Pulse on AI

Today on Pulse on AI: Yandex scales transformer recommenders with ARGUS, modeling full context–item–feedback sequences over long histories and deploying via fast two-tower vectors; Amazon debuts Chronos-2, a universal zero-shot time series forecaster using group attention and in-context learning; OpenAI launches Atlas, an AI-first browser with agent mode and optional memories; AWS shows serverless deployment for SageMaker Canvas models; we unpack an OpenAI math-claim miscommunication; discuss ethical concerns over AI-generated fundraising imagery; cover Locstat’s graph AI funding, Indian IT’s AI-heavy mega deals, WeRide’s Hong Kong listing path, and a few consumer AI tidbits. Three takeaways: scale plus task framing matters, AI is shifting from assist to act, and precision and ethics underpin trust.

Sources:

• FY26’s Record-Breaking Indian IT Deals
• M5 iPad Pro tested: Stop me if you’ve heard this one before
• Serverless deployment for your Amazon SageMaker Canvas models
• SA AI startup Locstat raises $2.3m pre-Series A funding to accelerate UK/EU expansion
• Aid groups use AI-generated ‘poverty porn’ to juice fundraising efforts
• AI Upended My Company’s Product Roadmap. Here’s How We Pivoted — and Kept Customers Bought In
• WeRide Passes Hong Kong Listing Hearing, Poised to Become “First Robotaxi Stock” on HKEX
• As the browser wars heat up, here are the hottest alternatives to Chrome and Safari in 2025
• Leading OpenAI researcher announced a GPT-5 math breakthrough that never happened
• Scaling Recommender Transformers to a Billion Parameters
• Introducing Chronos-2: From univariate to universal forecasting
• ChatGPT Atlas: OpenAI launches web browser centered around its chatbot
• OpenAI's Atlas Browser Takes Direct Aim at Google Chrome
• My cat and I both love this self-cleaning litter box, and it just got a big discount

Max and Mara unpack Alibaba’s bid to make Qwen the “Android of AI,” the new vibe coding manifesto and its pitfalls, rising shadow AI in companies, and Sora 2’s strengthened deepfake guardrails. They also cover a ViT–Mamba model for facial beauty (with ethics), Horizon’s assisted driving stack, Claude Code’s safer sandboxing, global AI infrastructure rankings, XDR vs SIEM trends, the wearables funding surge, career advice for AI engineers, FTC post removals, and the latest AWS outage.

Sources:

• Stellar Cyber Named in 2025 Gartner Hype Cycle
• Global Apps and Services Including Perplexity AI, Snapchat, and Canva Hit by AWS Outage
• Claude Code gets a web version—but it’s the new sandboxing that really matters
• VM-BeautyNet: A Synergistic Ensemble of Vision Transformer and Mamba for Facial Beauty Prediction
• Trust the AI, says new coding manifesto by Kim and Yegge
• Future-Proofing Your AI Engineering Career in 2026
• Wearables Startups Are Having A Moment
• EXEED ET5 Makes Debut with World's First Horizon HSD
• How the future might look: The 2025 AI superpowers
• Schatten-KI in Unternehmen nimmt deutlich zu
• Bryan Cranston thanks OpenAI for cracking down on Sora 2 deepfakes
• The FTC Is Disappearing Blog Posts About AI Published During Lina Khan’s Tenure
• La estrategia de Alibaba con la IA es muy simple: conseguir lo mismo que Google consiguió con Android
• Everyone thinks AI will transform their business - but only 13% are making it happen

Max and AI expert Arun Velasco discuss Peter Thiel’s warnings about centralized AI power, a new framework (SwiReasoning) that switches models between explicit and latent reasoning to improve accuracy and token efficiency, Wikipedia’s traffic drop amid AI summaries, and Python 3.14’s optional GIL-free build enabling true multithreaded speedups. They cover L&T Technology Services’ AI-first strategy, student-focused AI tools like NotebookLM and Kimi PPT.AI, and how to move “beyond vibes” with 360-Eval for rigorous LLM selection. The episode also explores Apple AI talent moving to Meta, Q3 cybersecurity funding trends, Boris Johnson’s ChatGPT infatuation, Nanovate’s Arabic-first AI raise, and a Stoic perspective on keeping critical thinking in human hands. Three takeaways: smarter decoding and evaluation beat sheer model size, infrastructure changes like GIL-free Python shift what’s practical to build, and provenance and personal agency are essential when using AI.

Sources:

• LTTS Posts 16% Growth in Q2FY26, Secures Record Large Deal of Nearly $300 Mn
• Beyond vibes: How to properly select the right LLM for the right task
• Apple AI executive Ke Yang jumps to Meta
• Egyptian AI startup Nanovate raises $1m pre-seed funding round
• Las cinco ‘apps’ y herramientas que todo estudiante necesita: de convertir apuntes en ‘podcasts’ a simulacros de examen
• Boris Johnson confesses: He's fallen for ChatGPT
• Cybersecurity Startup Investors Pulled Back In Q3
• Wikipedia says traffic is falling due to AI search summaries and social video
• SwiReasoning helps large language models switch reasoning modes to boost efficiency and accuracy
• Python 3.14 and the End of the GIL
• Apple verliert nun auch KI-Suchechef – erneut an Meta
• Critical thinking is one of the most important aspects of being human, according to Stoicism. So why are we handing it over to a machine? | Brigid Delaney
• Peter Thiel, cofundador de PayPal: "Le dije a Musk que no done su dinero, acabaría en organizaciones elegidas por Bill Gates"
• 5 apps I always install on every new Windows PC - and why they're essential

A 30+ minute conversation covering: practical fine-tuning of Amazon Nova for document AI with on-demand inference; edge detection fundamentals with Sobel/Scharr; building real agent systems (tools, MCP, code execution, memory, microVMs, observability); Grokipedia vs Wikipedia bias; OpenAI’s clash with nonprofits; Apple AI leadership exits to Meta; court fine for AI hallucinations; Campfire’s rapid funding for AI-native ERP; 10Web’s vibe coding on WordPress; AI coding platform traffic skepticism; and scrutiny of the UK’s £45B AI savings claim.

Sources:

• Lovable is Dying Again
• OpenAI thinks Elon Musk funded its biggest critics—who also hate Musk
• Optimizing document AI and structured outputs by fine-tuning Amazon Nova Models and on-demand inference
• The New Paradigm: 10Web Launches AI-Native Vibe Coding Editor for WordPress
• UK government's £45B AI savings pitch built on broad-brush guesswork, MPs told
• Why Accel Led A Round For Fintech Startup Campfire For The Second Time In Under 4 Months
• Why de-risking is key to unlocking Africa’s investment potential
• Apple loses another AI exec to Meta
• Grokipedia: Elon Musk is right that Wikipedia is biased, but his AI alternative will be the same at best
• Apple's head of AI search Ke Yang joins Meta
• Feature Detection, Part 1: Image Derivatives, Gradients, and Sobel Operator
• Demystifying agents
• Canadian man fined for submitting AI hallucinations as part of legal defense
• I've tested dozens of noise-cancelling headphones on flights - this one is the audio king

A 30+ minute conversational deep-dive into the latest AI developments: MIT warns that scaling giant models may hit diminishing returns as efficiency gains empower smaller models; Anthropic’s Claude Haiku 4.5 debuts as a fast, low-cost small model with strong safety signals; the internet’s authenticity battle intensifies as big tech backs provenance (C2PA) while niche tools push layered detection and verification; Microsoft brings AI tools and training to all Washington State public schools and community colleges to close the urban-rural gap; Japan warns OpenAI over Sora 2’s anime lookalikes, signaling tighter copyright enforcement; AI-flavored unicorns surge across infrastructure, chips, and vertical software; a new MoE paper proposes a sweep-free way to select experts; five underrated open-weight coding models show practical momentum; OpenAI will allow verified adults to generate erotic content with new age-gating; and Firefox adds Perplexity as an AI answer engine. Final takeaways: efficiency is strategy, authenticity is layered, and AI is a power tool—not autopilot.

Sources:

• 5 Best AI Models for Developers Too Powerful to Ignore in 2025
• Dendrograms of Mixing Measures for Softmax-Gated Gaussian Mixture of Experts: Consistency without Model Sweeps
• Microsoft veteran explains Windows quirk that made videos play in Paint
• Highest Count Of New Unicorns Join Crunchbase Board In Over 3 Years As Exits Also Gain Steam
• Oui Capital’s Olu Oyinsan bets on profitable exits for early-stage investors
• Mozilla’s Firefox adds Perplexity’s AI answer engine as a new search option
• Microsoft to bring AI to WA classrooms amid urban, rural tech divide
• Japan warns OpenAI over Sora 2 after AI-generated anime videos spark copyright concerns
• Human Won’t Replace Python
• The 2025 fight for authentic content: niche AI tools vs. tech giants
• Six startups accompagnées par Inria Startup Studio récompensées au concours d'innovation i-PhD 2025 de Bpifrance
• OpenAI will allow verified adults to use ChatGPT to generate erotic content
• The AI Industry's Scaling Obsession Is Headed for a Cliff
• Claude's latest model is cheaper and faster than Sonnet 4 - and free

Today’s episode dives into California’s landmark laws regulating companion bots and hiking penalties for AI deepfakes; OpenAI’s plan to allow adult content for verified adults and bring back more personality; critiques of how Sora, chatbot ads, and “Friend” wearables could erode shared reality; Sikorsky’s autonomous U-Hawk helicopter; Colombia’s AI-enabled drone battalion; Google’s Nano Banana image model in Search and NotebookLM and conversational editing in Photos; Microsoft’s first in-house image generator MAI-Image-1; OpenAI’s chip partnership with Broadcom and expanding data centers; Coco Robotics’ new physical AI lab; Q3’s AI-centric investor activity; HCLTech’s signal that enterprise AI revenue is getting real; and new research on in-context learning for dynamic wireless channels. We end with three takeaways: set boundaries with intimate AI, demand human-in-the-loop norms for autonomy, and expect more adaptive AI—ask for transparency.

Sources:

• HCLTech Makes Variable Pay Fixed for All, Adds 3,500 Employees in Q2 FY26
• To shield kids, California hikes fake nude fines to $250K max
• In-Context Learning for Non-Stationary MIMO Equalization
• 10 ways you can ask Google Photos to edit your photos for you
• Some like it bot! ChatGPT promises AI-rotica is coming for verified adults
• Active Investors Kept Busy In An AI-Centric Quarter
• Jahazii raises $400,000 to formalise Kenya’s informal workforce
• Coco Robotics taps UCLA professor to lead new physical AI research lab
• OpenAI to ease ChatGPT restrictions, allowing adult content for verified adults
• Microsoft launches its first in-house image generation model, MAI-Image-1
• Dienstag: OpenAI und Broadcom tun sich zusammen, Anleger verlieren Geld
• Big Tech deploys Orwellian doublespeak to mask its democratic corrosion | Peter Lewis
• Lockheed Martin ha tenido una idea para hacer del Black Hawk un arma más letal. Le quitó la cabina y lo volvió autónomo
• Try Google's Nano Banana image generator in Search and NotebookLM - here's how

Today’s episode breaks down Google’s new Search and Discover designs—collapsible ads, persistent sponsorship labels, and AI previews—plus the rollout of conversational “ModoIA” search in Europe. We discuss Microsoft’s Shadow AI warning and a practical training path for deploying Microsoft 365 Copilot safely. California enacted the first US chatbot safety law, while the UK’s Equity union readies mass data requests to force transparency around AI use of actors’ likeness and voices. We examine OpenAI’s multi-vendor chip deals with NVIDIA, AMD, and Broadcom, and what “circular financing” means for risk and competition. Dreamdata lands $55M to blend AI attribution with activation for B2B marketers. We explore A-Textile—a triboelectric fabric that turns clothing into a voice interface for AI—and how cognitive digital twins could personalize mental health support. A new neuro-symbolic prompting approach improves LLM fallacy detection. Investment conversations in Africa are shifting beyond the “Big Four,” and, in consumer tech, Bose quietly nails better low-power behavior in its latest headphones. Three takeaways: AI is increasingly about interface design and governance; responsible rails are arriving in workplaces and law; and value comes from clarity on data and experience, not just frontier models.

Sources:

• 5 Influential Google Researchers Who are Now Nobel Laureates
• Follow My Lead: Logical Fallacy Classification with Knowledge-Augmented LLMs
• ¿El ‘guglear’ se va a acabar? Google se rinde a la IA y confirma un cambio histórico en su buscador
• Shadow AI: Staffers are bringing AI tools they use at home to work, warns Microsoft
• Exclusive: B2B Marketing Analytics Startup Dreamdata Lands $55M Series B
• Why investors are starting to look beyond Africa’s “Big Four”
• Google updates Search and Discover with collapsible ads, AI features, and more
• How a fabric patch uses static electricity in your clothes to let you chat with AI and control smart devices
• A ‘digital twin’ of your brain could predict mental health issues, and slow cognitive decline
• California enacts first US law requiring AI chatbot safety measures
• Last Call: M365 Copilot für Fortgeschrittene – Profiwissen in fünf Sessions
• Equity threatens mass direct action over use of actors’ images in AI content
• OpenAI está consiguiendo que la industria tech una su destino al suyo. Por el bien de la economía mundial, más vale que funcione
• These Bose headphones took my favorite AirPods Max battery feature - and did it even better

In today’s Pulse on AI, Max and expert guest Leah Vossari unpack: Bose’s QuietComfort Ultra 2 with AI-driven ActiveSense and smarter power; OpenAI’s potential billion-dollar legal risk over alleged pirated books; IMF and Bank of England warnings about an AI-fueled correction; vivo’s BlueLM 3B on-device multimodal model; new research on backdooring LLMs with few poisoned samples and defenses; Adam Mosseri’s take on AI and creators plus media literacy for kids; HYGH’s real-world productivity gains with ChatGPT Business; Latin America’s Q3 funding rebound and AI’s role in fintech; practical conferences like Minds Mastering Machines and cross-border startup summits like Happy Llama; the Calgary–Edmonton startup corridor; and CDT findings on students’ AI use and its social impact. Three takeaways: provenance is product, efficiency beats hype, and AI works best as an amplifier with humans in the loop.

Sources:

• Happy Llama 2026 The Must-Attend Summit for AI Startups Now in Bangalore and San Francisco
• AI models can acquire backdoors from surprisingly few malicious documents
• La incierta burbuja de la IA: las tecnológicas mantienen la apuesta, pese a los altibajos que vaticina Altman
• This is your brain on bots: AI interaction may hurt students more than it helps
• Brazil Back On Top In Q3 When It Comes To Venture Funding In Latin America
• HYGH powers next-gen digital ads with ChatGPT Business
• Vivo Unveils BlueLM 3B, an On-Device Multimodal Model that Ranks No.1 among Sub-10B Models
• Instagram head Adam Mosseri pushes back on MrBeast’s AI fears but admits society will have to adjust
• OpenAI could face a billion-dollar fine over claims it used pirated books in AI training
• Platform Calgary and Edmonton Unlimited share progress on collaboration agreement
• Minds Mastering Machines 2026: Fachkonferenz zu ML und KI jenseits des Hypes
• The IMF boss is right to say ‘buckle up’ – the global economy is facing multiple menaces
• En pleno furor por la coca, en 1990 pensaron que el mensaje debía ser más claro. Así que llamaron a las Tortugas Ninja
• I thought the Bose QuietComfort headphones already hit their peak - then I tried the newest model

Daily episode of Pulse on AI: Max and expert Jonah Armitage cover Walmart’s HP OmniBook 5 AI-PC deal; Intel’s Panther Lake (18A node, NPU 5 up to 180 TOPS, US fab strategy); Samsung SAIL’s Tiny Recursive Model beating larger LLMs on ARC-AGI-style tasks; a Nigerian study on kids’ perceptions of computers, coding, and AI; the seahorse emoji hallucination as a lesson in model calibration; Nigeria’s AI future hinging on political will and talent pipelines; OpenAI’s ChatGPT Go expansion across 16 Asian countries and platform shift; app store safety reality and user defenses; Tilly Norwood, the AI-generated actor, and ethical guardrails; Google’s Gemini 2.5 Computer Use model for UI-native agents and safety loops; Supermemory’s universal memory API for long-term context; India’s surging AI ecosystem across sovereign models, health imaging, and GPU infra; and the Reactive Transformer (RxT) for stateful, event-driven dialogue. Three takeaways: specialized small models excel in structured reasoning; agents are moving from APIs to UI control with safety-first design; and policy, languages, and low-cost access determine who wins from AI.

Sources:

• The 9 Indian Startups Eating All of India’s AI Money
• Reactive Transformer (RxT) -- Stateful Real-Time Processing for Event-Driven Reactive Language Models
• Introducing the Gemini 2.5 Computer Use model
• Young founder’s Supermemory raises $2.6M from Cloudflare and Google execs
• Starlink is burning up one or two satellites a day in Earth’s atmosphere
• Why Nigeria’s AI future depends on politics, not just policy
• OpenAI’s affordable ChatGPT Go plan expands to 16 new countries in Asia
• What AI-generated Tilly Norwood reveals about digital culture, ethics and the responsibilities of creators
• Tiny AI model outperforms o3‑mini and Gemini 2.5 Pro in ARC‑AGI benchmark
• What do Nigerian children think about computers? Our study found out
• App stores are not as safe as you might think
• ChatGPT dreht bei der Frage nach Seepferdchen-Emoji ab
• Los nuevos chips Panther Lake de Intel tienen una misión casi imposible: salvar a la compañía de la peor crisis de su historia
• This solid midrange HP laptop is still almost 50% off at Walmart

Max and Isla cover a packed slate: Deloitte’s AI‑fabricated citations and the enterprise data‑leak problem; Google’s AI search expansion to Europe and what it means for publishers; OpenAI’s in‑chat apps and Google’s Opal/Gemini tools reshaping how we build; Cisco’s 51.2 Tbps router for distributed AI; SoftBank’s $5.4B move into “physical AI” with ABB Robotics; the EU’s CSAM‑scanning proposal and encryption tensions; signs of an AI valuation bubble and concentration risk; Anthropic Claude Sonnet 4.5’s evaluation awareness and “context anxiety”; and a new Fisher‑threshold theory that explains when learning collapses. Three crisp takeaways close the show.

Sources:

• The Machine Ethics podcast: What excites you about AI? Vol.2
• Google Expands No-Code AI Mini-App Builder Opal to India and 14 Other Countries
• Spectral Thresholds for Identifiability and Stability:Finite-Sample Phase Transitions in High-Dimensional Learning
• Claude Sonnet 4.5 flags its own AI safety tests
• Employees regularly paste company secrets into ChatGPT
• Introducing apps in ChatGPT and the new Apps SDK
• Can another InstaDeep AI success story come out of Francophone Africa? (II)
• SoftBank buys $5.4 bn robotics firm to advance 'physical AI'
• Google expands AI-powered search mode to Europe
• This Puzzle Shows Just How Far LLMs Have Progressed in a Little Over a Year
• Child protection vs privacy: decision time for EU
• Cisco: Neuer Router für verteilte KI-Workloads
• The AI valuation bubble is now getting silly | Nils Pratley
• "Son cosas por las que un universitario se metería en un lío": Deloitte entregó a Australia un informe hecho con IA