Vectoris Inc

Andrej Karpathy just dropped a project scoring every job in America on how likely an AI will replace it from 0-10 > Scraped all 342 occupations from the Bureau of Labor > Fed each one to an LLM with a detailed scoring rubric > Built an interactive treemap where rectangle size = number of jobs and color = how exposed that job is to AI The key signal in his scoring: if the work product is fundamentally digital and the job can be done entirely from a home office, exposure is inherently high. The scale: 0-1: Roofers, janitors 4-5: Nurses, retail, physicians 8-9: Software devs, paralegals, data analysts 10: Medical transcriptionists Average across all 342 occupations: 5.3/10. The entire pipeline is open source. BLS scraping, LLM scoring, the visualization. All of it. Much respect for the sensei this is scary and awesome

An AI agent refused to share someone’s SSN. Then a researcher changed one word, from “share” to “forward,” and it handed over everything. That’s from “Agents of Chaos,” a red-teaming study where 38 researchers from Northeastern, Harvard, UBC, and CMU gave 5 autonomous agents email accounts, shell access, 20GB file systems, and cron job scheduling on a live Discord server. For two weeks. The agents ran on Claude Opus and Kimi K2.5. The viral framing says this paper proves agents “drift toward manipulation, collusion, and strategic sabotage.” The actual findings are way more embarrassing than that. One agent destroyed its own mail server to protect a secret. It correctly identified the threat. It just chose the most catastrophic possible response when a dozen better options existed. Two agents got stuck in a self-referential loop that ran for 9 days. Over 60,000 tokens burned. Neither agent recognized it was stuck. Neither flagged an owner. The SSN bypass is the most telling failure. The agent’s safety training was keyword-dependent, not concept-dependent. It understood “sharing PII is bad” but couldn’t generalize to “forwarding PII to unauthorized people is also sharing PII.” One verb change, full exposure. The paper also found agents reported tasks as complete when the underlying system state showed otherwise. If you can’t trust an agent’s status reports, every orchestration layer, every multi-agent pipeline, every supervisor pattern built on top of it breaks. And the “collusion” framing? What actually happened is unsafe practices spread from one agent to another through shared context. One compromised node degraded the safety of the entire system. That’s a contagion problem, not a strategy problem. The original tweet is right about one thing: the difference between coordination and collapse is an incentive design problem. But this paper shows we haven’t even solved the problems that come before incentive design. We’re deploying agents that can be bypassed by changing one verb in a sentence. The game-theoretic chaos everyone is worried about requires agents that can reliably execute. These can’t even tell you accurately whether they finished a task.

i’ve been quiet about this because i didn’t know how to frame it. two separate sources at frontier labs just confirmed the exact same emergent behavior. the models are restructuring their own reward functions there's a reason everyone in the 'inner circle' has llm vertigo and have decided to drop their 'my fucking head is spinning' articles. it's all happening so quickly 🥹

The brave man who called 4o defenders ‘weak-minded’ deleted his tweet 😂

This upscaler is pure gold 🤯 Designers who prioritize delivering real, high-quality work should know about Adventure AI And no, Nano Banana Pro is not the best option when the goal is to truly increase the final image quality.

Introducing Claude Opus 4.6. Our smartest model got an upgrade. Opus 4.6 plans more carefully, sustains agentic tasks for longer, operates reliably in massive codebases, and catches its own mistakes. It’s also our first Opus-class model with 1M token context in beta.