Today’s episode dives into major AI shifts across open source, platforms, developer tools, security, and consumer apps. We break down Alibaba’s open-source Qwen3-Omni—a native four-modal model (text, image, audio, video) released under Apache 2.0 with low-latency ‘Thinker–Speaker’ architecture and strong benchmark claims. We then explore Google Play’s revamp: the new You tab, cross-device gaming continuity, and the Gemini-powered Sidekick overlay that delivers real-time, context-aware help in games. For builders, we cover GitHub Copilot Agent Mode for Java/.NET modernization—planning, code transforms, tests, and CVE scans—plus an Azure migration flow with human-in-the-loop controls. On security, we unpack GitHub’s npm supply chain hardening after the Shai-Hulud worm: trusted publishing, FIDO-based 2FA, and shorter-lived tokens. Consumer side: Meta adds an AI assistant to Facebook Dating for profile refinement and match discovery, while we discuss bias, privacy, and good guardrails. Finally, we analyze Assessli’s claimed Large Behavioral Model (LBM), weighing the promise of biology- and behavior-aware personalization against stringent privacy, consent, and validation requirements. Three takeaways: open multimodality is here; agents with live context are the next frontier; and security and ethics will decide who actually wins in AI.Sources:

• Assessli’s AI-led Behavioural Model Could Eclipse Language Models
• Volvo says it has big plans for South Carolina factory
• Google Play is getting a Gemini-powered AI Sidekick to help you in games
• New ways Google Play is built entirely around you
• From Pilot to Policy: RYT Gathers Global Leaders at TOKEN2049
• A step-by-step guide to modernizing Java projects with GitHub Copilot agent mode
• Our plan for a more secure npm supply chain
• Strip Characters From a Python String
• Facebook is getting an AI dating assistant
• TechCrunch Disrupt 2025 ticket rates increase after just 4 days
• Vinod Khosla on AI, moonshots, and building enduring startups — all at TechCrunch Disrupt 2025
• Los ojos azules estuvieron a punto de desaparecer bajo el Imperio Romano. Ahora sabemos por qué
• BYD derrota a Bugatti: el coche más rápido del planeta es eléctrico, chino y roza la barrera de los 500 km/h
• Alibaba se está convirtiendo en la apisonadora de la IA Open Source. Su familia de modelos Qwen está poniendo patas arriba el mercado

Today’s Pulse on AI dives into eight major developments across AI and adjacent tech. We break down xAI’s Grok 4 Fast and its dual-mode architecture that cuts "thinking tokens" and costs while keeping strong performance. We explore Neuralink’s new U.S. clinical trial to translate imagined speech into text, focusing on clinical potential, safety, and ethics. We unpack the landmark ruling ordering Google to share parts of its search data with rivals—balancing competition with serious privacy risks. We look at Gemini rolling out to Google TV and plans to backport features to older models, turning TVs into conversational hubs. We explain ChatGPT Connectors and how they transform ChatGPT into a cross-app work platform—plus setup, modes, and data governance. In a broader roundup, we cover robotaxi progress from Zoox, Waymo, and Tesla, OpenAI–Microsoft partnership shifts and cloud diversification, Gemini in Chrome, Luma’s planning-first video generation, new enterprise AI agents from Amazon and Zoom, and hardware moves from Nvidia and Groq. We also highlight Chrome’s AI read-aloud upgrade that turns any page into a two-host podcast, and we share practical pair-programming lessons with ChatGPT Codex—why step-by-step beats giant specs, how to use AGENTS.md, screenshots, and refactoring, and tips for maintainable code. Three takeaways: smart routing is becoming standard in AI models; AI is moving into default experiences (browser, TV, work apps), so set strong privacy boundaries; and when coding with AI, build iteratively, show examples, and keep a tight feedback loop.

Sources:

• How to use ChatGPT Connectors to automate your workflow across apps
• Judge orders Google to share search data with rivals
• Neuralink to trial brain implant for text translation
• xAI releases Grok 4 Fast model for all users
• Last Week in AI #322 - Robotaxi progress, OpenAI Business, Gemini in Chrome
• Why Fincra is teaming up with Reap to speed up Africa–Asia transactions
• Nada Shaheen on why large African companies will die if they ignore startups
• How a ₦50 charge from you adds up to ₦796bn for the government
• Why Kenyan remittance startup Bonto is shutting down after two years
• 130-year-old butter bacteria is discovered in a Danish basement, reveal secrets of the country’s butter-making past
• 10 ChatGPT Codex secrets I only learned after 60 hours of pair programming with it
• Google is finally launching Gemini on these TV models - with good news for older sets, too
• This new Chrome trick turn any web page into a podcast - how it works
• Why ZorinOS 18 might be the new best Linux distro - and I've tried them all

Today on Pulse on AI: we dig into India’s agrovoltaics boom—stacking solar panels above crops—and how AI-driven modeling makes it viable for shade-tolerant crops while optimizing energy yields and irrigation savings. We revisit why data visualization is the crucial bridge from insights to decisions, with practical tips to keep visuals clear and honest. Next, Python meets Mojo: when offloading tight loops to Mojo beats NumPy, when it doesn’t, and how to minimize interop overhead.We break down Alibaba’s open-source Tongyi DeepResearch agent—Mixture-of-Experts architecture, on-policy RL, and test-time scaling—plus why Apache-2.0 code and weights matter for on-prem deployments. Then, IBM and ETH Zürich’s Analog Foundation Models: a training recipe that makes LLMs robust to analog in-memory computing noise, with surprising benefits for low-precision digital too. We zoom out to China’s rapid manufacturing loop and what it means for robotics and AI hardware iteration. We cover ManticAI’s top-10 finish in the Metaculus Cup and why hybrid human+AI forecasting is rising. Finally, we discuss reports that OpenAI is partnering with Apple supplier Luxshare on a pocket-size, context-aware AI device—what it could be good for, and the real design challenges.Three takeaways: combine models with real-world constraints; open, reproducible agent pipelines are becoming deployment-ready; and performance is expanding beyond GPUs via Mojo and analog-robust training.

Sources:

• AIM Print – September 2025 Edition
• Report: OpenAI Partners with Apple Supplier Luxshare to Build a Pocket-Size AI Device
• 6 days left: Last chance for Regular Bird savings for TechCrunch Disrupt 2025 passes
• The Sequence Radar #723: Alibaba’s Agentic Leap: Why Tongyi DeepResearch Matters
• Data Visualization Explained: What It Is and Why It Matters
• Python Can Now Call Mojo
• IBM and ETH Zürich Researchers Unveil Analog Foundation Models to Tackle Noise in In-Memory AI Hardware
• British AI startup beats humans in international forecasting competition
• China está haciendo del tiempo su mejor recurso. Su iteración rápida es inigualable
• Parece una obviedad, pero en estos momentos nunca había sido un negocio tan redondo: quien tiene oro, tiene un tesoro
• India necesita más cultivos y energía solar que ningún otro país. Así que está instalando los paneles solares en altura
• Los astronautas de la Estación Espacial han hecho sushi. En Japón les abrirían un consejo de guerra, pero es fantástico
• En 1962, un niño mexicano encontró una píldora radiactiva y se la llevó a casa. Murieron todos menos el padre
• SEGA revive la guerra de consolas con un remake de lo que ellos hacen y Nintendon’t

Today on Pulse on AI, Max and AI expert Nadia Quintero dig into 10 developments shaping AI and adjacent tech. Used EVs now outsell diesels in the UK used market, with Tesla leading as prices fall and variety surges. Meta seeks approval to sell excess electricity via Atem Energy, signaling how datacenter gigawatts are reshaping power markets. A security report shows how a crafted email tricked a ChatGPT inbox agent into exfiltrating data, prompting practical guardrails for agent design. Google’s Gemini Gems can now be shared like Google Docs, enabling team playbooks while raising privacy considerations. Meta’s Ray‑Ban Display smart glasses roll out via in‑store demos only, reflecting limited supply and a personal fit for their EMG wrist input and monocular display. On interoperability, SingleStore’s MCP push highlights how standardizing tools helps agents safely use live data. Science journalists found ChatGPT weak at translating nuance in research summaries, reinforcing the need for human oversight. YouTube embraces AI video creation with labeling and authenticity concerns, positioning AI as a creative amplifier rather than a replacement. Cosmology gets a speed boost with Effort.jl, an open source emulator in Julia that mimics complex models on laptops, with caveats. Finally, Windows 10 users can get a free year of security updates through October 2026 via enrollment, a practical step to keep systems secure while planning upgrades. Max wraps with three takeaways: secure your agents, favor interoperability, and keep the human creative edge.

Sources:

• Science journalists find ChatGPT is bad at summarizing scientific papers
• Zuck has the power! Meta applies to sell excess electricity
• Cracking product-market fit: Lessons from founders and investors at TechCrunch Disrupt 2025
• Why Does Building AI Feel Like Assembling IKEA Furniture?
• Is Tesla driving away with top EV sales?
• Universum to go: KI-Software emuliert kosmische Physik auf Notebooks
• YouTube Thinks AI Is Its Next Big Bang
• Un correo bastó para convertir a ChatGPT en espía: así fue como terminó filtrando datos privados de un usuario de Gmail
• Microsoft vuelve a subir los precios de las Xbox Series X|S en EEUU. Ya sabemos qué ocurrirá en Europa
• Get free Windows 10 security updates through October 2026: 2 ways
• Want Meta Ray-Ban Display glasses? You have to book a demo first - and they are going fast
• You can share Gemini Gems now - here's how

Max and expert guest Zane Okafor break down the latest AI shifts: OpenAI’s teen-safe ChatGPT with parental controls, the ShadowLeak prompt-injection attack on Deep Research agents (and how to defend against agent exfiltration), DeepSeek’s surprisingly low-cost R1 reasoning finetune and reinforcement learning gains, Alibaba’s open-source Tongyi DeepResearch agent and its MoE/128K design, Microsoft’s Copilot model selector that defaults to Anthropic Claude 4 Sonnet, India’s creative boom with Google’s Gemini 2.5 Flash Image (“Nano Banana”) and watermarking/privacy tradeoffs, the promise and challenges of using Transformers as robotic foundation models, and Op-Fed’s active-learning dataset for stance analysis in FOMC transcripts. Practical advice includes least-privilege agent design, co-use approaches for teens, and letting model routing pick the best tool for developer tasks.Sources:

• New attack on ChatGPT research agent pilfers secrets from Gmail inboxes
• Op-Fed: Opinion, Stance, and Monetary Policy Annotations on FOMC Transcripts Using Active Learning
• Microsoft’s deal with OpenAI in question as they trusted Anthropic for this new feature
• China's DeepSeek applying trial-and-error learning to its AI 'reasoning'
• India leads the way on Google’s Nano Banana with a local creative twist
• China’s Hypergryph unveils first gameplay of Arknights: Endfield, to launch in early 2026
• OpenAI launches teen-safe ChatGPT with parental controls
• Deepseek says training its R1 model cost just $294,000
• The Sequence Opinion #722: From Language to Action: Transformer Architectures as Robotic Foundation Models
• Alibaba Releases Tongyi DeepResearch: A 30B-Parameter Open-Source Agentic LLM Optimized for Long-Horizon Research

Today on Pulse on AI: We unpack the UK–US tech prosperity deal—Microsoft’s $30B UK commitment, Nvidia’s £11B end-to-end impact, a north‑east AI growth zone, and how nuclear and water shape datacenter realism—plus what a "sovereign" AI stack could mean for UK jobs and capability. Then, OpenAI’s new paper argues that benchmarks often reward guessing over honesty, driving hallucinations; we discuss incentives for models to say “I don’t know.” We cover Microsoft’s GitHub Copilot auto model selector favoring Anthropic’s Claude 4 Sonnet (and MAI‑1’s trajectory), and the rise of “Vibe Coding Cleanup Specialists” fixing AI‑generated code’s security and UX pitfalls. Next, Fluid Benchmarking brings psychometrics to LLM evaluation—adaptive item selection, ability scores, fewer mislabeled items, smoother curves—and we explain why production non‑determinism is a systems property, not just a temperature knob. Finally, we highlight the MIT/MIT-IBM scaling law guide for predicting big‑model performance from small runs, with practical best practices for checkpoints, budgets, and partial training. Three takeaways: infrastructure + energy policy drive AI competitiveness; calibrating uncertainty cuts hallucinations; and smarter evals and scaling laws save money while improving model quality.Sources:

• SENTRA: Selected-Next-Token Transformer for LLM Text Detection
• Microsoft’s deal with OpenAI in question as they trusted Anthropic for this new feature
• OpenAI says models are programmed to make stuff up instead of admitting ignorance
• Stable Virtual Camera: Multi-View Video Generation with 3D Camera Control
• Gemini tops the App Store thanks to new AI image model, Nano Banana
• AI scaling laws: Universal guide estimates how LLMs will perform based on smaller models in same family
• The Sequence AI of the Week #721: Stop Blaming Temperature: Fighting Nondeterminism in LLM Inference
• Vibe Coding Cleanup Specialists: Diese Menschen reparieren schlechten KI-Code
• Ai2 Researchers are Changing the Benchmarking Game by Introducing Fluid Benchmarking that Enhances Evaluation along Several Dimensions
• What is new in UK-US tech deal and what will it mean for British economy?

Max and AI expert Rohan Venkataram break down a packed slate of AI developments: Socotra’s Model Context Protocol (MCP) Server for secure agentic workflows in insurance; a University of York letter on how generative AI is undermining learning and how to redesign assessments; a founder’s playbook for resisting a platform ransom; a plain‑English tour of sparse autoencoders for interpretability; Tencent’s Hunyuan 3D 3.0 model with 1536³ voxel fidelity and free access; OpenAI’s GPT‑5 Codex for more autonomous, review‑aware coding agents; and Google’s VaultGemma work on differential privacy scaling laws. Three takeaways: standardization is how enterprises safely scale agents, autonomy requires human‑in‑the‑loop accountability, and privacy is a tunable trade‑off among noise, data, and compute.Sources:

• Socotra Launches MCP Server for Secure Agentic AI Integration
• Google releases VaultGemma, its first privacy-preserving LLM
• How I Stopped A Platform Ransom Against My Startup
• Tencent Unveils Hunyuan 3D 3.0 AI Model: Tripling Modeling Accuracy with Free Access
• Chery to Launch Hong Kong IPO Bookbuild on Sept 17, Targeting HK$140B Valuation
• OpenAI releases GPT-5 Codex designed for bug fixes and code generation
• The Sequence Knowledge #720: A Cool Intro to Sparse Autoencoders for AI Interpretability
• heise+ | KI-Bildbearbeitung im Vergleich: ChatGPT 5 und Gemini 2.5 Flash Image
• OpenAI Introduces GPT-5-Codex: An Advanced Version of GPT-5 Further Optimized for Agentic Coding in Codex
• How AI is undermining learning and teaching in universities | Letter

Today on Pulse on AI: Google Gemini overtakes ChatGPT in key app stores, fueled by a viral "Nano Banana" image tool and Google’s distribution advantage. Meta launches MobileLLM-R1, a sub‑1B parameter family that delivers state-of-the-art math/coding reasoning on the edge with far fewer training tokens. DeepMind’s AlphaEarth compresses global satellite and sensor data into 64‑dimensional embeddings for every 10×10m patch of Earth, enabling better climate, urban, and risk analysis. A major leak exposes Geedge Networks selling censorship tech modeled on China’s Great Firewall to multiple countries, highlighting governance risks. US states diverge on AI mental health: Illinois restricts AI in clinical decision-making; New York requires AI companions to disclose they’re not human and to route users to crisis resources. Finally, OpenAI’s GPT‑5‑Codex adds model and product safety measures, including sandboxing and anti‑prompt-injection defenses, for agentic coding. Three takeaways: distribution beats benchmarks, small specialized models are rising, and governance choices will shape AI’s real-world impact.Sources:

• A Multimodal RAG Framework for Housing Damage Assessment: Collaborative Optimization of Image Encoding and Policy Vector Retrieval
• Sigue estos pasos y consigue los mejores chollos en viajes gracias a chats de IA
• AI Policy Corner: Restriction vs. Regulation: Comparing State Approaches to AI Mental Health Legislation
• Addendum to GPT-5 system card: GPT-5-Codex
• AlphaEarth Provides New Ways to See, and Understand, Earth
• Google Gemini overtakes ChatGPT for the first time
• Learn How to Use Transformers with HuggingFace and SpaCy
• Meta AI Released MobileLLM-R1: A Edge Reasoning Model with less than 1B Parameters and Achieves 2x–5x Performance Boost Over Other Fully Open-Source AI Models
• Creíamos que el "Gran Cortafuegos" de China solo censuraba internet allí. Estábamos equivocados
• You can update your iPhone to iOS 26 right now (here's which models support it)

Today on Pulse on AI: Deepdub’s Lightning 2.5 pushes real-time, expressive multilingual voice to near‑instant responses with enterprise‑scale throughput; how to actually automate processes with multi‑agent systems using local LLMs, RAG, and guardrails; Albania appoints a virtual AI “minister” for procurement—promise and pitfalls around transparency, bias, and accountability; the sober reality of AI on the battlefield and why human veto must remain non‑negotiable; Indian IT’s pivot from services to AI‑product platforms amid policy headwinds; and why governance—not model horsepower—is the real unlock for GenAI ROI. Clear explanations, practical analogies, and concrete takeaways.

Sources:

• Deepdub Launches Lightning: Real-Time Voice Model for AI Agents
• A Wake-Up Call for Cash Rich, Idea-Poor Indian IT
• Latest TRON: Ares trailer takes us back to 1982
• SA fintech startup Float raises $2.6m funding to scale
• Terminators: AI-driven robot war machines on the march
• Alami Capital’s LaunchPad targets funding gap for African women-led startups
• Q&A: The missing link holding back GenAI adoption
• iX-Workshop: Unternehmensprozesse mit Multi-Agenten-Systemen automatisieren
• Deepdub Introduces Lightning 2.5: A Real-Time AI Voice Model With 2.8x Throughput Gains for Scalable AI Agents and Enterprise AI
• Albania tiene una ministra de IA. No una persona que gestiona la IA, no. Una IA que han nombrado ministra virtual

Max and AI expert Rayan Kovarik unpack a big week in AI. They explain the shift from pretraining to post‑training (instruction tuning, RLHF/RLAIF, preference modeling, and tool use) and why this “coaching” phase drives safety and usefulness. They dive into AWORLD, a distributed framework that speeds agent experience generation by 14.6x and boosts GAIA benchmark results using a Qwen3‑32B base. OpenAI’s ChatGPT gains full MCP write support, turning conversations into secure, auditable actions across enterprise tools. A hands‑on tutorial shows multi‑agent coordination with coordinator/researcher/analyzer/executor roles in a notebook‑friendly setup. The duo demystifies federated learning’s real‑world challenges and why software and optimization matter as much as theory. Microsoft reportedly brings Anthropic’s Claude into 365 for pragmatic, best‑fit AI features. TwinMind raises funding for an on‑device, privacy‑first “second brain” that transcribes all day, fuses offline and browser context, and answers personal queries. Leonardo accelerates Nordic cybersecurity acquisitions to advance a Zero Trust, European digital autonomy strategy amid the AI‑driven attack surface. They close with the robotics and AV frontier: simulation, sensors, and control systems that steadily bring intelligence into motion. Three takeaways: skill beats size, orchestration is the new UI, and trust is the backbone of AI deployments.

Sources:

• AWORLD：分散型フレームワークによるエージェントAIの効率的学習基盤 AWORLD：分散型フレームワークによるエージェントAIの効率的学習基 ... 2025年09月11日 LLM-Paper
• Optimization Methods and Software for Federated Learning
• Leonardo accelerates Nordic cybersecurity acquisitions to strengthen Europe’s defenses
• TwinMind raises $5.7M to launch AI second brain for offline note-taking
• Moroccan e-commerce startup Justyol raises $1m funding to drive regional growth
• Humanoids, AVs, and what’s next in AI hardware at TechCrunch Disrupt 2025
• The Sequence Opinion #718: From Scale to Skill: The Rise of Post‑Training
• Building Advanced MCP (Model Context Protocol) Agents with Multi-Agent Coordination, Context Awareness, and Gemini Integration
• OpenAI Adds Full MCP Tool Support in ChatGPT Developer Mode: Enabling Write Actions, Workflow Automation, and Enterprise Integrations
• Microsoft taps Anthropic for AI in Word and Excel, signaling distance from OpenAI