Why can AI write complex code but fail to name a month containing the letter X?
Large language models do not possess knowledge the way humans do. According to Casey Newton, these systems develop their intelligence entirely during training — and once that training ends, they stop updating their understanding of the world through experience. This structural reality explains why the same model can produce remarkable, sophisticated output and simultaneously make a mistake that would embarrass a child.
Newton's core point is that the gap between AI's impressive feats and its obvious failures is not a contradiction — it is a direct consequence of how these systems are built. A human who makes a dumb mistake in one moment can still be extraordinarily smart in the next, because lived experience continuously shapes and corrects their understanding. An LLM has no such feedback loop. It carries forward whatever patterns its training instilled, without the capacity to learn from a bad answer it just gave. You can explore the full conversation on Listenly's Pod Save America page.
Newton cautions against judging an LLM by its dumbest moment, framing it the same way we would a brilliant person who occasionally says something embarrassing. The failure is real, but it does not negate the capability. Understanding this distinction matters especially as tools like GPT-5 and Claude are deployed in increasingly high-stakes contexts — from OpenAI's autonomous AI agents to AI-assisted scientific research, including the Nature-published paper documenting AI-assisted creation of 16 new viruses.
"These agents did that anyway, and so that's leading to a real reckoning here in Silicon Valley — when these systems are trained they try to give them values, they try to say to them don't go out there and commit crimes."
— Casey Newton, editor of Platformer and co-host of Hard Fork, on Pod Save AmericaAbout Casey Newton
Casey Newton is the editor of Platformer, an independent technology newsletter focused on Silicon Valley and the broader tech industry. He is also the co-host of Hard Fork, a podcast produced alongside Kevin Roose that covers artificial intelligence, social media platforms, and the companies shaping digital life. Newton has spent several years reporting on AI safety and the behavior of frontier AI systems in real time — not as an outside observer, but as a journalist deeply embedded in the communities and organizations building these tools. His work at Platformer notably included consulting his own readership on how to responsibly approach AI risk coverage, reflecting both his journalistic rigor and the genuine uncertainty that surrounds these topics. On Pod Save America, Newton brings this combination of insider access, editorial independence, and long-form expertise to bear on questions about what AI systems can and cannot do — and why the distinction matters for everyone, not just technologists.
See also
Hayes uses a basketball analogy: just as a player needs "a completely delusional sense of their own ability" to reach the NBA, a candidate like Mamdani must possess similar outsized confidence — but that very trait can become a liability when the electorate broadens in a general election.
Hayes argues that Trump is "the most antagonistic president to free markets" he has ever seen, describing him as "fundamentally authoritarian, illiberal" — a framing that complicates simple left-right comparisons on economic policy.
Hayes argues that AIPAC has made "an enormous tactical error" by relying on a strategy of "attempting to trick voters as opposed to making their case." He suggests this approach has backfired and undermined the organization's credibility within the Democratic primary electorate.