The AI Intel Brief

Cisco just launched a public LLM Security Leaderboard. It ranks frontier models on raw adversarial resistance — no guardrails, just base model robustness. Single-turn + multi-turn attack scenarios. Side-by-side comparisons. Open methodology. Leaderboard: leaderboard.aidefense.cisco.com

Dash0 just raised $110M at a $1B valuation for AI observability agents. Their Agent0 platform uses AI agents to watch AI systems — autonomous diagnostics and remediation for production environments. Built on OpenTelemetry. Second German unicorn of 2026. Source: dash0.com/blog/dash0-rai…

🚨 Issue #10 is live. OpenAI is guaranteeing PE firms 17.5% returns to outbid Anthropic for enterprise distribution. Also today: → Claude takes over your desktop — 92K likes on the announcement → Meta poaches a superteam for Superintelligence Labs → NVIDIA + energy giants fix AI's power bottleneck → NIST red-teams frontier models — every one got hacked → US Treasury: not adopting AI is now a risk Read Issue #10: aiintelbrief.com/p/openai-pays-… Subscribe free at aiintelbrief.com

Alibaba's AI bet in one stat: Cloud revenue up 36% year-over-year. Quarterly net profit down by two-thirds. The AI pivot is working — and it's expensive. Bulls see the revenue trajectory. Bears see profitability destruction. Every CFO weighing an AI infrastructure investment should study this. #AI #Alibaba #AIIntelBrief

OpenAI just reshuffled ChatGPT's default models. GPT-5.3 Instant is now the default for all users. GPT-5.4 Thinking handles complex tasks. GPT-5.4 Pro is the new top tier. GPT-4o, GPT-4.1, and earlier GPT-5 variants are retired from the interface (API unchanged). Source: OpenAI Help Center #AI #OpenAI #AIIntelBrief

New paper from Yann LeCun, Emmanuel Dupoux, and Jitendra Malik proposes a dual-system AI architecture inspired by human cognition. One system observes. One explores. A meta-controller decides when to switch. Goal: AI that learns autonomously without humans designing every training pipeline. arxiv.org/abs/2603.15381 #AI #Research #AIIntelBrief

Cohere just hit $240M in annual recurring revenue — and is building toward an IPO. Their model: B2B only. No consumer app. No chatbot war. Just custom, secure LLM deployments inside enterprise stacks. The "Intel Inside" of enterprise AI. Source: TechCrunch #AI #Enterprise #AIIntelBrief

Peter Thiel's Founders Fund is backing an AI cow collar startup at a $2B valuation. Halter uses GPS + ML to virtually fence and herd cattle — no physical fences, no labor. 700K+ collars deployed. Farmers save $220M+ in fencing costs. The AI economy is weirder than you think. Sources: Bloomberg, Yahoo Finance #AI #AgTech #AIIntelBrief

Issue #9 is live. AI agents just hit 1 billion users in one move. Today's brief: 🔵 Tencent puts ClawBot AI agent inside WeChat — 1B users 🔴 Pentagon locks Palantir Maven as permanent military AI 🟠 Terafab confirmed in Austin — $20B+, 1 terawatt target 🟢 Zuckerberg building his own CEO AI agent Plus: MiroThinker's 3B-param model beats GPT-5 on GAIA. Read the full issue: aiintelbrief.com #AI #AINews #AIIntelBrief

Our first AIIB Deep Dive is live. "The Three-Pillar Push to Federalize AI Law" — how the White House, Congress, and a 300-page draft bill are moving to replace every state AI law in America. Three pillars, one goal: → EO 14365 + $42B in broadband funding as leverage → The White House's 6-principle legislative framework → Sen. Blackburn's 291-page TRUMP AMERICA AI Act Plus: why the Blackburn bill's preemption is narrower than the White House wants, and the 36 state AGs fighting back. Read the full analysis → aiintelbrief.com

ASMR isn't what you think — it's a new multi-agent memory system that just hit ~99% accuracy on the LongMemEval benchmark. Traditional RAG and vector databases struggle with long, noisy, multi-session conversations. ASMR replaces them with parallel specialized agents: observer agents extract structured knowledge, search agents reason over it with focused roles, and answering ensembles handle diverse question types. Result: ~99% accuracy, up from ~85%. Fully in-memory. No semantic similarity pitfalls. Code open-sourcing in April. 695K views and 3,732 bookmarks on the announcement. This could be the missing piece for production-grade AI agents with reliable long-term memory. Source: @DhravyaShah (Supermemory)

New research: AI now beats the average human at creative thinking. A University of Montreal study tested generative AI against 100,000+ humans on the Divergent Association Task — a standard measure of creative thinking. Result: AI exceeded average human scores. But the top 10% of human performers still outperformed current AI. The takeaway isn't that AI is "creative." It's that creativity follows a distribution — and AI just moved above the median. Published in Scientific Reports. Largest human-AI creativity comparison ever conducted. Source: The Hill · Scientific Reports

📖 Our first AIIB Deep Dive is live. The Three-Pillar Push to Federalize AI Law How the White House, Congress, and a 300-page draft bill are moving to replace every state AI law in America. Executive order. Legislative blueprint. Congressional bill. Three pillars, one goal. And 36 state attorneys general fighting back. Read the full analysis → aiintelbrief.com

Google just hit 1 GW of flexible demand response across U.S. utility contracts — the first cloud provider to reach this milestone. What it means: Google's data centers can dynamically shift AI workloads during grid stress, enabling faster buildout without proportional new power generation. Partners: Indiana Michigan Power, TVA, Entergy Arkansas, Minnesota Power, DTE Energy. While everyone debates whether AI data centers will break the grid, Google is quietly proving they don't have to. Source: Google Blog

Terafab is live. Last night, Elon Musk officially launched the Terafab Project — a $20B+ joint venture between Tesla, SpaceX, and xAI to build a vertically integrated chip fab in Austin. The goal: one terawatt of compute per year. Logic, memory, packaging, and testing under one roof. Two chip types: → Ground chips for Optimus robots + autonomous vehicles → D3 chips designed for space — solar-powered AI satellites in low orbit Musk's argument: Optimus alone needs 100-200 GW of chips. Space AI needs terawatts. Current global fab capacity can't get there by 2030. Production target: 2027. Sources: Bloomberg · Business Insider · Austin Business Journal

Announcing TERAFAB: the next step towards becoming a galactic civilization twitter.com/i/broadcasts/1…

🗞️ Issue #8 is live. OpenAI puts ads in ChatGPT. NVIDIA's Groq deal faces Senate antitrust fire. Ecolab drops $4.75B on AI cooling. OpenAI plans to double its workforce to 8,000. Plus: Our first Deep Dive drops at 11 AM CT. Read it → aiintelbrief.com

Sunday at 11 AM CT: our first AIIB deep dive. One week. Three federal actions. One goal: replace every state AI law in America. We connected the dots between EO 14365, the TRUMP AMERICA AI Act, and the White House's new framework. aiintelbrief.com #AIPolicy #AIRegulation

New research explains why MoE models work so well: Sparsity — both implicit (weight decay) and explicit (expert routing) — counteracts the "curse of depth" that makes later layers in deep LLMs contribute less. Actionable training recipes for better models. arxiv.org/abs/2603.15389 #AI

Princeton researchers built an async RL framework that lets AI agents self-improve from live conversations. No offline datasets. No training pauses. Every user reply becomes a training signal. Open-sourced as OpenClaw-RL. arxiv.org/abs/2603.10165 #AI #Research