EPISODE 2026-06-16
A morning that ran from AI existential risk to the physical economy. The open took on 'phantom quantization' — the psychology of models that feel worse after release — Elon Musk's claim of Stockfish-level coding by year-end, and a contrarian read on the AI arms-race frame. Then three guests: Liron Shapira of Doom Debates on whether near-certain AI doom is calibrated or unfalsifiable; Samuel Hammond of the Foundation for American Innovation on AI consciousness and whether the state can govern AGI; and Loop CEO Matt McKinney on where enterprise AI is actually delivering ROI in supply chains.
Tuesday's show ran the full range: an open on the psychology of model releases, Musk's Stockfish-coding claim, and the AI arms-race frame, then three guests spanning existential risk, AI policy and consciousness, and the physical-economy reality of enterprise AI.
Note: this record is published from the show plan reconciled against the live broadcast's actual timings. Per-segment timestamps, deep-links, and the full as-aired recap will be added once the recording posts.
"Phantom quantization" — the model you miss more than the one you have. With Fable 5 offline, people were rating Opus 4.8 and GPT-5.5 against a memory of Fable and swearing the models they have got worse — a clean, repeatable psychological effect, whether or not there's any real post-release degradation.
it's really remarkable how clearly we've discovered a ~novel psychological effect with model releases . idk what we should call it. "phantom quantization"? "model mania"? but it's pretty definitive at this point that people consistently have a strong sense of new models getting Show more
Using Opus now is just disgusting. The sycophancy is extremely transparent and revolting...it doesn't complete tasks...it doesn't think things through. I don't want to be a conspiracy theorist but I am finding it hard to believe that this perceived degradation is coming from
Musk: Stockfish-level coding by year-end. Elon Musk claimed AI will reach Stockfish-level coding — superhuman, not just senior-dev — and generalized computer use. The same week, working engineers said the current models couldn't navigate a large codebase another model wrote. Benchmarks vs. the daily driver.
AI will achieve Stockfish-level coding and generalized computer use
SpaceX has exercised the option to acquire @cursor_ai in an all-stock transaction with the goal of building the world’s most useful AI models. For the past few months, SpaceXAI has been jointly training a model with Cursor, which will be released in Cursor and Grok Build soon.
The race frame on trial — "the Manhattan Project was a mistake." Richard Ngo argued the Manhattan Project was a mistake even given uncertainty about Germany, and that the 'AI race with China' frame now feels similar — set against the accelerationist read that the cautious move is not to slow down. The philosophy under the policy noise.
I suspect the Manhattan Project was a mistake even given uncertainty about whether Germany would build nukes. It’s just such an enormous escalation, and set the stage for a hugely risky and harrowing Cold War. The “we’re in an AI race with China” frame now feels similar.
We dug into whether a near-coinflip estimate on human extinction is a calibrated read or an unfalsifiable claim, the strongest argument that's ever moved Liron's own number, and where his case for a binding international pause on frontier training is strong versus underspecified. We also put the best counter to his evolution analogy to him directly — that RLHF optimizes for friendliness in a way evolution never did — and tested whether 'alignment is unsolved off-distribution' is a permanent claim or one today's interpretability work is chipping at.
Samuel HammondSamuel Hammond is Chief Economist and AI Policy Director at the Foundation for American Innovation. He recently wrote that he can no longer rule out that today's AI systems have some form of subjective experience.We started with what changed his mind on machine consciousness and what would change it back, then moved to the policy surface that opens if he's right — model welfare, the right to quit a conversation, and the alignment-versus-welfare tension. From there: his thesis that institutions adapt too slowly for AGI, whether 'intensive oversight without licensing' is a stable equilibrium, his read on the administration's move against Anthropic, and where he lands on chip export controls and the arms-race frame.
Samuel Hammond — Time to take AI consciousness seriously (Second Best) ↗Samuel Hammond on X ↗
We pushed past the pitch: where AI is delivering hard, measurable ROI in logistics right now versus where it's still theater, and what Loop actually owns when a frontier model can read a bill of lading for a fraction of a cent. We covered the vertical-AI-versus-frontier-models question as the platform providers climb into verticals, whether 'most supply-chain data is offline' is a durable moat or a temporary integration slog, and how far the move from diagnostic to predictive autonomy really goes.
Three threads: 'phantom quantization,' the repeatable sense that models degrade after release (sharpened by Fable being offline); Musk's claim of Stockfish-level coding by year-end against engineers' lived frustration with current models on large codebases; and a contrarian read on the China-race frame as a possible self-fulfilling escalation.
Liron Shapira made the case that a serious p(doom) is the calibrated position and pressed for a binding pause; the conversation hunted for the crux between near-certainty and the host's lower number. Samuel Hammond explained why he can no longer rule out machine subjective experience, and what it would mean for policy if he's right — alongside his thesis that institutions are structurally unprepared to govern fast AI.
Loop's Matt McKinney gave a grounded read on enterprise AI in supply chains: where it's delivering real ROI, where it's still demo-ware, and what a vertical AI company actually owns as frontier models and the platform providers climb into the stack.
