How agents will change banking forever | E2260

What It Covers

Jason Calacanis and Alex Wilhelm cover three converging stories: Andrej Karpathy's AutoResearch tool demonstrating recursive AI self-improvement, a NBC poll showing 46% of Americans hold negative views on AI, and live demos of AI agents executing real banking transactions, automating smartphones, and running self-optimizing website testing workflows across multiple platforms.

Key Questions Answered

• •Recursive AI Self-Improvement: Karpathy's AutoResearch tool on GitHub runs 5-minute LLM training loops, retesting and retaining only improvements. Shopify CEO Toby Lutke ran 37 experiments over 8 hours and achieved a 19% performance gain on an 800M-parameter model that outperformed a previous 1.6B-parameter model. Diminishing returns appear after roughly 83 experiments, but the core loop is functional and accessible to non-researchers today.
• •AI Banking Agents with Cryptographic Approval: NetxD's platform lets OpenClaw agents read bank balances and queue transactions, but executes payments only after the user approves via ECDSA private-key biometric sign-off in a mobile app. Practical use cases include auto-sweeping checking account buffers to maintain a $5,000 floor, capturing windfalls into savings buckets, and flagging underperforming interest rates — tasks humans routinely neglect due to tedium.
• •AI Public Trust Deficit: An NBC poll from March 2026 shows only 26% of Americans hold positive views on AI versus 46% negative — a minus-20-point gap. The hosts attribute this to a broken corporate social contract where surging profits now correlate with layoffs rather than raises, combined with gig-economy job erosion from autonomous vehicles and delivery robots threatening workers who previously relied on flexible platform income.
• •Career Defense Strategy Against Automation: To remain employable, workers should move up the task stack toward roles requiring physical presence and judgment. Skilled trades — carpentry, plumbing, electrical, fencing — currently pay $75–$100 per hour and remain beyond near-term robotic capability. Knowledge workers should position as AI orchestrators rather than individual contributors, managing agent workflows rather than executing repetitive tasks that LLMs now handle faster and cheaper.
• •Agent-Controlled Smartphones as Automation Layer: Phone Claw (getsupers.com) connects multiple Android devices to OpenClaw, enabling voice-commanded multi-phone automation through AR glasses. Agents can open apps, post to social platforms, and navigate interfaces across three simultaneous devices. The practical application extends beyond social posting to competitive intelligence — provisioning 10 Android devices to install, authenticate, and benchmark apps autonomously without human interaction at each step.
• •Self-Training Agent Governance Framework: Air Inc's AirTest platform runs multilayer recursive agent optimization, tracking token reduction per loop and routing expensive reasoning tasks to frontier models while offloading simple tasks to cheaper ones. Critically, the system includes a "blast radius" governance layer that flags agents receiving external signals as potential prompt-injection vectors and requests human permission before adding new high-risk tools to any compromised workflow.