AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Will we have Superintelligence by 2028? With Anthropic's Ben Mann

· No Priors

Ben Mann, Anthropic co-founder and GPT-3 co-author, joins No Priors to discuss the Claude 4 release, the lab's strategic bets, and his views on the path to transformative AI. He explains Anthropic's versioning philosophy (major bumps only when capability shifts feel significant), highlights Claude 4's dramatic coding improvements—especially the reduction of over-eager edits and reward hacking—and points to multi-hour unattended agentic workloads (including the Manus video-to-PowerPoint example) as the key new capability unlocked. On competitive positioning, Mann frames Anthropic as the "Adyen to OpenAI's Stripe": enterprise-strong but lower consumer mindshare. He defends forward integration into coding via Claude Code as essential for learning and improving the model, predicts a routing layer will eventually replace user model selection, and walks through MCP's evolution from internal hack to industry standard now adopted by OpenAI, Google, and Microsoft. He also describes Claude 4 Opus's ASL3 classification, citing measurable bio-uplift over Google search as the trigger for stricter safeguards. Mann endorses the AI 2027 forecast as "quite possible," defining transformative AI by when agents pass an "economic training test"—being hired in place of a human for roughly 50% of economically valuable tasks. He describes accelerating internal R&D via Claude Code (data analysis, log tailing, environment construction) as a recursive self-improvement loop, while noting that human expert feedback is hitting expertise ceilings, pushing Anthropic toward RL-AIF, Constitutional AI, preference models, and—where possible—real-world verifiers (e.g., the Novo Nordisk cancer-treatment-reporting partnership). He also discusses why computer-use agents aren't yet consumer-deployed (irreversible-action and prompt-injection risk) and why Anthropic published its deceptive-model alignment-faking research rather than suppress it.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.