{"id":"71f4849b-52c5-48e5-9013-9515bde5aed1","arxiv_id":"2502.19425","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A multi-logistic curve fit to historical AI milestones predicts the current AI wave peaks around 2024 and fades by 2035 to 2040, making an imminent singularity unlikely.","lead":"This paper fits a curve made of several S-shaped growth steps to the history of AI milestones and concludes that the current deep learning wave is peaking around 2024, with a decline by 2035 to 2040 if no new breakthrough appears. A generalist should read it to see one quantitative argument against an imminent technological singularity.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2035–2040 decline and 'no singularity soon' conclusion are not supported: the model fixes N=3, fits the current wave only on its pre-inflection branch, and has no mechanism to predict the overlapping fourth wave that Section 4.4 admits is uncertain.","rationale":"The reader's rejection is well founded. I add a sharper reason: the forecasted decline is not an empirical prediction but a consequence of imposing a logistic saturation on a pre-inflection trajectory and fixing the number of waves a priori. The paper's own qualitative admission about the fourth wave (Section 4.4) is a stated limitation and should be weighed as such. The comparative fits in Table 2 are real but not decisive for extrapolation: the AIC gap between Multi-Logistic and LPPL on the Total series is only 0.33, so the model-selection evidence for the specific multi-logistic structure is weak. The Arxiv cross-validation in Section 4.5 is likewise a within-sample fit of a logistic to a 16-year interval; it does not test the 2035–2040 horizon. My proposed N=4 check would settle whether the wave count is identified. Since this concern supports the reader's REJECT without changing it, I recommend UNCHANGED.","tokens_in":12276,"tokens_out":8680,"duration_ms":84428,"concrete_test":"Fit a four-logistic model (N=4) to the AI Historical Statistics total series using the same Levenberg–Marquardt procedure, with the fourth component's midpoint free in 2025–2045, and compare to N=3 via AIC/BIC. If the N=4 fit has ΔAIC < 2 or lower BIC and produces a non-negligible fourth wave beginning before 2040, then the fixed-N forecast and the 2035–2040 decline are not robust to the model's main structural assumption. If N=4 is not identifiable (flat likelihood over midpoints), that also demonstrates the data cannot fix the wave count.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.4 fixes N=3 for total/academia and N=2 for industry from the 'three peaks and two troughs' narrative, then fits Eq. 3. The headline quantities—fastest point 2024 and decline 2035–2040—come from the third logistic component's midpoint m3=2024.46 and width w3=4.33 (Table 3). This is load-bearing because both quantities are properties of a curve that has not yet shown saturation. The data end in 2023, before m3, so the current wave is observed only on the pre-inflection, accelerating branch. In that regime the ceiling a3=1664.85 and w3 are weakly identified: many (a3,w3,m3) triples reproduce the observed early trajectory with very different decline dates. The confidence intervals from Eq. 9 are linear-regression intervals around the fitted curve and do not propagate this non-identifiability, so the 2035–2040 window is an artifact of the assumed single-logistic form rather than a data-derived bound.\n\nSecond, even if the current-wave fit is accepted, the conclusion 'the technological singularity appears unlikely in the foreseeable future' does not follow from the fitted decline, because the fitted decline is explicitly conditional on 'no fundamental technological innovation emerges.' The model has no arrival process for new logistic components; it cannot predict when a fourth wave begins. Section 4.4 itself states that the timing of the fourth wave 'remains uncertain' and that historical waves overlap. A fourth wave beginning around 2030–2035 is consistent with the model's own historical overlap pattern and would invalidate the singularity conclusion, yet the model is silent on its probability. The paper's central claim therefore rests on an assumption about unmodeled innovation, not on the fitted data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a multi-logistic growth model for the cumulative number of famous AI systems, hypothesizing that AI development is a superposition of logistic components. Using the AI Historical Statistics, AI Arxiv Papers, GPU Transistor, and Internet User datasets, the authors fit the model with N=3 waves for total/academia and N=2 for industry/collaboration, compare against logistic, exponential, LPPL, and polynomial fits, and read off from the fitted third-wave midpoint (m3≈2024) that the current AI wave peaks around 2024 and will decline around 2035-2040 absent fundamental innovations. The paper concludes that the technological singularity is unlikely in the foreseeable future.","tokens_in":12663,"tokens_out":7041,"duration_ms":55384,"significance":"If the central forecast were supported, the paper would provide a quantitative, data-driven counterpoint to exponential-growth singularity forecasts and an interpretable decomposition of AI history into logistic waves. The in-sample fits are excellent (R² ≈ 0.999 on the AI Historical Statistics dataset), and the comparison against ordinary logistic and exponential growth is a useful demonstration that a single S-curve cannot capture the multiple waves. The cross-validation on Arxiv paper counts with a single logistic component is a reasonable check of the current-wave timing. However, because the headline decline prediction is extracted from parameters that are weakly identified from pre-inflection data, and because the model has no mechanism to forecast the fourth wave that the authors themselves acknowledge, the paper does not establish the 'objective law' it claims. The contribution is a plausible descriptive model plus an unsupported extrapolation.","major_comments":[{"comment":"The claims that the current wave peaks around 2024 and declines in 2035-2040 are direct readings of the fitted parameters m3=2024.46 and w3=4.33 in Table 3, not independent forecasts. Because the data end in 2023, before m3, the third logistic component is observed only on its pre-inflection, accelerating branch. In this regime, a3 and w3 are weakly identified: many (a3,w3,m3) triples reproduce the observed early trajectory with very different saturation behavior. The confidence intervals from Eq. (9) are linear-regression intervals around the fitted curve and do not propagate the joint uncertainty in the nonlinear parameters. The paper should provide profile-likelihood or bootstrap intervals for m3 and w3 and should treat the 2035-2040 decline window as a consequence of the assumed logistic form, not as a data-derived bound.","section":"4.4, Table 3, Eq. (9)"},{"comment":"The conclusion that 'the technological singularity appears unlikely in the foreseeable future' does not follow from the fitted decline of the third wave. The multi-logistic model is a sum of independent logistic components and contains no arrival process for future components. Section 4.4 itself states that the timing of a fourth wave 'remains uncertain' and that historical waves overlap; a fourth wave beginning around 2030-2035 is consistent with the model's historical pattern. The fitted decline can therefore support only the conditional statement 'if no fundamental technological innovation emerges,' not a judgment about the probable occurrence of a singularity, which would depend on future waves that the model does not predict.","section":"4.4, Section 5"},{"comment":"The number of waves N is fixed a posteriori from the 'three peaks and two troughs' narrative (N=3 for total/academia, N=2 for industry/collaboration), and no comparison is made to alternative values of N on the same data. Because the chosen N determines which model component produces the headline m3 and w3, the model comparisons in Table 2 do not select among multi-logistic models with different numbers of waves. The claim in Section 5 that the multi-logistic process is 'an objective law of AI technology development' therefore overstates the evidence, which supports only a descriptive fit to the historical waves.","section":"4.4, Table 2"},{"comment":"The cross-validation experiment on the AI Arxiv Papers dataset fits a single logistic function to cumulative counts from 2008 to 2023 and then uses the fitted function to assert a decline around 2035-2040. Only 16 data points are used, no data are held out, and the decline time is again an extrapolation of the fitted functional form beyond the observation period. This is a consistency check of the current-wave midpoint, not a validation of the forecast; the term 'cross-validation' is therefore misleading.","section":"4.5, Table 5"}],"minor_comments":[{"comment":"The research questions are misnumbered: 'RQ1' appears twice; the second should be RQ2, and the subsequent questions should be renumbered accordingly.","section":"Section 4"},{"comment":"The description of the four growth stages contains duplicated words ('which which') and a confusing statement that the development speed 'begins to decline exponentially' in both the maturity and saturation stages; the text should be edited for clarity.","section":"Section 3"},{"comment":"There are several typos: 'verity' should be 'verify', and 'are are' appears in the definition of N and Nv. Additionally, Table 1 labels four datasets while the text describes categories within datasets; the wording should be aligned.","section":"Section 4.1"},{"comment":"Figure 2 is described before it is introduced, and the claim that the logistic model has a higher R² than exponential growth for LLM papers is not quantified in the text; the authors should provide the numerical comparison.","section":"Introduction"},{"comment":"The sentence 'From the results in Table 4, we obtain the 2σ time range of total AI development in the third wave is [2009, 2040]' is grammatically incomplete; 'is' should be 'as' or the sentence should be restructured.","section":"Section 4.4"},{"comment":"Equation (9) is the standard formula for a linear-regression confidence interval; its use for a nonlinear model with estimated parameters should be justified or replaced with a nonlinear method (e.g., delta method or bootstrap).","section":"Eq. (9)"}],"recommendation":"reject","confidential_remarks":"The manuscript's data collection and descriptive fits are potentially useful, and the authors are honest about the conditionality of the forecast in places. However, the central inference—that the singularity is unlikely—is not supported by the model, and the weaknesses are structural rather than presentation-level. I recommend rejection, but a reframed version that presents the multi-logistic model as a descriptive decomposition and avoids the singularity verdict could be reconsidered as a new submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper fits a sum of logistic curves to the cumulative number of famous AI systems and uses the fitted third wave to predict that AI development peaks around 2024 and declines by 2035–2040, concluding the singularity is not near. The fitting is competently done and the historical decomposition into three overlapping waves (two for industry) is visually plausible and clearly explained. I also give the authors credit for comparing against several baselines and for cross-checking with arXiv paper counts, GPU transistors, and internet users. As a descriptive history of AI as a sequence of technology waves, this is reasonable work.\n\nThe soft spot is the central extrapolation. The 2024 peak is just the fitted midpoint m3 of the third logistic term, and the 2035–2040 decline is the 2σ interval computed from the fitted width w3. The data end in 2023, before that midpoint, so the current wave is only observed on its pre-inflection, accelerating branch. In that regime the ceiling and width are weakly identified—many parameter triples reproduce the early data with very different saturation dates. The confidence intervals in Eq. 9 are linear-regression intervals and do not propagate this non-identifiability, so the forecast window is essentially an artifact of the imposed logistic shape, not a data-driven bound.\n\nThe second issue is logical. The decline is explicitly conditional on \"no fundamental technological innovation emerges,\" and Section 4.4 admits that a fourth wave's timing is uncertain and that historical waves overlap. But the model has no arrival process for new waves. A fourth wave starting around 2030–2035 is entirely consistent with the historical pattern the paper itself documents. So the abstract's unqualified statement that the singularity is unlikely does not follow from the fitted curve—it depends on an unmodeled assumption about future breakthroughs.\n\nThe number of waves is also chosen post hoc from the \"three peaks and two troughs\" narrative, and calling the multi-logistic pattern an \"objective law\" overstates what a curve fit can establish. These flaws undermine the headline claim, but they do not destroy the paper's value as a modeling exercise.\n\nWho is this for? Readers interested in technology forecasting or AI policy debates will find a thoughtful, quantitative counterpoint to exponential singularity predictions, even if the certainty is unwarranted. It deserves a serious referee—the topic is timely and the methodology, while flawed at the extrapolation step, is transparent and testable. I would send it to peer review with a request for proper uncertainty quantification and a reframing of the forecast as a scenario, not a prediction. But as it stands, the central conclusion is not supported.","headline":"A decent curve-fitting exercise with an honest historical decomposition, but the headline singularity forecast is just a labeled logistic parameter, not a robust prediction.","tokens_in":13208,"tokens_out":2079,"would_cite":false,"duration_ms":21543,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Stacking S-shaped growth curves onto 70 years of AI milestones puts the current deep-learning wave's peak at 2024 and its fade at 2035–2040, implying the technological singularity is not near.","keywords":["technological singularity","artificial intelligence","multi-logistic growth","logistic growth process","technology forecasting","deep learning","large language models","innovation waves"],"falsifier":"If, after 2024, the annual increase in notable AI systems or AI papers does not begin falling along the fitted curve, or if a new wave of foundational AI systems appears before roughly 2035, the central claim is falsified.","tokens_in":12089,"feed_emoji":"📈","tokens_out":9782,"duration_ms":82174,"temperature":0.7,"pith_summary":"The paper sets out to test a contested question with data: is the technological singularity imminent? It hypothesizes that AI development is not one exponential takeoff but a sequence of overlapping S-shaped growth waves, and it fits a multi-logistic model to the cumulative number of famous AI systems from 1950 to 2023. The fitted third wave peaks around 2024 and reaches saturation around 2035–2040 unless a fundamental new breakthrough starts another wave. Independent counts of AI papers, GPU transistors, and internet users corroborate the same timing. The authors conclude that a near-term singularity is unlikely, while acknowledging the current wave is still transformative.","feed_headline":"AI boom peaks in 2024 and fades by 2040, model finds","feed_subtitle":"Seven decades of AI milestones, fit to stacked S-curves, put the wave's peak at 2024 and its fade near 2035.","key_machinery":"The multi-logistic growth function $$L_n(t) = a_0 + \\sum_{i=1}^{N} \\frac{a_i}{1+\\exp\\left(-\\frac{t-m_i}{w_i}\\right)}$$ sums one logistic curve per technology wave, each with ceiling $a_i$, midpoint $m_i$, and speed width $w_i$. The midpoint $m_i$ is the fastest-growth point of that wave, and the interval rules $\\sigma_i=\\pi w_i/\\sqrt{3}$ divide each wave into emergence, growth, maturity, and saturation stages. This object carries the argument because the fitted third-wave midpoint $m_3\\approx 2024$ and its $2\\sigma$ upper edge near 2040 convert the historical dataset into a dated forecast. The first derivative of the sum is treated as the speed or heat of AI development, and the Levenberg-Marquardt method is used to estimate the parameters.","core_discovery":"The paper's central claim is that the dynamics of AI development are multi-logistic: the cumulative number of well-known AI systems from 1950–2023 is well described by a sum of logistic curves, with three waves for overall AI and two for industry. The estimated midpoint of the current deep-learning wave is $m_3 \\approx 2024.46$, and its $2\\sigma$ saturation interval runs to about 2040, so the model predicts that the current wave will plateau in 2035–2040 absent a new fundamental innovation. The first derivative of the fitted sum, read as the speed of AI development, is near its maximum now, which explains the intensity of LLM progress without invoking runaway growth. The consistency of the fits across the system-count, paper-count, GPU transistor, and internet-user datasets is taken as evidence that the forecast is robust.","pith_inferences":["A testable extension is to fit the same stacked S-curves to capability-weighted AI measures, such as benchmark scores or economic output, to check whether the 2024 peak and 2035–2040 saturation hold beyond simple system counts.","Because the paper fixes the number of waves in advance and does not model how a fourth wave would be born, the 2035–2040 fade is best read as the no-breakthrough baseline rather than as a hard ceiling.","If the midpoint is really 2024, year-over-year growth in new notable AI systems should be visibly declining by the late 2020s; this is a concrete calendar check of the model."],"forward_implications":["The current LLM boom is the steepest part of the third S-curve, so the rate of new notable AI systems should start decelerating from here.","Without a fundamental theoretical breakthrough, cumulative AI output will approach saturation in the 2030s, with the growth phase ending around 2035–2040.","The second wave's long duration supplied the deep-learning foundations, and industry now drives the third wave, a shift that frames where the next wave would likely originate.","The exponential-growth view of AI, and with it the near-term technological singularity, is rejected by the data; any singularity scenario is pushed beyond the current wave's horizon."],"supporting_citations":[{"why":"supplies the original logistic growth law that the model superposes into multiple waves.","marker":"[28]"},{"why":"shows logistic curves fit US energy production and consumption, supporting the logistic premise for technology.","marker":"[19]"},{"why":"models Intel chip density with a logistic curve, a direct precedent for technology as an S-shaped process.","marker":"[20]"},{"why":"fits logistic curves to historical milestones to estimate future milestone timing, the closest forecasting predecessor.","marker":"[27]"},{"why":"states the exponential-growth singularity hypothesis that the comparison experiments reject.","marker":"[9]"},{"why":"supplies the data-scaling limits used to argue deep learning will hit resource bottlenecks by the 2030s.","marker":"[26]"},{"why":"provides the log-periodic power law baseline the multi-logistic model is compared against.","marker":"[41]"},{"why":"gives the Levenberg-Marquardt algorithm used to estimate the multi-logistic parameters.","marker":"[40]"}],"fun_headline_variants":["AI wave peaks 2024, fades by 2040: model","Singularity postponed: AI curve peaks now, declines by 2040","No singularity in sight: AI growth peaks 2024, plateaus by 2040","AI's multi-logistic model: peak now, fade by 2040, no singularity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The forecast stands on the assumption that the current deep-learning wave is one S-shaped curve whose shape is already set by data through 2023, and that the number of waves (three overall, two for industry) is known in advance.","fun_headline_variants_meta":{"raw":{"variants":["AI wave peaks 2024, fades by 2040: model","Singularity postponed: AI curve peaks now, declines by 2040","No singularity in sight: AI growth peaks 2024, plateaus by 2040","AI's multi-logistic model: peak now, fade by 2040, no singularity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000721,"raw_usage":{"total_tokens":3256,"prompt_tokens":984,"completion_tokens":2272,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":2183}},"tokens_in":600,"tokens_out":2272,"duration_ms":14592,"temperature":1.0,"reasoning_tokens":2183,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:26:00.306259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If, after 2024, the annual increase in notable AI systems or AI papers does not begin falling along the fitted curve, or if a new wave of foundational AI systems appears before roughly 2035, the central claim is falsified.","supporting_citations":[{"cited_title":"Verhulst, Recherches math´ ematiques sur la loi d’accroissement de la population, M´ emoires de l’Acad´ emie royale de Belgique 18 (1845) 1–40","cited_arxiv_id":null,"evidence_quote":"supplies the original logistic growth law that the model superposes into multiple waves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"shows logistic curves fit US energy production and consumption, supporting the logistic premise for technology."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"models Intel chip density with a logistic curve, a direct precedent for technology as an S-shaped process."},{"cited_title":"Modis, Links between entropy, complexity, and the technological singularity, Tech- nological Forecasting and Social Change 176 (2022) 121457","cited_arxiv_id":null,"evidence_quote":"fits logistic curves to historical milestones to estimate future milestone timing, the closest forecasting predecessor."},{"cited_title":"Kurzweil, The singularity is near, in: Ethics and emerging technologies, Springer, 2005, pp","cited_arxiv_id":null,"evidence_quote":"states the exponential-growth singularity hypothesis that the comparison experiments reject."},{"cited_title":"Dro˙ zd˙ z, F","cited_arxiv_id":null,"evidence_quote":"provides the log-periodic power law baseline the multi-logistic model is compared against."},{"cited_title":"Fletcher, A modified marquardt subroutine for non-linear least squares (1971)","cited_arxiv_id":null,"evidence_quote":"gives the Levenberg-Marquardt algorithm used to estimate the multi-logistic parameters."}],"review_version":1}