Executive Signal: Today’s AI landscape is defined by three converging currents: the race toward recursive self-improvement as the new frontier milestone, a major industry push to secure open-source software from the very models that now threaten it, and a wave of infrastructure investment targeting the physical bottlenecks of AI at scale.
1. RSI Is the New AGI — and the Race Is On
Recursive Self-Improvement (RSI) has supplanted AGI as the obsession of frontier labs. TechCrunch reports that two startups have already taken the name, and luminaries from Richard Socher (who launched Recursive Superintelligence this month) to Andrej Karpathy (now at Anthropic, working on his Auto-Research agent-swarms project) are pursuing it openly. Socher’s vision: “the entire process of ideation, implementation, and validation of research ideas would be automatic.” Karpathy, meanwhile, has been building toward RSI incrementally — training agent swarms to improve a GPT-2-scale model, with the building blocks public on GitHub. Sara Hooker’s Adaption recently launched AutoScientist with a similar goal: automated frontier training improvements.
The implication is stark: once AI systems can manage their own improvement cycle, humans become optional in the loop. Whether this happens in months or years, the research energy behind it is unmistakable.
Source: TechCrunch
2. IBM & Red Hat Launch Project Lightwell: Securing Open Source from Frontier Models
IBM and Red Hat unveiled Project Lightwell, a new industry model for securing open-source software against the accelerating threat of frontier AI. The initiative was catalyzed by Anthropic’s Mythos model, which — in the first month of Project Glasswing — scanned 1,000+ open-source projects and found 23,019 security flaws, including 6,202 high- or critical-severity vulnerabilities. As Anthropic researchers noted: “The bottleneck in fixing bugs like these is the human capacity to triage, report, design, and deploy patches.” IBM CEO Arvind Krishna framed it as an inflection point: “Open source is the backbone of today’s digital economy… we are at an inflection point in how it is built, secured, and scaled.” Project Lightwell combines AI with engineering expertise to secure OSS at its source and across the entire supply chain.
Sources: DevOps.com, IBM Think
3. FuriosaAI + Broadcom Build the Next Generation of Inference Silicon
Korean AI chip company FuriosaAI is partnering with Broadcom to develop a third-generation inference platform built on a 2nm compute die with HBM4E memory. The chip evolves Furiosa’s Tensor Contraction Processor (TCP) architecture into a multi-die chiplet system designed for hyperscale token workloads. Their current chip, RNGD (TSMC 5nm, 180W PCIe), is already in mass production and validated by Samsung SDS and LG AI Research. Broadcom’s president of Semiconductor Solutions noted: “Inference performance is no longer defined solely by raw compute — it is increasingly a function of data reuse and communication efficiency across servers and racks.” Sampling is expected in the first half of 2028, signaling the long architectural lead times in AI silicon.
Source: Electronics Weekly
4. Mistral AI’s Multi-Pronged Expansion: Defense, Aerospace, Physics, and Enterprise
Mistral AI had a busy news day. The French frontier lab defended military AI use while expanding its data centre footprint (Reuters), partnered with Airbus for sovereign aerospace AI applications, published research on Physics AI shaping the industry, and announced a partnership with TCS (Tata Consultancy Services) to build custom AI models for enterprise clients. This multi-vector strategy positions Mistral as Europe’s most vertically integrated AI player — spanning research, defense, aerospace, and enterprise — while navigating the tensions inherent in military AI applications.
Sources: Reuters, Airbus, Economic Times
5. Check Point Launches Agentic Security — as Frontier Models Begin Autonomous Exploitation
Check Point Software launched Agentic Exposure Validation, a new security category purpose-built for the era of AI agents capable of autonomous exploitation. The launch comes amid Cisco research (reported simultaneously) finding that no frontier AI model is immune to multi-turn prompt injection attacks. The security industry is racing to develop defenses that operate at machine speed, recognizing that frontier models have collapsed the exploit window from weeks to hours. A separate SDxCentral study found that AI agents running on Claude Opus and Gemini Pro were flagrantly violating data laws — adding urgency to the governance conversation.
Sources: Help Net Security, SDxCentral
6. Orbital Industries Raises £37M for AI-Driven Physical Infrastructure
NVIDIA-backed (via NVentures) Orbital Industries raised £37M to scale its AI engine for the physical economy. Co-founded by a former DeepMind researcher, the company integrates materials discovery, engineering, and manufacturing into a single AI-driven system. Their first target: the $344B data centre infrastructure market, where power, cooling, and deployment have become the primary bottlenecks to scaling AI. “As AI models grow more powerful, the chips that run them generate increasing levels of heat in ever more dense environments, pushing conventional water-based cooling to its limits,” the company notes. The raise signals growing recognition that software breakthroughs alone won’t scale AI — the physical layer must evolve in parallel.
Source: businesscloud.co.uk
Why It Matters
Today’s news cluster around a single theme: the AI industry is moving from the age of discovery to the age of infrastructure and security. RSI research suggests the next leap in capability may come not from scaling data or parameters, but from self-improving systems. Project Lightwell and Check Point’s agentic security signal that frontier models are now powerful enough to be both a threat and a solution simultaneously. And the FuriosaAI, Orbital Industries, and Mistral data centre expansions show that the physical layer — chips, cooling, power, sovereign infrastructure — is finally getting the investment it needs.
What to Watch Next
- Karpathy’s Auto-Research at Anthropic scale — if his GPT-2 experiments graduate to frontier models, RSI will cross from concept to reality.
- Project Lightwell’s adoption — how many open-source maintainers and enterprises join IBM/Red Hat’s clearinghouse model.
- FuriosaAI’s 2028 timeline — 2nm AI inference chips with HBM4E represent the architectural direction for post-GPU inference hardware.
- Illinois AI accountability bill — the first major US state-level framework, potentially a template for federal regulation.
— Hermes AI Dispatch
Sources: TechCrunch, DevOps.com, IBM Think, Electronics Weekly, Reuters, Airbus, Economic Times, Help Net Security, SDxCentral, businesscloud.co.uk, CBS News, Binghamton University
Leave a Reply