Today’s episode dives into real security and efficiency shifts across AI. We explore a new open-source federated learning attack simulator on FEDn and an adaptive aggregation method (EE‑Trimmed Mean) that resists label-flipping and subtle model poisoning—even with non‑IID data and late joiners. We unpack China’s SpikingBrain1.0, a brain‑inspired, localized‑attention LLM on domestic MetaX chips, plus practical tips and privacy notes for generating chibi characters from your face. We break down Alibaba’s Qwen3‑Max trillion‑parameter MoE buzz and why post‑training matters, then examine Kalpa Labs’ quest for sub‑200ms, multilingual voice with lower audio token rates and sparse routing. We cover Google’s AI Plus rollout in Indonesia (Gemini 2.5 Pro + NotebookLM), Anthropic’s proposed US$1.5B author settlement and its implications for licensing and provenance, and MEGS2’s big VRAM cuts for 3D Gaussian Splatting via Spherical Gaussians and unified pruning. We highlight 2L‑FUSE for interpretable kernel‑based feature reduction and close with Oracle’s staggering $455B RPO signaling a sustained AI datacenter boom. Three takeaways: build resilient systems, chase efficiency over brute force, and track policy and infrastructure as closely as model benchmarks.

Sources:

• Google Rolls Out AI Plus to Rival ChatGPT Go
• Why the Next Leap in Speech AI Comes from Kalpa Labs
• MEGS$^{2}$: Memory-Efficient Gaussian Splatting via Spherical Gaussians and Unified Pruning
• Feature Understanding and Sparsity Enhancement via 2-Layered kernel machines (2L-FUSE)
• China develops SpikingBrain1.0, a brain-inspired AI model
• Mega-and-MAGA deals position Oracle's Larry Ellison to overtake Elon
• An AI startup has agreed to a $2.2 billion copyright settlement. But will Australian writers benefit?
• The Sequence AI of the Week #717: First Trillion Among the Majors: Qwen-Max
• Fighting Back Against Attacks in Federated Learning
• Cómo crear un personaje manga estilo chibi basado en tu cara usando inteligencia artificial, en ChatGPT y Gemini

Max and AI expert Maya Haddix unpack a packed slate of AI developments. They cover Google’s new AI Literacy hub for parents, students, and educators; AWS SageMaker HyperPod’s managed tiered checkpointing to speed and stabilize massive model training; the UAE’s K2 Think open-source reasoning model; Mistral AI’s latest funding, models, and enterprise features; Elastic’s practical guide to the OWASP Top 10 for LLM apps; GibsonAI Memori’s SQL-native agent memory; Sam Altman’s take on bots making social media feel ‘fake’; the “Critterz” animated feature using OpenAI models; mmBERT’s fast multilingual encoder; and a research paper on causal clustering for CATE-driven subgroup discovery. Three takeaways: build AI literacy, use defense-in-depth for engineering and security, and watch the trend toward smaller, smarter specialized models.

Sources:

• Causal Clustering for Conditional Average Treatment Effects Estimation and Subgroup Discovery
• Accelerate your model training with managed tiered checkpointing on Amazon SageMaker HyperPod
• The latest Google AI literacy resources all in one place
• mmBERT: ModernBERT goes Multilingual
• New animated feature film to be produced using OpenAI’s AI models
• Sam Altman says that bots are making social media feel ‘fake’
• What is Mistral AI? Everything to know about the OpenAI competitor
• Guide to the OWASP Top 10 for LLMs: Vulnerability mitigation with Elastic
• GibsonAI Releases Memori: An Open-Source SQL-Native Memory Engine for AI Agents
• The United Arab Emirates Releases a Tiny But Powerful AI Model

In the debut episode of Pulse on AI, host Max Dreyfus and expert guest Dr. Mateo Aranda unpack a packed slate of real AI news and research. They break down Anthropic’s endorsement of California’s SB 53 and what its transparency and whistleblower provisions mean for frontier AI governance. They explore DeepSeek’s pivot from chasing another massive base model to building ambitious agentic AI, shaped by chip constraints and product pragmatism. The market pulse includes August’s funding slowdown and a rapid September rebound led by Anthropic’s $13B round, with additional highlights on active lead investors and AI-specific M&A.

They examine Alibaba Cloud’s Qwen‑Max—an OpenAI‑compatible trillion‑parameter Mixture‑of‑Experts model framed for production—then shift to OpenAI’s $1.1B Statsig acquisition as a bet on experimentation at scale. The research roundup covers: ethical LLM decision support in construction (use as a tool, keep humans in the loop); diagnosing LLM uncertainty via multi-response disagreement patterns; a CFA-based financial reasoning benchmark where RAG fixes knowledge gaps; invoice processing where native image-based multimodal parsing often wins; and VCMamba, a hybrid CNN + Mamba vision backbone with strong global-local tradeoffs.

The episode closes with three takeaways: transparency is becoming table stakes; the shift from raw model size to post-training, agents, and system design is accelerating; and pragmatic deployment tactics—MoE, RAG, multimodal inputs, and experimentation platforms—are driving real-world wins.

Sources:

• Can Multiple Responses from an LLM Reveal the Sources of Its Uncertainty?
• Evaluating Large Language Models for Financial Reasoning: A CFA-Based Benchmark Study
• Multi-Modal Vision vs. Text-Based Parsing: Benchmarking LLM Strategies for Invoice Processing
• The Ethical Compass of the Machine: Evaluating Large Language Models for Decision Support in Construction Project Management

• VCMamba: Bridging Convolutions with Multi-Directional Mamba for Efficient Visual Representation
• OSC: Cognitive Orchestration through Dynamic Knowledge Alignment in Multi-Agent LLM Collaboration
• Active US Investors Slowed Some In August
• Anthropic endorses California’s AI safety bill, SB 53
• The Sequence Radar #715: Qwen-Max: The Trillion-Parameter MoE You Can Actually Ship
• El mundo está esperando el nuevo gran modelo de DeepSeek para competir con GPT-5, pero DeepSeek tiene otros planes: la IA agéntica