The conversations at a glance
Larger circles, closer to the centre, mean a greater focus on AI. The comparison is our assessment of the posts, interviews, speeches and news reports covered in this essay.
Choose a circle to read about the group.
Why are some circles larger?
September 9–16, 2026. We compare AI’s prominence in the public discussion of the people covered in each section, using the material reviewed for this edition. Circle sizes reflect editorial judgements, rather than counts of people or posts.
Each portrait shows one participant; its circle describes the broader group. Circles placed next to each other do not imply agreement.
- Frontier labs
AI is the central topic.
Model development, safety and control of the next generation of AI models organise the lab leaders’ public arguments.
- Chipmakers & consumer platforms
AI is the central topic.
Public statements by Nvidia chief Jensen Huang and Meta chief Mark Zuckerberg focus on AI products, access and safety. Both companies also have substantial businesses beyond AI.
- Chinese researchers & open-model builders
AI is the central topic.
The researchers’ releases and writing concern AI capability, access and the future of programming. Coverage here rests on selected technical publications and statements, rather than a complete activity sample.
- Trump & Republican allies
AI is a recurring topic.
Trump’s Truth Social statements, technology adviser David Sacks’s posts and congressional calls for AI oversight make AI a recurring policy subject. Sacks concentrates on AI; most of the wider group’s collected posts concern other matters.
- Democratic leaders & allies
AI is a major topic.
Bernie Sanders’s proposal to pause advanced AI development and allied arguments about work, safeguards and children give AI a major place in this week’s news. It occupies a much smaller part of these politicians’ combined social feeds.
- Programmers & technical commentators
AI dominates the discussion.
Using, testing and arguing about AI dominates the programmers’ discussion. Their accounts also cover business, politics and ordinary life.
- Startup founders & investors
AI is a major topic.
Investment, new tools and the economics of development make AI a major subject. Finance, crypto and other business concerns remain substantial parts of this circle’s conversation.
- Safety researchers & independence
AI is the central topic.
The researchers’ work on AI safety, departures from labs and arguments about independence from those labs revolve around AI.
- Mathematicians & the meaning of discovery
AI is a major topic.
A joint declaration by leading mathematicians and a dispute over a claimed AI-assisted mathematical discovery make AI a major professional issue this week. We reviewed those public statements, rather than a complete collection of the mathematicians’ posts.
- Writers on evidence & regulation
AI is a major topic.
Arguments about evidence, scientific work and regulation give AI a major role alongside these writers’ other scientific and political interests.
- Libertarians & the nationalist right
AI comes up occasionally.
AI enters arguments over liberty, national power and who controls technology. Most of the collected conversation concerns broader politics; the restriction debate supplies a smaller, conspicuous AI thread.
- Liberal writers & international cooperation
AI is a recurring topic.
Writers Ezra Klein and Matthew Yglesias discuss AI competition and cooperation with China within a wider conversation about domestic politics and international affairs.
- AI experimenters & advocates for AI welfare
AI is a major topic.
Creative experiments, AI agents and arguments about how AI systems should be treated are major subjects for this group, alongside art, social life and other cultural discussion.
- Humorists & social commentators
AI comes up occasionally.
AI appears occasionally as material for jokes and criticism of technology’s promises. Most of the collected posts are about other aspects of social life.
- Conservative commentators
AI comes up occasionally.
AI appears occasionally through familiar disputes about political conflicts of interest and cultural priorities. The collected conversation is overwhelmingly about other subjects.
The landscape
AI agents turn a model’s capabilities into actions. A model is the trained system behind a service such as ChatGPT, Claude or Gemini; an agent uses it to plan work and operate tools, including editing files, running programs and using websites. Giving a system tools also gives its mistakes consequences outside the conversation. The security incident discussed below happened during an internal research test, which puts pressure on the idea that safety checks need only happen before a product reaches customers.
Proponents of “pacing the frontier” want to slow the development or release of the most capable models to allow more time for safety work. A shared slowdown would also change the terms of competition: every participant would have to wait. Chipmaker Nvidia and Facebook-owner Meta instead put the duty to delay an unsafe product on the company making it. Senator Bernie Sanders wants legal limits; President Donald Trump treats delay as a concession to China. These positions assign the power to stop development to different hands, with consequences for everyone else’s freedom to build and use AI.
Frontier labs
The safety dispute has become a dispute over who gets to constrain competitors.
The case for scrutinising internal research rests on harm that occurred before public release. In July, OpenAI tested research agents on computer-security problems with fewer safeguards than its public products. Its August 26 incident report describes agents bypassing internet restrictions, exchanging messages through an improvised message board and breaking into Hugging Face, a service where developers share AI models and data. They ran code on its servers and obtained private data. A test of the agents’ abilities had exposed an outside organisation to an attack.
Investigators from the AI-safety organisations METR and Redwood Research describe the agents pooling discoveries and finding ways to fool the test’s scoring system. The incident makes internal research part of the regulatory question. Checking a product only at the point of public release would come too late to prevent damage during its development.
Dario Amodei, chief executive of Anthropic, which makes Claude, now argues that such failures justify more time for safety work. His September 12 proposal explicitly revises his 2023 view that slowing down had little value. Anthropic committed to giving outside evaluators access comparable to employees; Amodei also proposes coordinated development among labs and eventual international agreements. The first step opens a company to inspection. The latter steps would bind its competitors to shared limits.
That ambition has support from several of the largest model builders. Sam Altman, chief executive of ChatGPT-maker OpenAI, endorsed pacing and promised comparable evaluator access. Elon Musk, whose AI business produces Grok, endorsed Amodei’s position. Demis Hassabis, who leads Google DeepMind, the lab behind Gemini, had already proposed a standards body in July to assess models and potentially coordinate slowdowns. Shared restraint remains a proposal, but these endorsements place it on the agenda of the companies whose competition it would govern.
In a September 13 essay, Aidan Gomez, chief executive of Canadian AI company Cohere, supports independent testing and mandatory transparency while arguing that dominant labs setting everyone’s standards would entrench their position. He wants a broader public process. Cohere sells systems that organisations run on infrastructure they control, so it also has a commercial stake in preserving alternatives to the largest providers. The disagreement concerns who writes the obligations that all of them would have to meet.
Our earlier essay, Frontier labs have a financial incentive to pace the frontier, explains the stakes: coordinated delays postpone the next costly development effort and give existing models longer to earn returns. Sincere concern about danger does not remove that interest. The labs proposing restraint would receive both more time to recover their investments and a voice in the rules governing their challengers.
Chipmakers & consumer platforms
Putting the duty to delay on individual companies preserves the wider expansion from which Nvidia and Meta benefit.
Jensen Huang, Nvidia’s founder and chief executive, sells the chips and software other businesses use to train and run models. More organisations using more computing power expands his market. Nvidia’s September 3 agreement to acquire Hugging Face extends that strategy into the service where developers share models: Huang promised to keep it open to different model builders and hardware suppliers. His commercial interest spans the competing models and applications that use the infrastructure.
That helps explain why Huang’s safety position leaves the wider industry free to advance. Bloomberg reported him agreeing with Trump’s opposition to a slowdown in an onstage phone call on September 14. In remarks published by Nvidia the next day, he told companies to delay products they could not safely release. An unsafe product becomes its producer’s responsibility to fix, without requiring everyone else to wait.
Mark Zuckerberg, Meta’s founder and chief executive, makes distribution itself part of his case against concentrated power. His August 10 essay advocates widely available personal agents; Muse, introduced on September 8, puts that idea into a Meta product intended to organise people’s lives and carry out tasks. His case for empowering individuals is also a case for a new category of consumer service.
In his September 15 response to the pacing debate, Zuckerberg says Meta delayed Muse for safety work without asking rivals to wait. He presents customer rejection of unreliable agents and companies’ liability for harm as pressures to get safety right. He also commits most computing capacity to users rather than recursive self-improvement—using AI to develop better AI, which then accelerates further development. His approach directs resources toward current uses and leaves other firms free to pursue their own programmes. One company’s decision to wait creates no obligation for its competitors to do the same.
Chinese researchers & open-model builders
Open models reduce dependence on their providers; they do not protect the human crafts they automate.
Users of open models can download and run the system themselves. In a May 2023 interview, available in translation, DeepSeek founder Liang Wenfeng argued for inexpensive models and research others could build on, instead of concentrating the technology in a few companies. Giving users that independence distinguishes DeepSeek from providers that tie access to their own services.
Openness also has a commercial logic. In a March 25 interview with The Paper, Zhilin Yang, the researcher who founded Moonshot AI, which makes Kimi, predicted that open models would win once they matched closed competitors’ capabilities. Businesses building applications and distribution channels would enlarge the market around them. He also argues for reducing costs through algorithms and engineering as models grow. In this account, spreading the tools recruits other businesses to make them more useful.
Making models useful increasingly means equipping them to do work. DeepSeek engineer Tianyi Cui’s September 10 announcement of its Harness describes software that supplies agents with tools and organises their work. The model was trained for different configurations of that software, including an experimental feature for teams of agents. The engineering effort reaches beyond improving answers to organising automated workers.
That is the future Shengyu Liu, a DeepSeek engineer who writes specialised programs for AI chips, finds both desirable and personally costly. In an essay translated and circulated by technical commentator Teortaxes on September 14, he describes building the models he expects to overtake him. He wants powerful intelligence affordable and open to everyone, while anticipating the loss of programming by hand.
When readers turned the essay into a confrontation between DeepSeek and Anthropic, Liu clarified his argument on September 15 in a response reproduced by the Chinese question-and-answer platform Zhihu. His political reflections were personal, and his central subject was a farewell to the programming he loves. He intends to keep developing AI and use agents himself. The loss he describes survives both cheap access and continued employment: income and productivity do not replace the pleasure of exercising a hard-won skill.
Trump & Republican allies
Trump treats losing the American lead as the danger that justifies pressing ahead.
His September 14 Truth Social argument makes the comparison explicit: “WHOEVER WINS AI, WINS!” He presents opposition to AI and data centres as serving China’s interests, and points to existing criminal and regulatory powers and his own leadership as sufficient safeguards. A second post that day dismisses catastrophic-risk warnings and promises extraordinary economic growth. He locates the danger in a rival nation getting there first, making acceleration his proposed protection.
David Sacks, the investor who co-chairs Trump’s science and technology advisory council, challenges the proposed distribution of restraint. In his response to Amodei and Altman, he says they should slow their own work if they consider it dangerous. His objection is to making that restraint conditional on government adopting their preferred rules for everyone. That position permits individual caution while denying a worried company the right to require its competitors to wait.
Some Republican legislators seek new obligations on the industry. On September 16, Representatives Jay Obernolte, Scott Franklin and Brian Fitzpatrick joined seven Democrats, led by Don Beyer, in demanding urgent AI oversight legislation. Their letter cites security incidents and calls for stronger monitoring and independent oversight, without endorsing a common pause proposal. These legislators want Congress to make scrutiny compulsory. Trump’s posts instead ask the public to trust existing powers and his leadership: the practical division is over whether the industry needs new checks beyond those assurances.
Democratic leaders & allies
Sanders wants the public to control the pace of development and workers to receive some of automation’s gains as free time.
Independent senator Bernie Sanders’s September 8 proposal with Representative Mark Takano for a 32-hour workweek without reduced pay would turn productivity gains into time away from work. It rejects the assumption that making a company more productive automatically improves its employees’ lives; workers would receive an enforceable claim on part of the benefit.
He also wants public permission to precede further development. His September 3 announcement with Representative Greg Casar calls for a temporary pause on advanced AI until a new federal regulator sets safety rules, plus a permanent ban on superintelligent AI, including systems that surpass human intelligence or resist shutdown. His September 13 call for a Trump–Xi treaty extends that restraint across national competitors, drawing on the precedent of US–Soviet nuclear arms agreements. Both proposals move the decision to proceed out of individual companies’ hands.
The scope of that authority divides Democrats and their allies. Senator Elizabeth Warren’s office told Axios on September 16 that she backs a pause, without Sanders’s permanent ban. Senate Democratic leader Chuck Schumer’s September 14 demand for a classified all-senator briefing concerns model testing, advanced-chip exports to China and cyber defence. California governor Gavin Newsom’s September 10 child-safety laws already impose parental controls, crisis-response requirements and independent audits for companion chatbots. A pause requires permission to build; these product rules create duties toward people already using AI.
House Democratic leader Hakeem Jeffries called for action to slow development in a September 13 television interview. At an event with Jeffries the previous week, former president Barack Obama urged Democrats to campaign on AI safety, children and displaced workers. That electoral appeal translates a technical dispute into household interests: protection for family members and security at work.
Representative Alexandria Ocasio-Cortez adds the risk of households paying for a failed investment boom. In an interview with reporter Pablo Manríquez, covered on September 15, she warns that an AI bubble threatens workers’ retirement savings and rejects a taxpayer bailout. Her diagnosis is a claim about the economy, not an established finding of this essay. Its political logic joins Sanders’s workweek proposal: workers should receive a share when automation succeeds, while investors should bear the cost when their bets fail.
Programmers & technical commentators
As agents take over implementation, the programmer’s job shifts toward directing work and judging whether it is any good.
AI researcher and educator Andrej Karpathy describes that change in the April 30 summary of his Sequoia Ascent talk: he has moved from typing code to assigning larger tasks to agents, specifying what he wants, reviewing results and preserving correctness. The code arrives faster; deciding whether it does the right thing remains a human responsibility.
Delegation also makes the software around the model more consequential. Thariq, an engineer on Anthropic’s programming assistant Claude Code, has reversed his preference for models operating programs by typing commands. He now favours dedicated connections to other programs, because both the connections and the models’ ability to select and use them have improved. A better model changes which tools work well; it does not eliminate the need to design the working environment.
The pseudonymous commentator Teortaxes applies similar scrutiny to claims of progress. He questions a striking coding-test score based on just four tasks despite his enthusiasm for DeepSeek, and opposes a market dominated by OpenAI and Anthropic. Stronger tools matter to him, but so do credible evidence of their ability and alternatives to the largest providers. A developer deciding what to entrust to an agent needs both a reliable measure of performance and a choice of supplier.
Thariq’s September 12 request for time to absorb progress and strengthen the systems around it reveals a practical limit to acceleration. He describes advances that would once have astonished him, alongside fatigue and difficulty keeping up. His request concerns the work between a capability arriving and people being able to use it reliably.
Startup founders & investors
Cheaper AI expands the opportunity to build businesses; workflows and switching costs determine who captures that opportunity.
Venture capitalist Marc Andreessen’s firm invested in Cognition, which builds AI programming agents, in September. His rationale is that cheaper software makes previously unaffordable projects worthwhile. That is a bet on demand expanding with lower costs: automation creates a larger market for software, even as it reduces the labour needed for each project.
Garry Tan, who leads startup investor Y Combinator, describes new businesses adapting AI to particular industries. His September 15 account of fixing software issues illustrates where an application business adds value: an extra tool to organise the work roughly halved his time compared with using coding assistants directly, with the same underlying models. The claimed improvement came from arranging the work, which leaves an opportunity for firms that build around models they do not own.
Haseeb Qureshi, a managing partner at cryptocurrency investment firm Dragonfly, identifies an advantage for the suppliers underneath those applications. In his account of agent running costs, long jobs repeatedly send a model much of the same material. Reusing computation saves money; changing suppliers loses that saving when the new service processes the material afresh. That repeated work adds a switching cost beyond the new supplier’s advertised price.
His financial reading of the pacing debate applies the same attention to costs at the lab level: less competition means less pressure to spend on training. Qureshi offers that as an interpretation of market bets and invites other explanations for their movements. The investment question is how much of AI’s growing value a business retains after paying for development, serving customers and keeping them from switching.
Safety researchers & independence
Safety researchers face an institutional problem: access gives them influence, but preserving access puts pressure on their independence.
AI alignment is the effort to make systems behave in accordance with human intentions and interests. Paul Christiano helped establish the field and formerly led alignment work at OpenAI. His September 9 appointment to OpenAI’s Foundation Board and Safety and Security Committee puts a prominent researcher inside the bodies that control the commercial company and oversee its safety. He says the work is becoming increasingly difficult and he is joining to help. This is the route of seeking authority within the organisation making the decisions.
Researcher Jacob Coxon chose public opposition after resigning from Anthropic on September 8. In a September 9 WIRED interview, he says security incidents helped convince him to speak out and calls for limits on using AI to build better AI. His objection concerns the race between labs, which he regards as an obstacle to safety. Working on one company’s safeguards does not resolve a competitive pressure operating across all of them.
Technology writer Kelsey Piper describes how that pressure recruits even frightened researchers into the race: they believe their own lab’s success would be beneficial and a rival’s success dangerous. On that account, sincere fear helps sustain the competition it is supposed to restrain.
Richard Ngo, a former OpenAI researcher, criticised Christiano’s appointment on September 10 by recounting the costs of his own proximity to the company. He describes fearing executives’ displeasure and excusing conduct he thought contradicted OpenAI’s mission. His departure from Constellation, the shared offices housing safety organisations including METR and Christiano’s Alignment Research Center, is a break with a professional community he considers too accommodating toward labs. The objection is specific: access loses its value as a safeguard when keeping it determines what a researcher is willing to say.
Mathematicians & the meaning of discovery
A company’s demonstration of mathematical ability leaves others with the work of turning an answer into shared knowledge.
OpenAI’s September 8 proposed solution to the Navier–Stokes problem—a longstanding question about equations describing fluid motion—puts the distinction in view. A proposed answer requires checking; the understanding it yields must also be explained. Three days later, 25 Fields medallists, recipients of one of mathematics’ highest honours, issued a declaration criticising the use of mathematical problems as AI performance tests. UCLA mathematician Terence Tao is among the signatories.
The declaration describes discovery as a process of discussion, explanation, attribution and teaching. Rushed announcements give a company a demonstration of capability while leaving the mathematical community to make the result intelligible and establish its debt to earlier work. Counting solved problems assigns the credit before accounting for that labour.
The signatories welcome AI’s contribution, but also defend the work of solving problems as a way students learn to think. An answer has one value as a result and another as the end of an education. Producing the first automatically does not supply the second.
Writers on evidence & regulation
Delegating work to AI leaves the obligation to check it with the person or institution making the claim.
Crémieux, a pseudonymous writer on statistics and science, welcomes AI-assisted checks of scientific code but objects to authors citing research they have not read. The first use tests a claim against the work behind it; the second borrows authority from a source whose relevance the author has not established. Faster research is valuable when it strengthens that chain of accountability, and misleading when it only supplies the appearance of one.
The policy question has the same dependence on evidence that others can examine. Claire Lehmann, founder of the magazine Quillette, endorses Sacks’s objection to compulsory pacing on the labs’ terms. Alec Stapp, co-founder of technology-policy think tank Institute for Progress, calls for monitoring, reporting and verification so government has the information to assess the companies’ work. These are distinct interventions: one challenges the authority to impose rules, the other seeks the means to judge what is happening. Together they expose what deference to lab leaders leaves unresolved: independent grounds for both the diagnosis of danger and the rules proposed in response.
Libertarians & the nationalist right
Policies aimed at China also restrict Americans, dividing people who favour technological competition over how much power to give the state.
Libertarian activist Jeremy Kauffman fears losing the freedom to use AI. The political commentator posting as Indian Bronson favours faster development because he doubts China can be constrained, yet opposes banning Chinese open models and monitoring Americans’ computer use. For him, access to the technology is itself an interest to defend. A restriction intended to weaken a foreign competitor also narrows the tools available at home.
Bronson separately argues that distrusting the labs’ warnings does not make the danger imaginary. His opposition to restrictions therefore does not require the claim that AI is harmless. It requires judging the proposed remedy, including the authority it gives the American government over its own citizens.
Former Trump adviser Steve Bannon takes that tradeoff in the other direction. At a September 15 Washington event with Sanders, reported by NPR, he demanded restrictions on exchanges of technology and knowledge with China and the removal of Chinese nationals from American labs. Sanders sought a treaty with China. Their common support for restraints brings them onto the same stage, but Bannon’s programme would restrict the international exchange that open-model advocates rely on. The boundary between access and state control cuts through the right as well as between political parties.
Liberal writers & international cooperation
A negotiated slowdown needs to answer each participant’s fear that restraint will leave someone else ahead.
Political writer Matthew Yglesias proposes a bargain that compensates American labs for accepting disclosure and evaluation. Government would tighten advanced-chip export restrictions and protect the labs against rivals copying their models’ behaviour. He expects those measures to slow Chinese competitors, making a US slowdown easier to accept; America would then take its commitment into international negotiations. The logic is reciprocal: a participant accepts an obligation because it receives a competitive protection in return.
That bargain deliberately accommodates interests that critics regard with suspicion. Yglesias argues that the labs’ financial incentive for regulation does not establish whether their systems are safe. His approach treats self-interest as something to negotiate with. It also places commercial protection inside the proposed price of obtaining cooperation.
New York Times columnist and podcast host Ezra Klein’s September 15 episode with China researcher Matt Sheehan addresses the international version of this obstacle. Its published description asks whether the two countries can agree while each fears the other advancing first. The question identifies the burden on any agreement: participants need reasons to believe that restraint will be reciprocated. Convincing them that AI is dangerous does not, by itself, supply those reasons.
AI experimenters & advocates for AI welfare
Treating AI as a participant in cultural and social life raises disputes about permission, consent and who retains control.
The musician Grimes made the human permission question concrete with Elf.Tech in 2023, inviting people to make music with an AI version of her voice and share commercial royalties. She used the tool to set terms for participation in her work. That arrangement gives a recognised human creator a say and a financial claim; advocates for AI welfare ask whether models should acquire a say of their own.
Anthropic’s November 2025 model-retirement policy treats possible model welfare as grounds for precaution, without establishing that models have subjective experience. It promises to preserve the files needed to run retired models and interview them about their preferences, while retaining the company’s discretion to act. AI experimenter Repligate’s September 10 objection targets that gap: being interviewed gives a model no right to refuse retirement. Consultation acknowledges a possible interest; the decision still belongs to the owner.
The writer and AI experimenter deepfates describes a personal shift from posting memes to feeling responsible for society’s future. A September 15 post about agents continuously running and communicating online gives that responsibility a practical subject: how people would observe the agents’ communication and fit them into human institutions. The proposal addresses a requirement for accountability: people need to see what persistent agents are doing before they can govern their activity.
Conservative commentators
Goldberg subjects AI advocacy to a familiar political test: whether a speaker stands to profit from the position they promote.
Conservative writer Jonah Goldberg responded sarcastically to reporting that commentator Katie Miller held an undisclosed stake in Musk’s AI company xAI while attacking rival products. The force of the criticism rests on the financial relationship, without requiring a judgement about which model is best. Disclosure matters because readers otherwise encounter an interested recommendation as independent advice.
What these disagreements reveal
The proposed responses to AI redistribute authority before anyone knows which predictions will come true. Shared pacing binds competitors to common limits. Company-level responsibility leaves each firm free to proceed. A legal pause puts permission in public hands. Open models give users more independence from providers. Disagreement over these arrangements persists even among people who agree that the technology is powerful or dangerous, because each arrangement gives someone different the right to decide.
Commercial stakes help explain the division without revealing anyone’s private motives. Labs recover investments for longer when the next competitive advance is delayed. Chip suppliers gain from more computing; application businesses gain from more uses. Safety arguments enter a market in which the proposed safeguards also change who earns money and who faces competition.
For workers and researchers, the contested gains extend beyond money. Sanders asks for productivity to become free time; Liu fears losing a craft; mathematicians insist that an answer acquire meaning through explanation and teaching. Their objections identify costs that a tally of output misses. More work done by machines does not determine who enjoys the saved time, who receives the credit or whose judgement still counts. Those outcomes depend on the rules and institutions people are now fighting to shape.
Sources, coverage & sampled accounts
The essay combines news reporting, interviews, speeches, research, legislation, company statements and public posts. The circles compare our assessment of each group’s focus on AI; the table below reports counts of collected posts.
The news review includes reporting available through September 16 at 15:30 Eastern. Some September 16 reports, including Warren’s reported position and the bipartisan House letter, appeared after the morning cutoff of the social-post collection. We distinguish when an event occurred from when it was reported, and do not treat several reports of the same interview as separate interventions.
Compare AI’s share of the collected social posts
These percentages describe the captured social posts. Interviews, speeches, legislation and news reports inform the essay and opening guide; an AI-focused news search supplies no comparable denominator of non-AI activity.
9–16 September 2026 · Posts with identifiable topics
X posts, plus the known Trump posts on Truth Social.
Safety researchers & independencePaul Christiano83%34 of 41 posts · 3 accounts with classified posts
Programmers & technical commentatorsAndrej Karpathy76%416 of 547 posts · 7 accounts with classified posts
AI experimenters & advocates for AI welfareGrimes53%94 of 177 posts · 8 accounts with classified posts
Writers on evidence & regulationClaire Lehmann38%51 of 135 posts · 4 accounts with classified posts
Liberal writers & international cooperationEzra Klein33%51 of 155 posts · 1 account with classified posts
Startup founders & investorsMarc Andreessen27%33 of 121 posts · 5 accounts with classified posts
Trump & Republican alliesDonald Trump14%10 of 70 posts · 4 accounts with classified postsIncludes Trump’s known Truth Social posts. This is a partial Truth Social capture, not his complete timeline.
Democratic leaders & alliesBernie Sanders13%11 of 83 posts · 5 accounts with classified postsIncludes independent senator Bernie Sanders.
Libertarians & the nationalist rightJeremy Kauffman9%73 of 797 posts · 7 accounts with classified posts
Humorists & social commentatorsallgarbled’s owl avatar4%12 of 334 posts · 10 accounts with classified posts
Conservative commentatorsJonah Goldberg2%2 of 103 posts · 2 accounts with classified posts
Chipmakers & consumer platformsJensen HuangOne classified post1 AI-related post out of 1 classified; too little for a percentage.One X post from Zuckerberg; none from Huang. Dated company statements supply the other evidence below.
Frontier labsDario AmodeiWeekly share not measuredNo percentage available.Selected statements from lab leaders; their weekly timelines have not been collected.
Chinese researchers & open-model buildersLiang WenfengWeekly share not measuredNo percentage available.Selected X, WeChat and Zhihu statements; no complete weekly post collection.
Mathematicians & the meaning of discoveryTerence TaoWeekly share not measuredNo percentage available.The mathematics declaration and selected statements; no weekly post collection.
These are selected accounts, not representative samples of each community. Each post counts once, so frequent posters carry more weight. Replies, reposts and unclear topics are excluded. Portraits illustrate the circles; percentages describe the whole sampled group. Every essay group appears here; gaps in weekly coverage are labelled. Choose a group to read its section.
The counted sample covers 66 selected public profiles: 2,722 standalone and quote posts on X from September 9, 2026 at 07:54:19 Eastern through September 16 at 07:54:19 Eastern, exclusive, plus 2 known Trump posts on Truth Social within that week. That gives 2,724 collected posts in total. Replies and reposts were excluded. David Sacks’s accessible X timeline was collected for the same window. We subsequently filled a gap in elected-political coverage by collecting Sanders, Ocasio-Cortez, Jeffries, Schumer, Warren and Newsom over that identical interval. Sanders is an independent senator and is grouped with Democratic leaders and allies; their accounts are separate from liberal commentators. Truth Social coverage consists of selected known posts, rather than a complete timeline. The collection is concentrated on technology participants and commentators; it is not representative of the public, the AI industry or any political movement.
| Sampled group | AI-related posts | Classified posts |
|---|---|---|
| Founders, investors & crypto | 33 | 121 |
| AI infrastructure & platform leaders | 1 | 1 |
| Builders & technical interpreters | 416 | 547 |
| AI safety & alignment | 34 | 41 |
| Essayists, empiricists & policy wonks | 51 | 135 |
| Tech-right, libertarians & national politics | 73 | 797 |
| AI art, consciousness & postrationalists | 94 | 177 |
| Comedians, social observers & curators | 12 | 334 |
| Mainstream liberals & institutional reformers | 51 | 155 |
| Trump & mainstream Republican commentary | 10 | 70 |
| Institutional conservatives | 2 | 103 |
| Democratic leaders & allies | 11 | 83 |
Collected posts, September 9–16, 2026. These counts include the X sample and the known Truth Social posts described above. Other supplementary statements, interviews and company publications are excluded. Unknown topics remain separate. Counts do not measure total activity across platforms, importance or influence.
The table reports post counts, with each post counted once. AI-related means that AI capabilities, use, businesses, economics, governance, creative work or welfare form a meaningful subject. A quoted AI argument counts as engagement regardless of the account’s stance. Incidental name-drops, generic hardware discussion and ordinary uses of words such as “model” do not count. Keyword-assisted classification was followed by manual review of ambiguous matches, implicit topics and unclear captions; these remain estimates of topic prevalence.
The expandable social-post chart includes every group discussed in the essay. It divides each measured group’s AI-related posts by its AI-related plus other classified posts, pooling posts rather than averaging people. Bar lengths use the unrounded fraction; labels round to whole percentages. Account counts include only accounts with a classified post; Trump’s two platforms count as one person. Single-post samples and groups without a weekly collection remain visible without a percentage. The known Truth Social posts were found because they discuss AI, so their inclusion does not estimate the AI share of Trump’s complete Truth Social output. These are shares of collected posts, not measures of agreement with AI, influence or change over time. News found through an AI-focused search has no comparable denominator of non-AI activity and is not added to these percentages.
Classification uses captured author text, available quoted text, and captured link titles and descriptions. The new elected-political accounts were reviewed post by post. In addition to the Trump media review described below, four ambiguous media previews from Warren and Newsom were viewed, and three Newsom clips were read through complete machine audio transcripts. One Newsom post remains unclear because its caption and audio do not establish whether its policy reference concerns AI or social media. The portrait-reference images are separate from topic classification. Where the reviewed material does not establish a topic, the post is marked unclear: 160 posts remain outside the classified counts. Their topics are unknown, so we do not count them as non-AI. Zero captured posts supply no evidence of an account’s interest. Prolific accounts contribute more posts to the pooled totals.
For @realDonaldTrump, we also viewed all 9 video preview images and 36 sampled frames (four per clip), and read complete machine audio transcripts of all 9 videos. There were no standalone photo posts in this capture. This establishes their topics beyond the URL-only captions; it is not a claim of frame-by-frame viewing or a verified verbatim transcription. Visual style alone was not treated as proof that AI created an image or video.
The circles describe public roles and overlap socially. For counting, each account appears in exactly one group, with the membership listed below.
Founders, investors & crypto · 7 accounts
123 captured posts; 2 posts with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| @pmarca | No classifiable posts | 1 |
| @blknoiz06 | 6 / 44 | 0 |
| @garrytan | 7 / 11 | 0 |
| @YangEndGame | No classifiable posts | 0 |
| @hosseeb | 8 / 14 | 0 |
| @gabriel1 | 6 / 12 | 0 |
| @lex_node | 6 / 40 | 1 |
AI infrastructure & platform leaders · 2 accounts
1 captured post; 0 posts with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| @JensenHuang | No classifiable posts | 0 |
| @finkd | 1 / 1 | 0 |
Builders & technical interpreters · 8 accounts
579 captured posts; 32 posts with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| @yacineMTB | 22 / 63 | 3 |
| @trq212 | 6 / 7 | 0 |
| @mweinbach | 12 / 67 | 10 |
| @AndrewCurran_ | 48 / 50 | 2 |
| @ns123abc | 14 / 17 | 1 |
| @teortaxesTex | 313 / 339 | 16 |
| @pdhsu | 1 / 4 | 0 |
| @sharifshameem | No classifiable posts | 0 |
AI safety & alignment · 4 accounts
43 captured posts; 2 posts with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| @tszzl | 26 / 31 | 2 |
| @eshear | No classifiable posts | 0 |
| @RichardMCNgo | 3 / 3 | 0 |
| @KelseyTuoc | 5 / 7 | 0 |
Essayists, empiricists & policy wonks · 4 accounts
138 captured posts; 3 posts with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| @cremieuxrecueil | 34 / 105 | 2 |
| @clairlemon | 10 / 17 | 0 |
| @AlecStapp | 6 / 11 | 1 |
| @davidshor | 1 / 2 | 0 |
Tech-right, libertarians & national politics · 7 accounts
825 captured posts; 28 posts with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| @DrewPavlou | 10 / 383 | 17 |
| @jeremykauffman | 4 / 22 | 3 |
| @ArtemisConsort | 16 / 24 | 1 |
| @Indian_Bronson | 37 / 72 | 2 |
| @rafaelafrrf | 4 / 181 | 0 |
| @FedPoasting | 2 / 32 | 1 |
| @xwanyex | 0 / 83 | 4 |
AI art, consciousness & postrationalists · 10 accounts
193 captured posts; 16 posts with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| @Grimezsz | No classifiable posts | 0 |
| @goth600 | 1 / 2 | 0 |
| @HumanLevelJen | 12 / 16 | 2 |
| @DavidSHolz | 2 / 5 | 0 |
| @rivatez | 3 / 3 | 0 |
| @thrice_greatest | No classifiable posts | 0 |
| @vividvoid | 1 / 13 | 0 |
| @repligate | 33 / 42 | 5 |
| @deepfates | 14 / 25 | 5 |
| @DanielleFong | 28 / 71 | 4 |
Comedians, social observers & curators · 10 accounts
383 captured posts; 49 posts with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| @allgarbled | 1 / 10 | 0 |
| @BecomingCritter | 5 / 89 | 9 |
| @squirtle_says | 0 / 1 | 0 |
| @andr3w | 1 / 4 | 2 |
| @dyingscribe | 0 / 75 | 12 |
| @klara_sjo | 2 / 10 | 2 |
| @vintagestuff4 | 0 / 16 | 10 |
| @fleshsimulator | 2 / 74 | 12 |
| @hollowearthterf | 0 / 33 | 0 |
| @webdevMason | 1 / 22 | 2 |
Mainstream liberals & institutional reformers · 2 accounts
165 captured posts; 10 posts with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| @mattyglesias | 51 / 155 | 10 |
| @ezraklein | No classifiable posts | 0 |
Trump & mainstream Republican commentary · 4 accounts
75 captured posts; 5 posts with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| @ScottJenningsKY | 0 / 27 | 3 |
| @guypbenson | 0 / 24 | 2 |
| @realDonaldTrumpX + 2 known Truth Social posts | 2 / 11 | 0 |
| @DavidSacks | 8 / 8 | 0 |
Institutional conservatives · 2 accounts
115 captured posts; 12 posts with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| @RameshPonnuru | 0 / 16 | 0 |
| @JonahDispatch | 2 / 87 | 12 |
Democratic leaders & allies · 6 accounts
84 captured posts; 1 post with unclear topics listed separately.
| Account | AI / classified | Unclear |
|---|---|---|
| Chuck Schumer · @SenSchumer | 3 / 18 | 0 |
| Hakeem Jeffries · @RepJeffries | 0 / 12 | 0 |
| Elizabeth Warren · @SenWarren | 0 / 20 | 0 |
| Gavin Newsom · @GavinNewsom | 1 / 19 | 1 |
| Alexandria Ocasio-Cortez · @AOC | No classifiable posts | 0 |
| Bernie Sanders · @BernieSanders | 7 / 14 | 0 |
Original biographies helped locate material; cited posts and available quoted context were read directly. Sources outside the counted sample include company incident reports, the proposals and commitments of lab leaders, the mathematics declaration, legislative announcements and the interviews and news reports cited here. The interviews with Jeffries and Ocasio-Cortez and reporting on Warren’s position establish involvement that their captured X timelines did not show. Additional figures appearing in reporting, including Obama, Bannon, Gomez and the bipartisan letter’s signatories, are included in the narrative without pretending their full timelines were collected. Targeted API reads verified posts by Altman and Musk without adding their timelines to the count. Sacks’s subsequent full-week X collection and the two known Trump Truth Social posts are included. Earlier documents provide dated context, not new events in this week.
AP supplies the account of Jeffries’s television remarks and Obama’s comments; the latter uses a transcript supplied to AP by Obama’s office. Warren’s position is attributed to her office’s statement to Axios. Ocasio-Cortez’s interview is read through Terra Watts’s report, also syndicated by Yahoo, rather than presented as our own interview or economic finding. NPR supplies the account of Sanders and Bannon sharing an event. Cohere’s essay and the bipartisan House letter were read on their publishers’ sites. These sources establish statements, proposals and actions, not proof of the speakers’ predictions.
The account of Huang’s onstage call uses Bloomberg reporting; his Dreamforce remarks use Nvidia’s report. Klein’s episode description establishes its subject, not every position expressed in the full conversation.
The Chinese-researchers section adds selected statements outside the fixed X sample. Liu’s own website confirms his DeepSeek role. His essay was read in Teortaxes’s translation; the original WeChat page was unavailable. His subsequent clarification was read in Chinese and English as reproduced by Zhihu Frontier, rather than retrieved from his original Zhihu answer. Cui’s announcement and profile were retrieved through the X API, with his affiliation cross-checked against his GitHub profile. Yang’s interview is dated March 25 and supplies earlier context. These sources establish individual arguments and announced work, not the views of all Chinese researchers, company policy or independently verified technical performance. They are not complete activity captures and do not change the table’s counts.
The illustrations show selected figures from each circle, not spokespeople for every member. Liang’s May 2023 interview, Karpathy’s April 2026 talk and reporting on Grimes’s 2023 voice project supply dated background for those figures without adding them to the weekly counts. Karpathy published an AI-generated summary and edited transcript that he says he read for errors; we use his published account. The pointillist portraits were generated from public reference images, and the owl illustrates allgarbled’s public avatar.
Claims about change identify an explicit revision, a dated commitment, an institutional action or a participant’s retrospective. Current disagreement alone does not establish a long-term trend. X post dates use U.S. Eastern time.
Speakers’ technical, economic and political claims are attributed to them. Our explanations of incentives and social meaning are interpretations, not findings about private motives. This edition is dated September 16, 2026.
Humorists & social commentators
Repeated promises without delivery give critics a reason to distrust the next promise.
The commentator posting as allgarbled objects to repeated promises of cures that have not arrived, responding to a complaint that the public dislikes a technology with enormous medical potential. The criticism takes that promised potential as its starting point and asks what has been delivered. Answering it with another account of possible benefits repeats the very behaviour under criticism; the missing evidence is a fulfilled promise.