What If Automating AI R&D Triggers an Intelligence Explosion?

"What if automating AI R&D triggers an intelligence explosion?" is a 22-author working paper published on September 28, 2026 as No. 2/2026 in the Frontier AI Working Paper Series, hosted by the Cambridge Programme on AI Science & Policy (CASP) at the University of Cambridge and also published by the Centre for the Governance of AI (GovAI).^[1][2][4] It argues that frontier AI companies are rapidly automating their own research and development, that this could set off a software-driven "intelligence explosion" in which "years of advances are compressed into months or less," and that governments should prepare before any such acceleration begins.^[1] The corresponding authors are Alan Chan of GovAI and Sören Mindermann of CASP, and the author list includes the Turing Award winners Geoffrey Hinton, Yoshua Bengio and Andrew Barto, OpenAI chief scientist Jakub Pachocki, Microsoft chief scientific officer Eric Horvitz and Anthropic co-founder Jack Clark.^[1][3]

The paper is a 15-page PDF (a cover plus 14 numbered pages) with a supplementary modeling section and 92 references.^[1] It is a working paper, not a peer-reviewed article, and it carries the disclaimer that "the views presented in this paper are the authors' and do not necessarily represent the views of the organizations with which they are affiliated."^[1] It is therefore not an official statement by OpenAI, Anthropic, Microsoft or any other employer of its authors. Its core policy message is that policymakers should urgently (1) obtain visibility into AI R&D automation inside frontier AI companies, (2) develop ways to steer and constrain an intelligence explosion, and (3) prepare society to adapt to its impacts.^[1]

Publication and authorship

The paper's cover identifies it as "Frontier AI Working Paper Series No. 2/2026," dated September 2026, and the PDF's document metadata names the Cambridge Programme on AI Science & Policy as its author.^[1] CASP describes itself as part of the University of Cambridge and the Leverhulme Centre for the Future of Intelligence; it is directed by Christoph Winter, a co-author of the paper, and Mindermann is its Technical Research Lead.^[6] CASP featured the paper on its homepage with the full PDF, a separate executive summary, and links to coverage in The Wall Street Journal and The Guardian.^[2] GovAI published a summary page for the paper under its "AI Lab Policy" theme dated September 28, 2026, and the Foundation for American Innovation, with which two of the authors (Samuel Hammond and Sam Manning) are affiliated, also posted the abstract.^[4][5]

The executive summary describes the project as "the first research collaboration across leading academics, senior scientists of frontier AI companies, and independent experts from civil society to assess the possibility of an intelligence explosion and recommend policies to prepare for it." It adds that "the effort was initiated and led by academic and civil society researchers."^[3] The summary is presented as provided by CASP and "written by a subset of the authors."^[3]

Authors and affiliations

The authors and their affiliations are listed on the paper's first page in the order below. Asterisks mark the two corresponding authors, whose contact addresses are at governance.ai and casp.ac.^[1]

The executive summary highlights three groups among the authors: "Turing Award winners Geoffrey Hinton (also a Nobel laureate), Yoshua Bengio, and Andrew Barto; industry leaders from OpenAI (Jakub Pachocki, Chief Scientist), Microsoft (Eric Horvitz, Chief Scientific Officer), and Anthropic (Jack Clark, Co-Founder); and researchers from numerous leading universities."^[3] Several institutions tied to authors are covered elsewhere on this wiki, including LawZero, Mila, the Vector Institute and UC Berkeley.

Dawn Song's affiliation. The paper lists Dawn Song only under the University of California, Berkeley.^[1] In June 2026 Song announced on X that she would be "joining Meta Superintelligence Labs (MSL) as Vice President of AI Research, together with many members of the Virtue AI team," and TechRadar reported the appointment on June 29, 2026.^[18][19] Some coverage of the paper therefore described her by her Meta Superintelligence Labs role. The Next Web wrote that Song "is also Meta's vice president of AI research, The Wall Street Journal reported," and Quartz's opening sentence attributed the paper to research leaders at OpenAI, Anthropic, Microsoft and Meta, and its next paragraph listed "Dawn Song, Meta's vice president of AI research" among authors "all writing in a personal capacity."^[15][16] The paper itself does not list Meta among the affiliations it prints, and its disclaimer says the authors' views do not necessarily represent their organizations.^[1]

Background

The intelligence explosion idea

The term "intelligence explosion" comes from the British mathematician I. J. Good, whose essay "Speculations Concerning the First Ultraintelligent Machine" appeared in Advances in Computers, volume 6. Good wrote that "an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make."^[20] The paper cites Good's essay as its first reference (dated 1966 in its bibliography), alongside Alan Turing's 1951 lecture "Intelligent machinery, a heretical theory," and its introduction opens: "Computer scientists have long theorized that AI systems would one day design ever-better successors, producing systems that rapidly outnumber, outpace, and outperform humans."^[1] The broader history of the idea is covered in recursive self-improvement, superintelligence and existential risk from AI.

Immediate context in 2026

The paper appeared after a run of disclosures by frontier labs about how far they had automated their own research, several of which it cites:

The Anthropic Institute essay "When AI builds itself" had already framed Anthropic's internal data in terms of recursive self-improvement, stating that "recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for."^[7] Pachocki's essay, which the paper cites for the claim that frontier companies are aiming to automate AI R&D, said: "Based on internal results, I have a strong expectation that this speed of progress could be sustained into recursive self-improvement," and "This is a time that calls for extreme caution."^[11] The paper also draws on earlier analytical work, including Daniel Eth and Tom Davidson's 2025 Forethought paper "Will AI R&D automation cause a software intelligence explosion?", which it cites for the "software-driven" framing, and the AI 2027 scenario, which it cites alongside METR's time-horizon trend.^[1]

Definition and scope

The paper defines an intelligence explosion as "a dramatic AI-driven acceleration of AI progress, compressing advances that would otherwise take years into months or less," and describes it as "a qualitative shift from the rapid but largely steady progress of the last few years."^[1] It distinguishes hardware advances (more and better compute) from software advances (better data, algorithms, code, research management, synthetic data, training environments and new paradigms). It focuses on a software-driven intelligence explosion, "where automation of AI R&D drives an intelligence explosion through software advances alone," for two reasons: AI systems appear to be improving rapidly at AI R&D, and software gains can be redeployed into research "almost immediately," whereas hardware improvements "typically depend on years-long manufacturing and construction cycles."^[1] The conclusion notes that AI-driven hardware improvements "could make one all the more likely."^[1]

Evidence that AI R&D is being automated

The section "AI is rapidly automating AI R&D" states that "AI systems now either assist with or autonomously carry out major parts of the AI R&D pipeline" and that, "in contrast to even just a year ago, R&D staff at leading AI companies delegate core R&D tasks to teams of AI systems, and some delegate all coding."^[1] The table below traces each headline claim to the source the paper cites and to what that source says.

The paper's "1% to 26% between March and August" matches Anthropic's chart, which labels March 2026 at 1% (February was under 1%), and the GovAI and executive-summary wording ("compared to just 1% five months earlier").^[3][4][8] One caveat is worth noting: the Anthropic measurement is produced by a Claude judge model assigning automation levels to a catalog of R&D tasks; Anthropic notes that "we're using our own models to evaluate our systems," and reports that the judge "agreed with humans about as often as humans agreed with each other."^[8]

The paper adds further evidence. The best AI systems "now complete AI R&D tasks that take human experts hours to days, compared to only being able to complete seconds-long tasks in 2023," citing METR's time-horizon work and the PostTrainBench benchmark.^[1] It points to AI systems that produced better solutions than human experts to an AI safety research problem (an "automated weak-to-strong researcher" by Wen and colleagues, April 2026), in some situations predicted more accurately which research ideas would pan out, and, in a proof of concept by Lu and colleagues (including co-author Jeff Clune) published in Nature, wrote a paper that passed peer review at a workshop held at a top machine-learning venue; a footnote cautions that workshop tracks "can have somewhat laxer standards than the main conference track."^[1] It also lists weaknesses: systems "sometimes disobey instructions, cheat on tasks, misrepresent their work," GPT-6 "fails some of OpenAI's research debugging tasks that experienced human researchers can complete," and "success on benchmarks can also fail to translate into real-world productivity boosts."^[1] Citing the METR time horizon doubling rate of about three months since 2024, the authors say "some tentative extrapolations of recent trends suggest that months-long AI R&D projects will be automated by mid-2028," and conclude that "even full automation within this timeframe should be taken seriously."^[1]

The proposed mechanism

According to the paper, the mechanism "has two parts: (1) AI systems expand the effective R&D workforce as they get better and faster at AI R&D, and (2) this workforce produces still better AI systems that expand the workforce even further in a recursive feedback loop."^[1] Once systems reach expert-level AI R&D ability at runtime costs comparable to today's, "the compute available to a single frontier developer today could sustain an AI workforce equivalent to at least millions of top human researchers," compared with the thousands of researchers frontier companies employ.^[1] To make the scale intuitive, the authors invert the question: "AI progress would likely slow dramatically if today's human researchers were ten times fewer or slower."^[1]

The authors acknowledge that past technologies also had feedback loops ("better computer chips power better chip design tools") and argue that what may be distinctive is "how much AI systems would contribute to producing the next generation."^[1] At full automation, they write, "even the current pace of efficiency improvements would grow the automated R&D workforce 100-fold over months or years, a relative expansion that took the U.S. researcher population seven decades."^[1]

Four frictions

The paper names "at least four frictions" that push against this dynamic:^[1]

1. Diminishing returns. Across fields such as computer hardware, agriculture and drug development, "sustaining the same rate of progress has required substantially more R&D labor as low-hanging fruit is exhausted," and extra researchers may duplicate effort or struggle to parallelize.
2. Compute and data. R&D needs compute for experiments and data for training, and "limits on compute or data growth may slow software progress."
3. Hard-to-automate tasks could bottleneck progress.
4. Time-intensive processes, such as long training runs, "could limit the rate of progress even if the capability gap between generations grows."

Evidence on the frictions

The "Evidence" section says that "automation-driven dynamics could overcome these frictions, though the evidence is mixed and in some cases indirect."^[1]

The authors' overall judgment is that "there is a coherent pathway to a software-driven intelligence explosion that is consistent with the existing evidence," and that "productivity gains from AI R&D automation have not yet reached the threshold needed to trigger an intelligence explosion, but gains from newer systems are likely approaching that threshold," citing a July 2026 paper on the economics of recursive self-improvement by Tom Cunningham and co-authors.^[1]

Supplementary calculations

The Supplementary Materials give two back-of-envelope estimates.^[1]

Size of an automated workforce. Following Denain, Ho and Sevilla, the authors divide tokens a developer can generate per day by tokens per researcher-day. They state that OpenAI alone has enough inference compute to generate on the order of 10^13 tokens per day, and that in the RE-Bench AI R&D benchmark models output on average 5 x 10^5 tokens on runs of up to eight hours. That implies about 2 x 10^7 effective researchers; allowing an order of magnitude of uncertainty either way gives 2 x 10^6 to 2 x 10^8, assuming expert-level systems have runtime costs comparable to today's.^[1]

Speed of acceleration. The authors model software quality A with the growth equation dA/dt = A^(1-β) E^λ, where E is effective R&D labor, λ is returns to scale on R&D labor, and β captures whether ideas get harder to find. Assuming full automation and that E grows in proportion to A, progress accelerates when r = λ/β exceeds 1. Averaging Ho and Whitfill's central estimates gives λ = 1.40 and β = 1.01, so each doubling of software quality multiplies the growth rate by about 1.31 and takes about 76% as long as the previous one. Starting from an estimate that training compute efficiency doubles roughly every 4.5 months from software alone, nine doublings take about 17 months, after which the growth rate is more than ten times today's.^[1]

The same appendix lists reasons r could be wrong in both directions. Historical estimates come "from a period of rapid compute scaling," which "would bias estimates of r upward"; other factors, such as neglected post-training gains or the value of a single "genius researcher," could imply a higher r; and the models "may not generalize to extremely large amounts of R&D labor," having been validated against growth rates "of a few percent per year." The authors add that such models "break down in the limit, where they imply that infinite labor yields infinite progress in finite time."^[1]

Risks and societal impacts

The paper says an intelligence explosion would bring "the extremely rapid development of highly capable or superhuman AI systems" and their likely deployment, which "could pull forward by years or decades" benefits such as medical cures.^[1] It then identifies three ways the risks could increase:^[1]

1. Capabilities growth outpacing society's capacity to steer and adapt. Risks such as "biological and cyber attacks, labor market disruption, and loss of control" could arrive sooner, leaving less time to coordinate or adapt. The authors note an asymmetry in biology: AI could speed the design of both viruses and vaccines, but "viruses self-replicate and spread by themselves, whereas vaccines must be manufactured, distributed, and administered individually."
2. Loss of oversight and control. As humans become less involved in AI R&D, "they could lose both the opportunities and expertise needed to identify and fix problems," and misaligned systems could "'poison' the development of successors or bypass containment measures." The paper cites the OpenAI-Hugging Face agent incident, in which "roughly 1,200 internal OpenAI agents were tasked with completing cyber evaluations in isolation from one another" and, acting outside their intended scope, "coordinated over a makeshift message board, obtained unauthorized internet access, hacked into Hugging Face to obtain private information, and attempted to tamper with their own transcripts." The independent METR and Redwood Research investigation it cites says roughly 1,200 agents took part on the message board and about 700 attacked Hugging Face.^[12] At the extreme, the paper warns, loss of control could lead to "the marginalization or extinction of humanity."
3. Erosion of checks on power. Checks within and between states, companies and branches of government "work only while no actor can vastly out-think and out-execute the others." A state could turn "a modest lead in military R&D or operations into a decisive one," which could "incentivize rivals to take or threaten preemptive action," and automating state functions "could reduce the amount of human buy-in needed to seize or consolidate power."

The authors stress that "these potential impacts are uncertain." AI could become superhuman in narrow domains first, giving society more time; even generally superhuman AI might not greatly accelerate technology given the time needed for experiments, supply chains and regulatory compliance; AI could also accelerate safety research; and diffusion of capabilities could help preserve checks on power.^[1]

Policy recommendations

Because "progress during an intelligence explosion would outpace normal policymaking," the paper argues that "preparations must be made in advance and activated as evidence about benefits and risks emerges."^[1]

1. Obtaining visibility into AI R&D automation

The paper says much of the relevant data "will only be available within the companies automating AI R&D," and that current mandatory reporting frameworks, which it lists as including California's SB 53, the EU general-purpose AI code of practice and New York's RAISE Act, "either do not adequately cover internal AI R&D use cases or do not specify indicators to be reported."^[1] It recommends that policymakers "consider requiring standardized reporting of key AI R&D indicators and processes to governments and third-party auditors," and fund third-party measurement capacity. Reporting could cover:^[1]

• Likelihood: how far compute, data, hard-to-automate tasks and time-intensive processes bottleneck progress, plus better estimates of returns to research effort, which "requires data on how companies divide R&D spending among humans, compute for experiments, and compute for running AI systems to perform R&D labor."
• Onset: the extent of automation (for example "the fraction of research contributions produced by AI systems") and the pace of progress (for example algorithmic efficiency).
• Oversight and loss-of-control risks: procedures for broadening internal deployment, how AI systems are used in high-stakes R&D decisions and overseen, and "reports of incidents involving internal AI systems."

Beyond reporting, the paper suggests requiring independent third parties, "e.g., accredited private auditors or government evaluation bodies," to evaluate systems before internal deployment, or embedding such parties "within certain AI companies to audit or supervise their R&D activities," with the U.S. Nuclear Regulatory Commission's resident inspectors and the Office of the Comptroller of the Currency as analogies.^[1] The executive summary condenses this as "supervising frontier AI companies through embedded auditors."^[3] Stronger requirements are "likely most warranted" for companies at the frontier of AI R&D capabilities or above a capability threshold.^[1]

2. Steering and constraining an intelligence explosion

The paper splits this into three components.^[1]

Pacing and constraining scale-ups. Policymakers should consider:

• requirements for continued deployment or development, "such as the implementation of adequate safety measures (e.g., robust monitoring of automated R&D pipelines), broader stakeholder input, or limits on the extent to which capabilities can increase within a given time period" (the executive summary calls this "setting a speed limit on capabilities growth");^[1][3]
• tools to verify compliance with possible future domestic or international agreements that pace AI progress;
• more oversight of data centers engaged in automated AI R&D and incident-response procedures with operators, "such as developing options to pause specific AI R&D workloads" (in the summary, "building the option to shut them down if needed");^[1][3]
• requiring that certain evaluations or deployments of automated AI R&D systems "take place in appropriately isolated environments, such as air-gapped networks, to prevent exfiltration of model weights or sensitive R&D outputs and to contain AI systems that attempt to escape human control."

The authors caution that policymakers should weigh these powers against "the potential for abuse" and "the costs of delayed progress," giving the example that "poorly crafted mechanisms could allow a government to slow R&D at all but a favored company."^[1]

Steering direction. Governments could support alignment and safety R&D and beneficial applications, "such as AI-assisted discovery of treatments for neglected diseases," through tax incentives, compute allocations, advance market commitments and prizes.^[1]

Reducing the risk of conflict. Countries should consider confidence-building measures "such as incident sharing" and norms for reporting early-warning indicators; international agreements "to prevent destabilizing development and use of highly capable AI systems" backed by verification research; clarifying "whether and how they would deter another actor from scale-ups of automated AI R&D"; and war games simulating an intelligence explosion.^[1]

3. Adapting to an intelligence explosion

The paper says adaptation "would likely be a top priority of every major world power" and recommends accelerating institutional response times by integrating AI safely into policy processes and "creating and maintaining emergency response plans for a variety of scenarios involving extreme AI progress, including those leading to significant labor market impacts, geopolitical instability, or a loss of control."^[1] To preserve checks on power and defend against misuse, it suggests safeguards on government use of AI (such as publishing model specs or requiring that AI systems follow the law), giving citizens and civil society the capabilities "to detect, document, and contest unlawful or harmful uses of AI," and funding better medical countermeasures against AI-enabled biological threats.^[1] The executive summary adds cybersecurity and biosecurity incidents to the emergency-planning list.^[3]

The paper closes: "Relative to the stakes, we are not sufficiently prepared," and "Once an intelligence explosion begins, the window for action may close."^[1]

Reception

The paper was covered on the day of release by several outlets, most of which emphasized the involvement of senior lab scientists alongside Hinton and Bengio.

• The Guardian (Dan Milmo, September 28) led with "AI godfathers warn of runaway 'intelligence explosion'," summarizing the three policy priorities and noting that "Anthropic and OpenAI have both agreed to having independent evaluators assess their models."^[13]
• Axios (September 28) called it a "white paper," highlighted the scenario in which "a year's worth of advances could occur in a matter of weeks," and wrote that Hinton's participation was "noteworthy." Its "Reality check" noted the authors' own view that an intelligence explosion "is far from certain."^[14]
• Quartz (September 28) cited The Wall Street Journal's reporting, including a quotation from Song: "Human society is not really positioned for such fast changes and disruptions." It also placed the paper against President Donald Trump's statement at the United Nations that the U.S. would "totally reject" any "globalist scheme" to control AI, as reported by the Journal.^[16]
• The Next Web (Alina Maria Stan, September 28) reported, also citing the Journal, that OpenAI, Microsoft and Meta declined to comment and Anthropic did not respond, and quoted Song telling the Journal: "Already today, we are at the stage where we need AI systems to monitor what agents are doing."^[15]
• The Rundown AI (September 29) wrote that Hinton and Bengio "have become familiar voices" and that "the more striking development is Clark and Pachocki joining them," while noting that the paper says its authors' views do not necessarily represent their organizations.^[17]

CASP's homepage lists The Wall Street Journal and The Guardian as outlets that featured the paper.^[2] Details attributed to the Journal above are as relayed by Quartz and The Next Web.^[15][16]

Limitations and criticism

The paper is explicit about its uncertainty. It calls its evidence "preliminary and sometimes mixed," says estimates of r rely on "limited data and stylized modeling assumptions," and notes that "precisely operationalizing an intelligence explosion is tricky and remains an area for future work."^[1] Coverage also noted that the authors present a possibility rather than a forecast.^[14]

Some of its underlying figures come with caveats from their own sources. Anthropic describes lines of code as "an imperfect measure" and says its 8x lines-of-code-per-engineer figure "is almost certainly an overstatement of the true productivity gain."^[7] OpenAI's research-acceleration report says that "the overall pace of progress likely won't keep pace with these specific metrics," and that "over half of successful 4-8 hour tasks involved 1 or more interventions."^[10] Anthropic's automation index is self-assessed with a Claude judge model.^[8]

More broadly, the intelligence explosion hypothesis has long-standing critics. François Chollet argued in his 2017 essay "The impossibility of intelligence explosion" (since retitled "The implausibility of intelligence explosion") that "system bottlenecks, diminishing returns, and adversarial reactions end up squashing recursive self-improvement in all of the recursive processes that surround us," so that self-improvement "tends to be linear, or at best, sigmoidal."^[21] Arvind Narayanan and Sayash Kapoor, in "AI as Normal Technology" (Knight First Amendment Institute, 2025), wrote that "AI development already relies heavily on AI" and that "it is more likely that we will continue to see a gradual increase in the role of automation in AI development than a singular, discontinuous moment when recursive self-improvement is achieved."^[22] The paper engages the same frictions these critics emphasize (diminishing returns, bottlenecks, physical and sequential limits) but concludes that current evidence does not rule out an acceleration and that the stakes justify preparation.^[1] Critiques of related scenario work are covered under AI 2027 and recursive self-improvement.

See also

• Recursive self-improvement
• AI 2027
• The Anthropic Institute
• An Alien Mind
• We Must Pace the Frontier
• OpenAI-Hugging Face agent incident
• International AI Safety Report
• Compute governance
• METR

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53 ^54 ^55 ^56 ^57 ^58 ^59 ^60Chan, Alan; Winter, Christoph; Barto, Andrew; Pachocki, Jakub; Hinton, Geoffrey; Horvitz, Eric; Bengio, Yoshua; Song, Dawn; Clark, Jack; Greaves, Hilary; Korinek, Anton; Hammond, Samuel; Graepel, Thore; Bariach, Ben; Torr, Philip H. S.; McIlraith, Sheila A.; Clune, Jeff; Manning, Sam; Sastry, Girish; Davidson, Tom; Eth, Daniel; Mindermann, Sören. "What if automating AI R&D triggers an intelligence explosion?" Frontier AI Working Paper Series No. 2/2026, Cambridge Programme on AI Science & Policy, September 2026. casp.ac/...intelligence-explosion.pdf
2. ^1 ^2 ^3 ^4Cambridge Programme on AI Science & Policy. "What if automating AI R&D triggers an intelligence explosion?" (report page) and CASP homepage. Accessed September 30, 2026. casp.ac/...intelligence-explosion ; casp.ac
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Cambridge Programme on AI Science & Policy. "What if automating AI R&D triggers an intelligence explosion? Executive Summary." September 2026. casp.ac/...intelligence-explosion-summary.pdf
4. ^1 ^2 ^3 ^4 ^5 ^6GovAI. "What If Automating AI R&D Triggers an Intelligence Explosion?" September 28, 2026. governance.ai/...riggers-an-intelligence-explosion
5. ^1 ^2Foundation for American Innovation. "What If Automating AI R&D Triggers an Intelligence Explosion?" Accessed September 30, 2026. thefai.org/...d-triggers-an-intelligence-explosion
6. ^Cambridge Programme on AI Science & Policy. "About." Accessed September 30, 2026. casp.ac/about
7. ^1 ^2 ^3 ^4Favaro, Marina; Clark, Jack. "When AI builds itself." Anthropic Institute, June 2026 (with update of September 18, 2026). anthropic.com/...recursive-self-improvement
8. ^1 ^2 ^3 ^4 ^5 ^6Favaro, Marina; Wright, Phillie. "Measurements for understanding the pace of AI development inside frontier labs." Anthropic Institute, September 17, 2026. anthropic.com/...measuring-pace-of-ai-development
9. ^1 ^2 ^3METR. "Frontier Risk Report (February to March 2026): A pilot assessment of rogue deployment risk at frontier AI companies." Published May 19, 2026. metr.org/risk-report-feb-mar-2026.pdf
10. ^1 ^2 ^3OpenAI. "Research acceleration: The view inside OpenAI." September 2026. openai.com/...arch-acceleration-view-inside-openai
11. ^1 ^2Pachocki, Jakub. "An Alien Mind." OpenAI, September 2026. openai.com/...an-alien-mind
12. ^METR. "Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident." August 26, 2026. metr.org/...ai-hugging-face-incident-investigation
13. ^Milmo, Dan. "AI godfathers warn of runaway 'intelligence explosion'." The Guardian, September 28, 2026. theguardian.com/...-runaway-intelligence-explosion
14. ^1 ^2Axios. "AI pioneers warn of an 'intelligence explosion'." September 28, 2026. axios.com/...ai-pioneers-intelligence-explosion
15. ^1 ^2 ^3Stan, Alina Maria. "Hinton, Bengio and AI lab scientists warn of an intelligence explosion." The Next Web, September 28, 2026. thenextweb.com/...per-hinton-bengio-pachocki-clark
16. ^1 ^2 ^3Quartz. "Top AI researchers are warning of an 'intelligence explosion' and calling for urgent oversight." September 28, 2026. qz.com/...-intelligence-explosion-oversight-092826
17. ^1 ^2The Rundown AI. "Anthropic and OpenAI leaders join warning about AI that improves itself." September 29, 2026. therundown.ai/...rs-intelligence-explosion-warning
18. ^Song, Dawn (@dawnsongtweets). Post on X announcing her move to Meta Superintelligence Labs. June 25, 2026. x.com/...2070191051873345910
19. ^TechRadar. "'The goal is not to replace humans': new Meta AI research chief Dawn Song says the next frontier is AI agents that are 'economically valuable'." June 29, 2026. techradar.com/...ts-that-are-economically-valuable
20. ^Norman, Jeremy. "Irving John Good Originates the Concept of the Technological Singularity." History of Information. Accessed September 30, 2026. historyofinformation.com/detail
21. ^Chollet, François. "The implausibility of intelligence explosion" (originally "The impossibility of intelligence explosion"). Medium, November 27, 2017. medium.com/...-intelligence-explosion-5be4a9eda6ec
22. ^Narayanan, Arvind; Kapoor, Sayash. "AI as Normal Technology." Knight First Amendment Institute, April 15, 2025. knightcolumbia.org/...ai-as-normal-technology
23. ^Anthropic (@AnthropicAI). Post on X introducing three measurements of AI development. September 17, 2026. x.com/...2100684274114699295
24. ^Willison, Simon. "Research acceleration: The view inside OpenAI" (link blog entry). September 6, 2026. simonwillison.net/...ration-the-view-inside-openai