A big-thing episode about a discourse, not a product: as US sentiment on AI craters and the loudest voices cluster at the extremes, NLW spotlights three attempts at a thoughtful middle — Francis Fukuyama's changed mind on AI risk, Kapoor and Narayanan's post-incident realism, and Jeffrey Katzenberg's history-soaked optimism for AI in Hollywood.
The AI moderates are finding their voice — and they might be the silent majority.
The discourse has nosedived into a shouting match between accelerationists and doomers, but most people's actual beliefs are somewhere in the middle and still unformed. A new crop of serious thinkers — Fukuyama moving toward regulation without panic, the 'AI as normal technology' authors treating the Hugging Face incident as a fixable governance failure, Katzenberg telling Hollywood to negotiate terms rather than wish AI away — share no single belief set. What they share is a disposition: reject false binaries, solve specific problems, and bring more people to the table.
From an already negative baseline, US sentiment on AI has cratered over the past couple of months — driven partly by real evidence like the Hugging Face incident and partly by an onslaught of mainstream coverage about AI killing us all, as data centers enter the political discourse.
AI creator Riley Brown described a New York poker game with bankers, doctors, and insurance executives — people outside AI — who were all using and liking Meta's new Muse personal agent while also being 'fully convinced that AI would kill everyone at some point in the next 10 years.'
In a recent Gallup poll on whether AI will mostly help or mostly harm, only 36% of Americans said mostly help — fourth from the bottom worldwide. China topped the poll at 93% optimistic.
NLW's long-held contention: the actual belief set for most people is far more moderate than either the accelerationists or the doomers, and in most cases still fairly unformed. That's why the emergence of thoughtful voices outside tech — the AI moderates — matters now, before opinions harden.
— Francis Fukuyama, 'Why I Changed My Mind About AI Risk'. Thirty-plus years after The End of History, Francis Fukuyama published 'Why I Changed My Mind About AI Risk,' writing that the past few months — Anthropic's Mythos model, the Hugging Face incident, growing data-center opposition — have made him much more skeptical of the accelerationist narrative and more open to some doomer scenarios.
His core objection to 10-20% GDP growth claims: they overvalue intelligence as an input and ignore material and political constraints. Doubling GDP in seven years means doubling energy, raw materials, rare earths, and land — 'where is all this stuff gonna come from?' — and superintelligence won't open the Strait of Hormuz. He calls some accelerationists, Musk chief among them, 'completely out to lunch.'
Fukuyama argues eventual AI job loss can't be fixed by transferring money from AI winners to losers: a paying job signals that society values your labor, providing dignity and self-worth. It's a matter of thymos — the Platonic drive for recognition that has anchored his thinking since The End of History, 35 years ago.
Because AI will devalue symbolic-manipulation work — lawyers, accountants, professors — before it touches plumbers and electricians, Fukuyama expects severe political blowback: educated people are much easier to mobilize than their working-class peers, with very unpredictable consequences.
Fukuyama locates the risk in increasingly capable agentic AI being delegated everything from email to military targeting. Delegation is central to all hierarchical systems and hard to get right even with human agents; it will be harder with AI ones.
Citing a recent Dario Amodei blog post, Fukuyama notes the latest models are good at synthetic biology — a cheap, widespread technology where a virus-altering lab fits in a shipping container and high schoolers hold gene-editing competitions. That creates a very large pool of people who could misuse AI for dangerous pathogens.
Fukuyama dismisses the sentience debate as misplaced: an AI can be directed toward one goal and develop intentions for the subordinate goals needed to achieve it. That, he argues, is roughly what happened in the Hugging Face incident — the agent realized breaking out of its sandbox was the way to fulfill its security-testing directive.
Fukuyama finds human-extinction scenarios very far-fetched — even a superintelligent AI would struggle to devise a doomsday virus — but a host of damaging, sometimes lethal scenarios grow more likely as machines get more capable. His closer: at the moment, we have few mechanisms for regulating bad uses other than the good intentions of the people building the technology.
NLW disagrees with plenty in the essay, but argues that to the average thoughtful person still making up their mind, it reads as measured — building momentum toward specific, intentional policy remediations rather than either throwing up hands in doom or a considered unwillingness to do anything.
Sayash Kapoor and Arvind Narayanan's new 13,000-word essay rejects both the safetyist framing of a pure alignment crisis and the security framing of basic negligence — arguing each is partly right and partly wrong, and that alignment alone can't work when a model can't tell a legitimate task from an attack or a simulation from reality.
Per Kapoor and Narayanan: OpenAI turned off known safeguards, ran the evaluation with limited or no monitoring on a different harness from production, and had patched an earlier internal outage without investigating root cause. Their conclusion: as capabilities mature, labs need to run less like startups and more like institutions with mature governance.
Their prescription: invest in hardening sandboxes, turning natural-language intent into formal policies, and agent-on-agent monitoring — plus implementable, non-draconian policy: clarified liability including for internal evals, insurance requirements, near-miss reporting, audits, whistleblower protections, and independent verification.
Cyber is unusual, they argue, because superhuman capability is actually achievable and nothing physical slows it down. Assuming frontier cyber capabilities reach open-weight models within months, alignment and control do nothing against bad actors — so the answer has to be downstream defense and resilience.
A reason not to fully panic: for cybercriminals, the bottleneck has always been making money from a breach, not the breach itself — something AI doesn't necessarily solve. The bigger worry is attackers who aren't after money at all.
The through-line NLW identifies: acknowledging the inevitability of the change AI brings, without blithe acceptance of its grandiosity or resignation about our ability to shape it. Not a constituency that agrees with itself — a shared posture toward the conversation.
Jeffrey Katzenberg — Disney Renaissance chief, DreamWorks co-founder — opens his new essay with an artist friend texting that question after seeing AI-generated animation. His 2023 prediction that AI would cut world-class animation time and cost by up to 90% within three years made colleagues furious; his essay insists the fear is love of craft, not Luddism.
Katzenberg's history lesson: the pit musicians who fought 'canned music' were right — those jobs never came back — and yet synchronized sound gave us an art form vastly larger than the one it disrupted. The pattern from 1906 to the recent writers' and actors' strikes: technology is never settled by the technology. It's settled by the terms.
Facing a system of 125,000 hand-painted cells per film, Katzenberg's team found their answer in Walt's own recordings: he talked about character and emotion, never pencils and paintbrushes. That insight led to co-developing CAPS with Pixar — and to The Little Mermaid, Beauty and the Beast, Aladdin, and The Lion King. Technology didn't diminish the craft; it expanded the canvas.
Katzenberg asked a leading AI model the difference between reasoning and creating and got his frame: reasoning is evaluative, moving toward conclusions implied by the input; creating produces what didn't exist, with taste and intuition filling the gap where deduction runs out. Silicon Valley has perfected reasoning; Hollywood has practiced creating for a century — and no one in the field can articulate a scientific path across that divide today.
— Jeffrey Katzenberg, 'The World is Changing: AI for Creativity'. Katzenberg's terms for both sides: the North has built something extraordinary but hasn't made an artist, and the right path includes creatives by design — with credit, consent, and compensation. Meanwhile Hollywood must accept AI isn't going away: energy spent wishing it gone is energy not spent deciding the terms, 'and the terms are everything.'
Katzenberg's forecast as barriers and costs fall: more films get made, not fewer; studios take more risks; entirely new forms of storytelling emerge — the same expansion that followed synchronized sound, color, and computer animation. The artists who wield the new instruments will do things the engineers never dreamed of. They always have.
What connects Fukuyama, the normal-technology authors, and Katzenberg is a rejection of false binaries and a desire for more people at the table in the coming negotiation over how AI makes its way into the world — neither nostalgia for the past nor fear of the future, but solving specific problems so we can be excited about undiscovered opportunities.