Ken Fricklas

AI Semantic Ablation: Just Remove the Meat? A while back in an upscale restaurant in Paris upon learning that my companion is a vegetarian, the waiter replied, "No problem, we'll just remove the meat". This opinion piece suggests that AI may be doing something similar to writing: semantic ablation. https://lnkd.in/gVEHmZrS

A major new benchmark puts frontier LLMs to the ultimate test. The Humanity's Last Exam (HLE) was built precisely because current benchmarks like MMLU (where models score >90%) are no longer challenging enough to measure true progress. This isn't just another test. It's a 2,500 question exam at the frontier of human knowledge in dozens of subjects, created by nearly 1,000 global experts from over 500 institutions. The key criterion is that Every single question had to stump the most advanced LLMs available during its development. When evaluated, all state of the art models showed very low accuracy. More tellingly, they exhibited poor calibration, often answering difficult questions incorrectly with high confidence instead of acknowledging uncertainty. This highlights a significant gap in expert level reasoning. HLE provides a precise measuring stick for the next leap in AI capability. It pushes us to build models that don't just recall information but can reason at an expert human level. The benchmark is now publicly available, setting a common goal for the research community. https://lnkd.in/gCKdurBg Explore the HLE benchmark: https://lastexam.ai #ArtificialIntelligence #MachineLearning #LLM #AIResearch #Benchmark #AGI #TechInnovation #FutureOfAI

10

Alphabet’s $4.75B acquisition of Intersect Power signals a fundamental shift in AI strategy: leadership is no longer just about models or algorithms—it’s about owning the infrastructure and energy that power them. Key insights & trends: • Energy as a strategic asset: Reliable, scalable, and sustainable power is now a core constraint for AI growth. • Vertical integration is the new moat: Hyperscalers are combining compute, data, and energy to control the AI supply chain. • Risk containment drives deals: Acquiring infrastructure reduces operational bottlenecks and protects product timelines. • Future of AI product strategy: Leaders must account for cost, scalability, and sustainability—not just features or speed. • Market pattern: Expect more M&A where AI companies secure critical infrastructure to maintain competitive advantage. The big takeaway: the AI race is evolving into a full-stack game, where winning requires designing resilient, scalable, and future-proof ecosystems. https://lnkd.in/gHx3pM5D #AI #GenerativeAI #ProductStrategy #AIInfrastructure #TechLeadership #BusinessImpact #SustainableTech #FutureOfAI #Google

5

During "Meaningful Use," we sponsored "undergoverned data creation." This was inefficient, clumsy, and expensive to map, but relatively benign. In the current state of interoperability, we are sponsoring "undergoverned data power." This is dangerous, balancing the value of data with its potential for harm. Without governance, convened data creation leads to a dangerous concentration of "data power" among a few convening entities, necessitating even more regulatory frameworks to ensure equitable benefits. And the cycle continues. When will policymakers learn that controls matter? "For health insurers, health data sharing consent is no longer a narrow compliance obligation. Modern regulations (like TEFCA and the 21st Century Cures Act) assume that because data can move via APIs, it will only be used for "lawful" purposes (like patient treatment) without the need for a national, standardized system to verify patient consent in real time. This is foolish, naïve, and dangerous for patients," says Jeff Rivkin, research director for Payer IT Strategies, IDC. https://lnkd.in/e3xWWvH8

11

1 Comment

Great article by Understanding AI Waymo’s next five cities are all in red states Liberal cities risk falling behind on driverless technology. by Timothy B. Lee and Kai Williams - This tells the story: Waymo’s fastest expansion is happening in states and cities that are willing to say yes to innovation. Meanwhile, some major urban centers risk sidelining themselves from one of the biggest transportation revolutions since the internal-combustion engine. Robotaxis are not just a tech novelty—they’re an economic earthquake. When transportation costs collapse 5–10× with autonomous fleets, we unlock entirely new business models, labor mobility, and urban design. PwC’s forecasts already show the shift: shared mobility, digital services, and AV platforms grow while traditional vehicle sales shrink. And it’s not just Waymo. Tesla is prepping its Austin robotaxi rollout. Uber, Nvidia, and multiple automakers are preparing 100,000 Level-4 vehicles powered by Nvidia’s DRIVE Hyperion 10 platform—2,000 FP4 TOPS of real-time safety compute in every car. My Recommendation for Policymakers and Cities: If the goal is safety, accessibility, and public trust, we need simple, practical policy: 1. Let customers choose human support. Robotaxi companies should allow riders to request a safety driver or human helper (for elderly, disabled, or anyone who prefers it). No complexity. No stigma. Just a checkbox. 2. No extra cost for choosing human involvement. If society wants a safe transition, we cannot make safety a premium feature. 3. Require remote human engagement in <1 minute. Any AV that encounters an unusual situation should escalate to a trained remote operator instantly. This protects safety and human jobs while preserving scalability. Cities that enable this blended model—human-optional autonomy backed by rapid remote assistance—will see earlier deployments, greater consumer trust, and faster economic growth. The future of transportation is arriving. The only question is which cities will benefit first.

39

15 Comments

AI is accelerating "vibe-coding", but Margaret-Anne Storey warns of a hidden cost: cognitive debt. When we let agents build without oversight, we lose the mental model of our own systems. Stop the fragmentation of shared theory. Dive into the risks at https://buff.ly/eKFDkj5 #ProductManagement #AI

7

AI interpretability: alarm system, or mechanic's toolkit? Good talk by Stanford’s Chris Potts on AI interpretability. We often hear Anthropic’s view: Interpretability is safety, alignment, and detecting "deception" in future superintelligence. A defensive posture—trying to understand black boxes so they can’t cause catastrophes. Potts offers more pragmatic contrast. He argues that if interpretability is "real" science, it shouldn't just help us watch models; it should help us build better ones. Of course, Anthropic doesn’t deny this, but he cautions about seeing their approach as inevitable given their prominence in the field of interpretability. He goes on to explore the skepticisms — i.e. that interpretability is impossible (due to explanations being at least as complicated as the models), don’t do much to improve models etc. An interesting discussion. https://lnkd.in/gdbrHhYn

10

A new standard for decision systems is now live. After extensive research and development, I’m sharing the formal publication of: AICOS Decision Authority Standard (DAS v3.0) This work introduces a governance-first framework redefining how decisions are evaluated, validated, and audited in complex systems. 📄 DOI: https://lnkd.in/dnrw6hQe Why This Matters Modern systems optimize prediction. Some optimize decisions. But almost none answer the fundamental question: 👉 What makes a decision valid? The Core Shift This work proposes a transition from: Prediction → Decision → Decision Authority Where decisions are no longer treated as outputs, but as validated and auditable objects Key Contributions • Formal definition of decision validity • Integration of risk, irreversibility, and governance • Economic impact mapping of decisions • Stochastic and control-theoretic framework • Multi-agent decision structure under constraints Real-World Relevance The framework is designed for high-impact domains: • Energy infrastructure • Financial systems • Strategic capital allocation Where decision errors are systemic, not isolated Positioning This is not an AI model. This is not a decision tool. 👉 It is a Decision Authority Infrastructure The work aligns with the broader movement toward governed systems in financial technology, including initiatives such as the Fintech Open Source Foundation (FINOS) under the Linux Foundation. References Control Theory · Decision Science · Risk Governance · Complex Systems · Economic Modeling #Dubai #UAE #AbuDhabi #USA #SiliconValley #NewYork #FinTech #AIGovernance #DecisionScience #EnergySystems #StrategicInfrastructure #AXION #AICOS #FINOS

1