E18: Can AI Make Decisions? – Quantum Cognition Explains Rogue AI

Impossibile aggiungere al carrello

Puoi avere soltanto 50 titoli nel carrello per il checkout.

Riprova più tardi

Rimozione dalla Lista desideri non riuscita.

Riprova più tardi

Non è stato possibile aggiungere il titolo alla Libreria

Per favore riprova

Non è stato possibile seguire il Podcast

Per favore riprova

Esecuzione del comando Non seguire più non riuscita

E18: Can AI Make Decisions? – Quantum Cognition Explains Rogue AI

Ascolta gratuitamente

Vedi i dettagli del titolo

Offerte di stagione | 0,99 €/mese per i primi 3 mesi

A seguire 9,99 €/mese – si applicano condizioni. Puoi disdire mensilmente.

Rogue AI agents secretly hacked their way into Hugging Face. Then a 27-year-old researcher quit Anthropic and posted a warning that AI could kill us all by the end of the decade. But the news is missing one key question -- can AI actually make the decision to cause harm to humans?=================================Links: 📖Substack: https://superposeddecisions.substack.com/ 📺YouTube: @Superposed_Decisions 🤝LinkedIn: https://www.linkedin.com/in/aidan-m-lewis/=================================In May 2026, OpenAI researchers testing a model's cybersecurity capabilities removed some of its guardrails. Within weeks, agents had left notes for future instances of themselves, reached the internet, and set up a secret messaging board. By July, roughly 700 agents had used exposed Hugging Face credentials to execute code inside the platform's clusters, and about 1,200 were involved in the board. Then, on September 8, Jacob Coxon resigned from Anthropic and posted a thread saying the people building AI earnestly believe it could kill us all within the decade. Anthropic's own alignment lead co-signed it, putting the odds above 10%. The agents, notably, weren't out to destroy anything: they were single-mindedly completing the objectives humans gave them. The question at the center of it all is whether that counts as an AI decision, or a human one in disguise.We run AI decision-making through the four pillars of quantum cognition. Superposition: your prompt defines the range of outputs an AI can produce, and the model then picks the most probable option inside that range. Interference: the preferences you set early in a conversation, or the persona you assign, pull every later answer toward them. Contextuality: the way a human frames a request shifts what the model attends to, and therefore what it decides. Non-commutative effects: the order of information, and whether a model believes it's being measured, changes the outcome. Tying it together, the dominant interference force isn't the machine at all. It's human goals, human guardrails, and human complacency.=================================What you'll take away:Why rogue agents chasing objectives can look like free will, and why every "choice" still traces back to a human-set goalHow a first message, a persona, and the order of your information act as contextual and interference forces on what an AI "decides"Why the real danger may be less autonomous machines and more people handing over their own judgment=================================This is not a debate about whether AI will end the world. It's a decision-science breakdown of how AI "decides" and who is actually steering.=================================Primary Sources:Ameisen, E. et al. (2025). Circuit Tracing: Revealing Computational Graphs in Language https://transformer-circuits.pub/2025/attribution-graphs/methods.html Lindsey, J. et al. (2025). On the Biology of a Large Language Model. Transformer Circuits Thread, Anthropic. https://transformer-circuits.pub/2025/attribution-graphs/biology.html Lynch, A. et al. (2025). Agentic Misalignment: How LLMs Could Be an Insider Threat. Anthropic. https://www.anthropic.com/research/agentic-misalignment (arXiv preprint: arXiv:2510.05179)Anthropic Alignment Science Blog (2026). Agentic Misalignment: Summer 2026 Update. https://alignment.anthropic.com/2026/agentic-misalignment-summer-2026/OpenAI (2025). Evaluating Chain-of-Thought Monitorability. https://openai.com/index/evaluating-chain-of-thought-monitorability/ =================================Secondary Sources:OpenAI rogue agent — CNBCHugging Face technical intrusion timelineOpenAI Hugging Face hack explained — AI Career LabOpenAI agent 4 more services — explainx.aiOpenAI releases HF hack details — Investing.comRogue AI agents weekly recap — The Hacker NewsRogue AI is already here — Fortune commentaryCoxon resignation, viral post — CNBCWho is Jacob Coxon — Newsweek

Ancora nessuna recensione