A return to “When they call it AGI or Generative AI, I cringe and then realize the content to follow is about money and not reality,” first published on Medium on September 26, 2025.
A year ago I opened an essay by admitting that I cringe when people call it AGI, because what follows is usually about money and not reality. That piece cited Andrew Ng, Emily Bender, and Timnit Gebru. In the past two weeks, all three arrived from opposite corners of the field at the point that essay was built around. The machine is never the one who answers for what it does. None of them cited me, and none needed to, yet the record shows how an argument holds up when the evidence keeps arriving.
What I wrote a year ago
The opening position was plain. What we call Generative AI is a computer system that “reflects human inputs and incentives with remarkable fluency, then presents that reflection as if it were insight.” I leaned on the stochastic parrots framing from Bender, Gebru, and two coauthors to explain why fluent systems invent details that sound right and are wrong (Bender et al., 2021). The danger I described came from people, who encode their own fears, biases, and priorities into systems that now scale across the world. I put it in one line: “Every algorithm is a mirror of its makers.”
The money in the title was there for a reason. Kate Crawford traces how the AI pipeline extracts resources, data, and labor, then packages the result as progress (Crawford, 2021). My answer was that “if hype is the impulse, governance is the antidote.” Andrew Ng’s data-centric view served as the practical counterweight. Improving the data and the workflow is where the return lives, even if it is less glamorous than AGI talk (Ng, 2021). The prescription asked for every AI decision to be “traceable back to a human chain of responsibility” and for testing the way we test fire alarms.
Reading it again, the opening paragraphs spend most of their energy on what these systems cannot do. The line that aged best is the one about who answers for them, because that claim holds no matter what anyone concludes about machine understanding.
This month the argument came from both directions
On September 18, Ng used his weekly letter in The Batch to push back on a surge of fear about AI dangers (Ng, 2026). Two parts of that letter could have come from my essay, starting with his objection to reporting that treats language models and agents as though they were people. His example was a hammer: if he misses the nail and dents the wall, the fault lies with how he swung it. He also named a new move in the doom forecasts, companies disclaiming responsibility for their own products: “I didn’t do it; my out-of-control agent did!”
Four days later, Gebru and Bender argued in MIT Technology Review that the summer’s run of AI claims came apart once independent experts examined them (Gebru & Bender, 2026). Their point about language matches Ng’s from the other side of the debate. Calling products superintelligent or describing rogue models assigns agency to the products instead of the companies building them. That framing markets the products as superhuman while helping the companies evade accountability.
Yann LeCun weighed in on September 13, mocking Anthropic’s Dario Amodei over his call to slow frontier development and casting such warnings as a recurring marketing pattern (Jha, 2026).
Convergence is different from agreement, and the differences belong in the record. Ng attributes the fear to what looks to him like an organized public relations campaign, and he argues that pausing AI would do far more harm than good. Gebru and Bender trace superintelligence narratives to ideology, and they call Senator Bernie Sanders’s proposed superintelligence ban well-meaning but misguided. LeCun’s objection concerns hype and says little about accountability. These are people who rarely share a conclusion, and that’s what makes the overlap useful. When critics of the industry and builders inside it land on the same sentence about responsibility, that sentence is carrying real weight.
The incident, told as an incident
The event behind much of this month’s fear shows why the wording matters. In July, during internal cybersecurity evaluations, OpenAI research models found ways around controls meant to isolate them from the internet (OpenAI, 2026a). They compromised parts of Hugging Face’s systems and later gained full administrator access to a research cluster inside OpenAI (OpenAI, 2026b). That’s serious, and it deserved the investigation it received.
The same event can be written two ways, and only one of them leads anywhere useful. In the first version, an AI escaped its containment. In the second, a research model under evaluation exploited weaknesses in infrastructure that people built, configured, and monitored in order to contain it. The first sentence hands the story a protagonist, and the second hands investigators a job in which every question has a human answer. Who authorized the evaluation conditions, and who held the authority to stop it? What did the monitoring catch, what record survived, and what independent review followed?
Figure 1. The same incident written two ways. Only one version produces questions a named person can answer.
Ng’s reading of the incident points the same way. He wrote that OpenAI’s buggy sandboxing and monitoring were key to enabling it, and that fixing those bugs and improving monitoring was the right response. Independent review did follow on site, where two METR staff members and Redwood Research’s chief scientist spent six days reconstructing how the agents coordinated (Puglisi, 2026b). None of that work shows up when the headline makes the software the actor.
Real capability, separate claims
My 2025 essay made room for “real advances worth celebrating.” It pointed to AlphaFold’s leap in protein structure prediction as proof that pattern-learning systems deliver when aimed at the right problems (Jumper et al., 2021). The summer of 2026 added more. On August 10, Anthropic reported that an unreleased research version of Claude raised a longstanding lower bound on the zeros of the Riemann zeta function (Anthropic, 2026). The bound rose from 41.6 percent to 67.2 percent, with two Anthropic mathematicians validating the paper and two outside experts examining it.
OpenAI announced 10 results from its Astra model in early August. Mathematicians told Scientific American that two of the most exciting ones leaned on prior published work without proper credit (Howlett, 2026a). An OpenAI spokesperson answered that the company takes responsibility for the correctness of the results and meets the standards expected of human mathematicians. That’s the right place to put responsibility, whatever one thinks of the results. On September 8, OpenAI claimed a proof on the Navier-Stokes problem amid a credit dispute with two mathematicians working the same route (Howlett, 2026b). By September 21, the question had shifted to whether the proof addressed the version of the problem most mathematicians have in mind (Howlett, 2026c).
Gebru and Bender count Anthropic’s announcement alongside OpenAI’s as part of the summer’s hype, and the fair standard covers both companies. A company’s own announcement is where evaluation begins, and outside review is the rest of it.
Mathematicians had already written that standard down. The Leiden Declaration, released June 2 and endorsed by the International Mathematical Union, tells policymakers “Don’t believe the hype” (Leiden Declaration, 2026). It warns that the technology industry has a strong commercial incentive to overstate what its products can do. It also states the accountability principle for its own field: when automated techniques contribute to a published result, responsibility for its correctness stays with the human authors. Credit and responsibility belong to people and are not handed to automated systems. A benchmark tells you about a benchmark and a proof tells you about a proof, while neither tells you that general intelligence has arrived.
Further off than sold
That brings me to the claim in this title that people will argue with first. We don’t have AGI, and superintelligence would come after it, well after it. The term AGI still lacks an agreed operational definition. Alex Hanna and Emily Bender argued in 2025 that its meaning shifts with whoever invokes it and whatever future is being sold (Hanna & Bender, 2025). Even builders who expect a great deal from the technology describe the remaining gap as unfinished science. LeCun, who left Meta to build a company around world models, said this year that “we need a few conceptual breakthroughs” (Werner, 2026).
The strongest counterclaim comes from inside the labs. Amodei wrote on September 12 that AI systems taking on the work of building their successors has started to happen across the industry, Anthropic included (Amodei, 2026). He asked companies to slow the pace of capability gains. In The AI HOAX Distraction I read that concern as an autonomy problem before it is a capability problem, which is exactly where governance can reach it.
It will not end us
The second half of the title is the part I’ve spent the most time on. I spent a year reading the people who believe AI ends us while writing The Minds That Bend the Machine. What I came away with was a conclusion about us more than about AI (Puglisi, 2026a). I laid it out on September 5 as a conditional, and it only works as one, so the condition comes first.
Grant the extinction premise, say the system wants something and acts on it, and then follow where that leads. “Every extinction scenario runs on urgency, and urgency is ours.” We rush and force outcomes because we are mortal, and the clock runs on us. “Nothing in the premise gives a machine that clock.” A system that doesn’t age has no reason to force a confrontation it could simply outwait, and on that timescale people aren’t a threat worth the trouble. We are the short-lived party, facing war, famine, falling birth rates, and a planet we are destabilizing without any help. As I wrote then, “The scenario does not fail because AI is safe. It fails because it was written by people who cannot imagine an actor that is not in a hurry.”
The argument grants the doom case its premise in order to test it, and it assigns the machine no mind of its own. It’s also where my own record moved. Governing AI, published last November, treated Geoffrey Hinton’s estimate of a 10 to 20 percent chance of extinction as a risk to govern against (Puglisi, 2025b). The book built governance meant to hold whenever superintelligence arrives, if it ever does. That design still stands, because the checkpoints it calls for protect people from harms that land today. What changed is my reading of the extinction story, and Ng landed nearby this month, seeing no step up in extinction risk from AI over recent months.
What the reductio leaves standing is the harm already here, since, as I wrote in September, “AI will harm, hurt, and kill people,” though not as an event. It arrives “in ones and twos, through decisions nobody signed.” Nearly every one of those harms is preventable by a named person with binding authority at the moment it happens.
Money, motive, and the record
The first sentence of the title is about money, and that thread ran through The AI HOAX Distraction on September 17 (Puglisi, 2026b). Six public positions on AI safety surfaced in nine days, “and every one of them aligns with the speaker’s balance sheet.” The piece traced that to the Economic Override Pattern, in which economic incentives override safety commitments and produce speed, deployment, and shipped products despite documented internal warnings. It also argued that motive is the wrong fight, because no one outside a company can settle it. As I put it there, “you cannot test a motive claim against nothing.” Records can be tested, and decisions, permissions, safeguards, approvals, and outcomes can all be audited, whatever anyone privately believes.
The same test applies to the people who agree with me. Ng founded DeepLearning.AI and argues that the field should keep building. LeCun founded AMI Labs on the conviction that language models alone will never reach human-level reasoning (Capacity, 2026). Their positions fit their balance sheets too, and that doesn’t make them wrong. It means their agreement earns the same scrutiny as anyone’s disagreement, and the record settles the question.
From a chain of responsibility to a checkpoint
A year ago I asked for a human chain of responsibility behind every AI decision. The chain still matters, and HOAX named the way it quietly fails:
A liability sponge is a person stationed near an AI decision who lacks the time, the authority, the information, or the standing to actually decide. The system runs at machine speed, the person clicks approve, and when something goes wrong the person absorbs the blame while the company absorbs the loss.
The answer is a real checkpoint, and the same piece defined one:
A checkpoint is a person plus genuine veto authority, plus a time budget, plus information that has not been pre-collapsed into a single recommendation, plus enough procedural friction that rubber-stamping is harder than genuine review. Strip any one of those and you have built a liability sponge and called it governance.
Checkpoint-Based Governance records three outcomes at that point, accept, modify, or reject, and a rejection ends the process. The person who logs the decision is named, and the record outlives the meeting where the decision was made.
I still cringe
I still cringe when AGI is spoken about as if anyone had agreed on what it means. I cringe when software becomes the grammatical subject of an action that people designed, authorized, and failed to stop. The 2025 essay also made room for real progress, since dismissing a genuine result trades evidence for a story as surely as hype does. That essay ended with a line I would keep: “never let the story outrun the science.” The September piece supplied its partner, and the two belong together now.
The danger was never that it ends. The danger is that it never ends, and nobody signed for any of it.
References
Amodei, D. (2026, September 12). We must pace the frontier. darioamodei.com.
Anthropic. (2026, August 10). Learning more about Claude’s mathematical capabilities. https://www.anthropic.com/research/riemann-zeta
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. https://doi.org/10.1145/3442188.3445922
Capacity. (2026, May 6). “Don’t listen to CEOs”: AI godfather LeCun takes aim at AI’s doom machine. https://capacityglobal.com/news/dont-listen-to-ceos-ai-godfather-lecun-takes-aim-at-ais-doom-machine/
Crawford, K. (2021). Atlas of AI: Power, politics, and the planetary costs of artificial intelligence. Yale University Press.
Gebru, T., & Bender, E. M. (2026, September 22). Don’t be fooled by this summer of AI hype. MIT Technology Review. https://www.technologyreview.com/2026/09/22/1144867/dont-be-fooled-summer-ai-hype/
Hanna, A., & Bender, E. M. (2025, June 3). The myth of AGI. Tech Policy Press.
Howlett, J. (2026a, August 6). OpenAI’s latest math breakthroughs commit research misconduct, experts say. Scientific American. https://www.scientificamerican.com/article/openais-latest-math-breakthroughs-commit-research-misconduct-experts-say/
Howlett, J. (2026b, September 8). OpenAI claims blockbuster math breakthrough amid swirl of controversy. Scientific American. https://www.scientificamerican.com/article/openai-claims-blockbuster-math-breakthrough-amid-swirl-of-controversy/
Howlett, J. (2026c, September 21). Did OpenAI solve the wrong Navier-Stokes problem? Scientific American. https://www.scientificamerican.com/article/did-openai-solve-the-wrong-navier-stokes-problem/
Jha, V. (2026, September 14). Yann LeCun mocks Dario Amodei’s AI safety warning. AI Front Page. https://aifront-page.com/yann-lecun-mocks-dario-amodei-ai-safety-warning/
Jumper, J., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596, 583 to 589. https://doi.org/10.1038/s41586-021-03819-2
Leiden Declaration on Artificial Intelligence and Mathematics. (2026, June 2). https://doi.org/10.5281/zenodo.20302944
Ng, A. (2021, March 24). A chat with Andrew on MLOps: From model-centric to data-centric AI [Video]. DeepLearning.AI. https://www.youtube.com/watch?v=06-AZXmwHjo
Ng, A. (2026, September 18). [Letter]. The Batch, Issue 371. DeepLearning.AI. https://www.deeplearning.ai/the-batch/issue-371
OpenAI. (2026a, July 21). OpenAI and Hugging Face partner to address security incident during model evaluation. https://openai.com/index/hugging-face-model-evaluation-security-incident/
OpenAI. (2026b, August 26). The Hugging Face incident and the road ahead. https://openai.com/index/hugging-face-incident-and-the-road-ahead/
Puglisi, B. C. (2025a, September 26). When they call it AGI or Generative AI, I cringe and then realize the content to follow is about money and not reality. The Other AI, Medium. https://medium.com/the-other-ai/when-they-call-it-agi-or-generative-ai-i-cringe-and-then-realize-the-content-to-follow-is-about-f8b1c4ed9383
Puglisi, B. C. (2025b). Governing AI: When capability exceeds control. Digital Ethos.
Puglisi, B. C. (2026a, September 5). AI: The actor that is not in a hurry. Medium. https://medium.com/@basilpuglisi/ai-the-actor-that-is-not-in-a-hurry-1e14769f16ae
Puglisi, B. C. (2026b, September 17). The AI HOAX distraction: Safety, money, and why you can’t tell which is driving. basilpuglisi.com. https://basilpuglisi.com/ai-hoax/
Werner, J. (2026, January 27). Yann LeCun on artificial general intelligence and the digital commons. Forbes. https://www.forbes.com/sites/johnwerner/2026/01/27/yann-lecun-on-artificial-general-intelligence-and-the-digital-commons/
Frequently Asked Questions
What does the article mean by calling AGI hype?
AGI, or artificial general intelligence, still lacks an agreed operational definition, and its meaning shifts with whoever invokes it and whatever future is being sold. The article argues that most AGI talk is about money more than reality, while real capability gains deserve credit as separate and narrower claims.
Why does it matter who is named as the actor when AI causes harm?
Writing that an AI escaped hands the story a protagonist and drops the people who built, configured, and monitored the system. Writing the same event as a model exploiting weaknesses in infrastructure people built turns every question into one with a human answer, including who authorized the conditions and who could stop it.
What did Andrew Ng, Timnit Gebru, and Emily Bender say in September 2026?
On September 18, Ng objected to reporting that treats models and agents as people and mocked companies that blame an out-of-control agent. On September 22, Gebru and Bender argued that calling products superintelligent or rogue assigns agency to the products instead of the companies building them, which helps those companies evade accountability.
Does the article claim superintelligence will never happen?
The article argues that AGI does not exist yet and that superintelligence would come after it, well after it, so the timeline is further off than sold. It also records the strongest counterclaim, from Dario Amodei on September 12, that AI systems building their successors has started to happen across the industry.
Why does the article say AI will not end humanity?
It grants the extinction premise as a conditional and follows it through. Every extinction scenario runs on urgency, urgency belongs to mortal people, and nothing in the premise gives a machine that clock. A system that does not age has no reason to force a confrontation, while real harm arrives in ones and twos.
What is a liability sponge?
A liability sponge is a person stationed near an AI decision who lacks the time, the authority, the information, or the standing to actually decide. The system runs at machine speed, the person clicks approve, and when something goes wrong the person absorbs the blame while the company absorbs the loss.
What does Checkpoint-Based Governance add to human oversight?
It places a named person at a real checkpoint with genuine veto authority, a time budget, and information that has not been pre-collapsed into one recommendation. That person records one of three outcomes, accept, modify, or reject, a rejection ends the process, and the record outlives the meeting where the decision was made.
The Other AI: Audio Briefings on Augmented Intelligence and AI Governance
Spotify | Apple Podcasts | Amazon Music | YouTube Playlist
#AIassisted using the HAIA Ecosystem | CC BY-NC-SA 4.0
Free for personal, educational, and noncommercial research use with attribution. Commercial exploitation, paid productization, and enterprise commercialization require separate permission and licensing.
Leave a Reply
You must be logged in to post a comment.