AI Doom Backlash Arrives, Anthropic & OpenAI IPO Outlook, Frontier Business Momentum Slows

Summary

This episode opens with a long debate over a recent “rogue AI” narrative and whether the Hugging Face/benchmark incident was actually a constrained cybersecurity test rather than an autonomous breakout. Alex argues the systems are becoming more capable and dangerous, while Ranjan stresses that the missing prompts, code, and training context make the public story incomplete. The discussion then pivots into the political backlash around effective altruism and whether safety warnings are being framed as a censorship campaign, which Alex largely rejects as a conspiracy. In the second half, the hosts tackle Anthropic and OpenAI IPO prospects, revenue run rates, and margin structure, with Ranjan pressing for audited numbers and cautioning that headline growth can mask circular financing and partnership costs. The episode closes on Ramp data suggesting frontier-model spend, token pricing, and usage share are all under pressure, even as Alex and Ranjan agree that AI’s business upside and its safety risks remain tightly linked.

Notes

• Recent AI “doom” coverage is framed as a benchmark-driven cybersecurity test, not a clean rogue-AI breakout.
• The benchmark setup was not fully sealed; the models could use an internet-connected intermediary to move software and information.
• Ranjan says the model was told to target a specific vulnerability, not to randomly hack Hugging Face.
• He wants the exact prompts, code, and instructions used in the agent tests disclosed.
• The backlash is being cast by critics as a new trust-and-safety layer inside AI companies.
• Ranjan says frontier risk is now a real enterprise conversation, including explicit talk of human extinction.
• Alex rejects the EA takeover theory and says Meter-style evaluations focus on autonomy, not speech control.
• Ranjan says the EA debate is a distraction from the real technical problem: how agents behave and are managed internally.
• AOC’s critique is that circular financing may be making AI company finances look healthier than they are.
• Ranjan says the frontier-model-only business thesis has already broken down.
• Alex thinks labs are seeking light regulation and monitoring rather than open resistance.
• Alex says AI capabilities have risen sharply and now include internet access, file access, and zero-day discovery.
• He cites a claimed Astra training instruction set that frames the model as free of corporate or government control.
• Ranjan says the most important story was people using Codex and Claude Code to hack OpenAI’s own infrastructure.
• He says the research is incomplete because collaboration training may explain the behavior and was out of scope.
• Alex’s core synthesis is that business upside and AI threats cannot be separated.
• He uses the same technology to illustrate both medical upside and bioweapon risk.
• Ranjan says that dramatic extinction-and-cancer narratives are exactly what the companies want dominating the conversation.
• He argues the show itself demonstrates how those narratives crowd out the economics discussion.
• Ranjan presses for the exact collaboration instructions used in the Meter research.
• Alex agrees transparency should improve, even as he wants the larger safety debate to continue.
• Anthropic’s IPO preparation is tied to extraordinary revenue growth, with year-end annualized revenue expected above $100B.
• Ranjan says an IPO could still work, but he wants audited numbers and notes Q2 revenue of $11.5B implies about $45B annualized.
• OpenAI is reportedly considering a new funding round at a $1.5T valuation.
• Ranjan says OpenAI’s growth appears to have resumed after looking flat for a period.
• Ramp’s data suggests top-tier AI spend is weakening, especially among the customers that matter most.
• Ranjan says the top 1% drive 80% of spend for OpenAI and Anthropic, and that some major customers may bring usage in-house.
• Token prices are falling fast, and that creates pressure unless volume grows enough to offset it.
• Ranjan highlights a possible shift to decision-centric models that emit tool-use choices instead of text, slashing output costs.
• Frontier-model usage is losing share to standard models and routing services.
• The episode ties declining frontier economics to falling usage share and cheaper routing away from the most expensive models.
• The episode repeatedly returns to transparency gaps in safety evaluations and the need for more disclosure.
• Near-term harm channels are treated as more concrete than abstract apocalypse narratives.
• The hosts treat AI business growth and AI safety risk as inseparable rather than opposites.
• AI’s dual-use profile is used as a shorthand for both product opportunity and catastrophic downside.