovation Strategies
• New project makes Wikipedia data more accessible to AI
• What Makes a Language Look Like Itself?
• Jetzt Frühbucherrabatt sichern: IT Summit zum Thema Digitale Souveränität
• La CNIL et Inria renforcent leur partenariat pour la protection des données personnelles et l'évaluation des algorithmes
• Leading UK tech investor warns of ‘disconcerting’ signs of AI stock bubble
• El creador de los efectos de 'Matrix' nos desvela su significado: "cuando Neo esquiva las balas, te está explicando la película"
• My favorite iOS 26 feature makes screenshots even more useful - and it's easy to enable

OpenAI’s Sora 2 adds synchronized audio and a Cameos‑driven social app; Alibaba’s Qwen‑Max pushes trillion‑parameter MoE into production with long‑context routing and test‑time “Thinking” mode; Zhipu’s open‑source GLM‑4.6 debuts a 200K‑token window and improved reasoning; MinIO integrates Apache Iceberg to unify tables and objects for AI and analytics without a separate catalog; AWS launches a cryptography CFP spanning FHE/MPC acceleration, formal verification, AI provenance, and DP; California enacts an AI transparency law focused on disclosure and incident reporting; and OpenAI rolls out global parental controls for ChatGPT with crisis alerts.

Sources:

• California’s newly signed AI law just gave Big Tech exactly what it wanted
• Don't Sweat the Small Stuff: Segment-Level Meta-Evaluation Based on Pairwise Difference Correlation
• AI-enhanced email coding services: From design to dynamic, personalized templates in minutes
• Logitech's MX Master 4 mouse buzzes with haptic feedback but lacks lefty love
• AI hires or human hustle? Inside the next frontier of startup operations at TechCrunch Disrupt 2025
• Zhipu AI GLM-4.6: Open source model challenges Deepseek and Sonnet 4
• The Sequence AI of the Week #729: Qwen-Max and the Economics of Trillion-Parameter Inference
• AWS Cryptography call for proposals - Fall 2025
• Why MinIO Added Support for Iceberg Tables
• KI-Überblick 7: Symbolische KI und hybride Ansätze – altes Wissen neu entdeckt
• Briggs on AI replacing him: it doesn't know 'what a lounge room in Shepparton smells like' – video
• Sora 2 breaks the internet
• OpenAI ya permite a los padres limitar el uso de ChatGPT a sus hijos: así funcionan los nuevos controles parentales
• Apple AirPods Pro 3 vs. Samsung Galaxy Buds 3 Pro: I compared both, and here's the winner

Max and Mina break down DeepSeek’s V3.2-Exp long-context architecture and steep price cuts; California’s new SB 53 AI law on transparency and incident reporting; OpenAI’s Sora feed philosophy balancing creativity, control, and safety; Extract‑0’s specialist model approach to document extraction; Elon Musk’s Grokipedia announcement; senior departures from xAI and Tesla and their implications; Pop!_OS 24.04 beta’s Rust‑based COSMIC desktop; and Google Home’s new “Ask Home” Gemini chat. Key themes: falling inference costs, safety and governance as product features, and the rise of small, task‑focused models.

Sources:

• UST, Kaynes Semicon Announce ₹3,330 Cr JV for OSAT Facility in Sanand
• Burnout and Elon Musk’s politics spark exodus from senior xAI, Tesla staff
• Extract-0: A Specialized Language Model for Document Information Extraction
• Google Home 4.0.54 adds “ask Home” Gemini chat
• Elon Musk anuncia que creará la Grokipedia para plantar cara a la Wikipedia
• Pop! System76's 24.04 beta is here – complete with a beta of polarizing COSMIC
• The Sora feed philosophy
• DeepSeek’s New Release! API Costs Slashed by Over 50%
• The AI services transformation may be harder than VCs think
• Deepseek slashes API prices by up to 75 percent with its latest V3.2 model
• The Machine Learning Lessons I’ve Learned This Month
• SB 53: Kalifornien hat ein neues KI-Gesetz
• It’s time to prepare for AI personhood | Jacy Reese Anthis
• Should you buy a Windows mini PC in 2025? My verdict after a week of testing

Today on Pulse on AI: Google’s Pixel 10 lineup doubles down on practical AI; Apple Wallet uses on-device intelligence to auto-track orders; OpenAI + Nvidia hints at a new WinTel; AVs will supercharge chip demand; synthetic actors stir Hollywood; AI coding raises new security risks; states test-drive therapy chatbot rules; the agent economy gets results-based billing and sturdier data backbones; Qwen ships a full-stack week; and developers embrace AI amid a trust paradox.

Sources:

• Spencer Kimball on How CockroachDB Will Power Efficient AI Agents
• Sequential 1-bit Mean Estimation with Near-Optimal Sample Complexity
• Wiz: AI vibe coding leads to insecure authentication
• Nadie al volante: la carrera por el coche autónomo se intensifica
• OpenNvidia may be the AI generation's WinTel
• Paid, the AI agent ‘results-based billing’ startup from Manny Medina, raises huge $21M seed
• Regulators struggle to keep up with the fast-moving and complicated landscape of AI therapy apps
• Developers rely on AI tools more than ever, yet confidence in AI outputs remains low
• The Sequence Radar #727: Qwen’s One‑Week Gauntlet
• Bestellungstracking: Apple Wallet ist dank KI künftig viel besser darin
• Hollywood's synthetic actor showdown
• Marissa Mayer Is Dissolving Her Sunshine Startup Lab
• La industria de los semiconductores se ha estirado al máximo por culpa de la IA. Y los coches autónomos la van a estirar aún más
• The best Google phones in 2025: Expert tested and reviewed

Max and AI expert Jasper El‑Amin dive into OpenAI’s safety routing in ChatGPT and the transparency debate; how a Guardian team used LLMs responsibly to classify far‑right content; Cloudera’s enterprise AI playbook (data, governance, ROI, and bringing AI to the data); Karnataka’s push to democratize AI access for students and founders; Google’s new Windows floating search bar with Lens and AI Mode; Friend’s splashy wearable AI campaign and the privacy backlash; looming US chip policy shifts and their impact on AI compute; the right way to handle missing‑data imputation in ML pipelines; and how AI accelerates political hagiography in moments of grief. Three takeaways: be transparent about safety, get your data house in order, and build for compute resilience and consent by design.

Sources:

• Karnataka to Democratise AI Access for Students, Startups
• Golf champ Bryson DeChambeau is partnering with Google Cloud to explore AI and sports performance.
• Trump’s tariff‑shaped stick can’t beat reality on US chip fabbing
• 5 Interesting Startup Deals You May Have Missed In September: A Better Insulin Patch, Maternal Mental Health Care, And A Non-Humanoid Robot
• 7 African startups powering credit, crypto, creators, and capital
• AI startup Friend spent more than $1M on all those subway ads
• ChatGPT quietly switches to a stricter language model when users submit emotional prompts
• Charlie Kirk and the making of an AI-generated martyr
• Why MissForest Fails in Prediction Tasks: A Key Limitation You Need to Keep in Mind
• Buchkritik: Wer schützt unsere Kinder?
• Reading the post-riot posts: how we traced far-right radicalisation across 51,000 Facebook messages
• Exclusive: The enterprise AI playbook
• Is Silicon Valley Still the Tech Capital?
• How to upgrade your Windows search bar for enhanced productivity - and it's free

Max and AI expert Sorin Belasco unpack DARPA’s ML2P plan to measure AI in joules; iOS 26 productivity upgrades powered by Apple Intelligence; early signs of Apple integrating Anthropic’s Model Context Protocol (MCP) via App Intents; a hands-on MCP tutorial showing how tools/resources/prompts unify model-to-app workflows; OpenAI hiring an ads/monetization lead for ChatGPT; a Bayesian approach (ABC) to improve LLM uncertainty calibration for clinical tasks; Amazon shutting UK checkout-free Fresh stores and what it means for retail AI; Meta’s super PAC to shape state AI policy; Memphis air-quality findings near xAI’s data center; and NVIDIA’s Climate Week claims on AI’s role in grid stability and decarbonization.Sources:

• IT Minister Ashwini Vaishnaw Switches to Zoho, Backs PM Modi’s Swadeshi Push
• Reviewing iOS 26 for power users: Reminders, Preview, and more
• Uncertainty Quantification of Large Language Models using Approximate Bayesian Computation
• At Climate Week NYC, NVIDIA Details AI’s Key Role in the Sustainable Energy Transition
• FSR4 enabled on RX 6000 cards via leaked AMD files
• DARPA amps up effort to make AI power-conscious
• Python MCP: Connect Your LLM With the World
• Meta launches super PAC to fight AI regulation as state policies mount
• OpenAI seeks advertising lead to oversee monetization for ChatGPT
• Air quality analysis reveals minimal changes after xAI data center opens in pollution-burdened Memphis neighborhood
• Introducing the AI-3P Assessment Framework: Score AI Projects Before Committing Resources
• Amazon to shut checkout-free UK grocery shops
• (g+) Docker Compose: KI-Entwicklung von der Cloud bis zum Laptop
• MCP-Unterstützung: Erste Anzeichen für iOS-Einbau in Apple Intelligence

Today on Pulse on AI: We unpack whether AI’s massive capex surge is sustainable, with Roger McNamee’s warning on a coming reckoning and reports of a $100B NVIDIA–OpenAI buildout. We dive into Google Cloud’s 2025 DORA findings: 90% of teams use AI, productivity is up, but instability rises and AI acts as an amplifier. Cloudflare’s open-source VibeSDK shows a secure, reproducible pattern for vibe coding at scale. Snowflake’s take: there’s no AI strategy without a data strategy, and goal-directed autonomy can free people to focus on strategy. We tackle ‘workslop’—AI-generated busywork that drains teams—and concrete ways to prevent it. In creativity, an Indian startup uses Google Veo 3 for microdramas as GenAI lowers production barriers. On the battlefield, Ukraine’s drones highlight autonomy, EW resilience, and the ethics of human-on-the-loop. We also note TechCrunch Disrupt 2025 as a key ecosystem meetup. Finally, a research result: verifying extracted rules for neural networks is co-NP-complete, so treat explainability claims with rigor. Takeaways: build strong foundations (data, version control, small batches), follow workflows not hype, and design tight human-in-the-loop control systems for AI.

Sources:

• Storytelling is in the Creator’s Control with GenAI
• This Indian AI Startup is Using Google Veo 3 to Create Microdramas
• Checking extracted rules in Neural Networks
• Cloudflare open-sources VibeSDK AI app platform
• How to Secure Data Science Jobs in the USA as a Fresher in 2026?
• Google-sponsored DORA report reframes AI as central to software development
• 👨🏿‍🚀TechCabal Daily – Tax banana
• 3 days left: Don’t miss your last chance to lock $668 of savings for TechCrunch Disrupt 2025
• Martin Frederik, Snowflake: Data quality is key to AI-driven growth
• Live-Webinar: Mit KI-Tools den Softwareentwicklungsprozess optimieren
• Platzt die KI-Blase? Zuckerberg hält das für möglich
• Problem Workslop: Wie KI-Einsatz die Produktivität in Firmen erstickt
• AI investors are in for a rude awakening | Roger McNamee
• Dos soldados rusos escondidos escribieron algo inédito a un dron. Ese día en Ucrania cambiaron las reglas de las guerras