AI Tops Math, Coding Targets Years Early as Physical Adoption Stalls

The Experts Were Wrong About AI. Again. — Sabine Hossenfelder

Sabine Hossenfelder

Algorithms shatter expert timelines because pure software faces zero friction, while physical constraints and industrial bottlenecks stall automation across the broader economy.

AI-assisted summary of The Experts Were Wrong About AI. Again. — original source. Check the original for full context.

Source published:

Key ideas

  • AI benchmark forecasts in pure software have collapsed years ahead of schedule, including AI substantially assisting in solving a Millennium Prize Problem by September 2026 despite expert surveys assigning only a 10% probability by late 2027. (0:00)
  • Coding and reasoning capabilities vastly exceeded expectations, with top models reaching 53.8% on LiveCodeBench Pro in May 2026 against a year-end forecast of 14%, and matching International Mathematical Olympiad gold-medal performance five years ahead of projections. (0:00)
  • In the physical world, AI adoption has significantly underperformed forecasts: the World Economic Forum's estimate of 50% task automation across the global workforce by 2025 yielded a real-world substitution rate of just 22%. (3:24)
  • The immediate future of scientific discovery will likely be propelled by AI cross-indexing and synthesizing fragmented academic literature across theoretical fields where physical hardware friction does not apply. (5:05)

Detailed notes

## Background Published on YouTube by Sabine Hossenfelder on October 7, 2026, this video investigates how experts have persistently miscalculated the trajectory of artificial intelligence. According to the show description, benchmarks once projected to take a decade collapsed within a single year, even as forecasters made the opposite mistake in other domains.

## TL;DR AI specialists have consistently underestimated progress in purely digital and symbolic domains, shattering multi-year benchmarks in competitive programming, math proofs, and vulnerability detection far ahead of schedule due to academic forecasters failing to anticipate private capital velocity. In contrast, prognosticators radically overestimated AI's immediate impact on physical labor and infrastructure, where the real-world friction of power generation, hardware construction, and institutional resistance significantly slowed adoption compared to predictions. Because pure software exhibits zero friction, the near future offers a rapid revolution in scientific literature synthesis, allowing distributed researchers to solve long-standing theoretical puzzles before confronting broader societal crossroads.

## Atmosphere & style The presentation is marked by Sabine Hossenfelder’s trademark dry wit, skeptical pragmatism, and deadpan comedic delivery. She pairs rigorous survey data with biting satire aimed at circular Silicon Valley hype and bewildered academic researchers who openly submit papers they cannot comprehend. The video feels like an engaging and grounding reality check, shifting smoothly from brisk debunks of tech hyperbole to genuinely curious optimism about the immediate future of scientific discovery.

## Bulletin - [00:00](https://www.youtube.com/watch?v=xdwFncdu8t0&t=0s) AI benchmark forecasts in pure software have collapsed years ahead of schedule, including AI substantially assisting in solving a Millennium Prize Problem by September 2026 despite expert surveys assigning only a 10% probability by late 2027. - [00:00](https://www.youtube.com/watch?v=xdwFncdu8t0&t=0s) Coding and reasoning capabilities vastly exceeded expectations, with top models reaching 53.8% on LiveCodeBench Pro in May 2026 against a year-end forecast of 14%, and matching International Mathematical Olympiad gold-medal performance five years ahead of projections. - [03:24](https://www.youtube.com/watch?v=xdwFncdu8t0&t=204s) In the physical world, AI adoption has significantly underperformed forecasts: the World Economic Forum's estimate of 50% task automation across the global workforce by 2025 yielded a real-world substitution rate of just 22%. - [05:05](https://www.youtube.com/watch?v=xdwFncdu8t0&t=305s) The immediate future of scientific discovery will likely be propelled by AI cross-indexing and synthesizing fragmented academic literature across theoretical fields where physical hardware friction does not apply.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=xdwFncdu8t0&t=0s) Society has normalized rapid AI breakthroughs so quickly that the public routinely overlooks how consistently expert forecasts have failed. - [00:00](https://www.youtube.com/watch?v=xdwFncdu8t0&t=0s) More than 200 AI researchers surveyed in 2025 assigned only a 10% median probability to AI solving or aiding a Millennium Prize Problem by late 2027, a milestone reached in September 2026. - [00:00](https://www.youtube.com/watch?v=xdwFncdu8t0&t=0s) On LiveCodeBench Pro, models reached a 53.8% score by May 2026, delivering nearly four times the median expert forecast of 14% well ahead of schedule. - [00:00](https://www.youtube.com/watch?v=xdwFncdu8t0&t=0s) The AI task-completion time horizon reached 3 hours and 6 minutes in spring 2026 before the survey forecasting 3.4 hours by year-end had even closed. - [01:56](https://www.youtube.com/watch?v=xdwFncdu8t0&t=116s) A 2023 survey of over 2,000 researchers predicted it would take 22 years for AI to prove publication-grade math theorems, yet preprint archives were flooded with AI proofs within three years. - [01:56](https://www.youtube.com/watch?v=xdwFncdu8t0&t=116s) Academic forecasters systematically underestimate progress because their mental models are anchored to slow academic research rather than private markets backed by hundreds of billions in capital. - [03:24](https://www.youtube.com/watch?v=xdwFncdu8t0&t=204s) Forecasters committed the opposite error in physical domains, overpredicting autonomous trucking adoption and projecting 50% workforce task automation by 2025 when actual replacement reached only 22%. - [03:24](https://www.youtube.com/watch?v=xdwFncdu8t0&t=204s) Speculative 'intelligence explosion' models, such as Leopold Aschenbrenner's 2027 thesis, rely on flawed circular logic regarding natural gas drilling, power plant construction, and robotic automation. - [03:24](https://www.youtube.com/watch?v=xdwFncdu8t0&t=204s) Technological velocity is dictated by friction: pure software and symbolic math experience near-zero friction, while industrial supply chains, hardware labs, and power infrastructure require decades to scale. - [05:05](https://www.youtube.com/watch?v=xdwFncdu8t0&t=305s) Theoretical physics and mathematics may experience rapid breakthroughs not from a new Einstein, but from AI systems capable of reading, verifying, and connecting the entirety of existing literature. - [05:05](https://www.youtube.com/watch?v=xdwFncdu8t0&t=305s) Because access to software models is widely distributed, independent researchers globally are poised to participate in a temporary golden age of scientific discovery.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=xdwFncdu8t0&t=0s) Underestimating the Digital Frontier: Benchmarks Shattered Years Ahead of Schedule

A decade ago, the suggestion that computer software could generate an entire book—much less an academic research paper—would have sounded preposterous to almost anyone. By September 2026, headlines casually announced that a university professor had used AI to produce 200 papers in a single year, and the general public barely blinked. Society has normalized artificial intelligence so swiftly that people easily overlook how dramatically the technology has outpaced official forecasts. While experts have consistently attempted to forecast the trajectory of artificial intelligence, their benchmark projections have routinely collapsed in record time.

The extent of this forecasting failure is underscored by several surveys. When more than 200 AI researchers were surveyed in 2025 regarding the likelihood of an AI solving or substantially helping solve one of the Millennium Prize Problems by the close of 2027, the median expert estimated the probability at a mere 10%. AI crossed that milestone in September 2026. Similarly, when the same cohort evaluated how top-tier models would perform on LiveCodeBench Pro—an exceptionally rigorous coding benchmark—the existing high score stood at 7.7%. The median expert expected that figure to inch up to 14% by the end of 2026. Instead, models surged to 53.8% by May 2026, delivering nearly four times the predicted capability more than half a year ahead of schedule.

A parallel miscalculation emerged around the AI task-completion time horizon, a metric capturing how long a complex programming job would take an experienced human engineer while still allowing an AI system to solve it reliably. When queried in April and May 2026, specialists estimated the horizon would hit roughly 3.4 hours by the end of the year. In reality, a newly released model clocked 3 hours and 6 minutes before the polling survey had even closed.

- **Millennium Prize Problems:** 200+ researchers gave a 10% median probability of AI solving or aiding a solution by late 2027; benchmark achieved in September 2026. - **LiveCodeBench Pro:** Baseline stood at 7.7%; experts forecast 14% by year-end 2026; actual performance reached 53.8% in May 2026 (~4x the forecast). - **Task-Completion Horizon:** Forecasted to reach ~3.4 hours by late 2026; hit 3 hours and 6 minutes while the survey was actively underway in spring 2026.

---

[01:56](https://www.youtube.com/watch?v=xdwFncdu8t0&t=116s) The Academic Blind Spot: Decades of Progress Compressed into Months

These broken forecasts are not an isolated fluke from the current year; they reflect a persistent multi-year pattern. In the autumn of 2023, a comprehensive poll of over 2,000 artificial intelligence researchers found an average estimate that it would take 22 years before AI could independently prove mathematical theorems worthy of publication in top-tier mathematics journals. Instead, mathematics and physics preprint servers are already inundated with submissions written by AI—manuscripts that their own submitters openly admit they cannot comprehend. The real novelty is not academic authors publishing work they fail to understand, but rather that "the new part is that they admit it." The debate in academic publishing has ceased to be about whether machine-generated proofs are technically publishable; it has shifted entirely to who bears legal and scholarly responsibility for the content.

Other domains reveal identical gaps between expert skepticism and subsequent reality. In 2023, leading cybersecurity professionals claimed that software capable of discovering and weaponizing zero-day software exploits on its own remained years in the distance; autonomous tools demonstrated those capabilities the very next year in 2024. In 2022, forecasters projected that competitive AI would not achieve a gold medal at the International Mathematical Olympiad (IMO) before 2030; competitive systems matched gold-medal performance by 2025.

The root cause of these serial underestimates lies in the sociological background of the forecasters. Because so many credentialed AI specialists work inside academia, their expectations are anchored to institutional research speeds where projects move at a crawl. They routinely fail to anticipate the sheer velocity of commercial markets when accelerated by hundreds of billions of dollars in private capital.

- **Automated Mathematical Proofs:** 2023 survey of 2,000+ researchers predicted a 22-year timeline; preprint servers were overwhelmed with AI-authored proofs within three years. - **Cybersecurity Exploits:** Autonomous finding and exploiting of vulnerabilities labeled "years away" in 2023; realized in 2024. - **International Mathematical Olympiad (IMO):** Gold-level achievement estimated for 2030 in 2022; achieved in 2025 (five years early).

---

[03:24](https://www.youtube.com/watch?v=xdwFncdu8t0&t=204s) The Physical World Strikes Back: Why Truckers, Jobs, and Power Plants Resisted AI

While experts repeatedly lowballed algorithmic capability in purely digital environments, they made the exact opposite mistake regarding the physical economy and the labor market. Long-standing predictions that commercial trucking would be among the first sectors eliminated by autonomous driving have proven hollow, with widespread driver displacement barely underway. Back in 2020, the World Economic Forum (WEF) estimated that almost 50% of all tasks across the global workforce would be automated by 2025. A follow-up analysis published in 2025 revealed the real-world substitution rate sat at roughly 22%. Practical experiments highlight ongoing limits; attempting to have ChatGPT write an entire script still flounders because the software struggles to distinguish meaningful points from irrelevant trivia—having merely perfected "the important skill of sounding very confident while wasting my time."

The starkest conceptual failure in AI forecasting belongs to Leopold Aschenbrenner and his thesis of an impending "intelligence explosion" arriving by 2027. This speculative scenario depended on generating hundreds of gigawatts of new power by simply drilling more natural gas, using AI-designed robots to construct the gas power plants, while running the AI models on the gas-powered supercomputers assembled by those very robots. The realization that Silicon Valley investors entrusted roughly $30 billion to an individual operating on such circular logic demonstrates how detached tech culture has become from physical logistics.

```mermaid flowchart LR accTitle: Leopold Aschenbrenner's Circular Scaling Logic accDescr: Diagram mapping the recursive feedback loop between gas extraction, power plants, supercomputers, and robot construction. subgraph PhysicalLoop["The Unrealistic Industrial Feedback Loop"] A["Drill for Natural Gas"] --> B["Build 100 GW Power Plants"] B --> C["Power AI Supercomputers"] C --> D["AI Engineers Autonomous Robots"] D -->|"Builds plants & drills wells"| A end ```

The governing principle of technological diffusion is simple: the pace of transformation is governed by physical constraints and institutional friction. Pure software environments—symbolic mathematics, code execution, and manuscript text—offer virtually zero friction, enabling exponential leaps. Conversely, erecting wet labs, building fleets of self-driving hardware, retooling industrial supply chains, laying power lines, and mining raw minerals involve friction that requires decades, not months, to resolve.

- **World Economic Forum Labor Projections:** 2020 estimate projected ~50% workforce task automation by 2025; actual 2025 audit measured 22%. - **Industrial Bottlenecks:** Physical construction, environmental permits, resource extraction, and corporate organizational habits do not scale at the velocity of digital code.

---

[05:05](https://www.youtube.com/watch?v=xdwFncdu8t0&t=305s) The Paper-Reading Revolution and the Future of Science

Recognizing the distinction between digital software fluidity and physical inertia makes it possible to predict the immediate future of research. Anything bound purely to software will advance at breakneck speed, unlocking tasks that were historically impossible due to human cognitive limits. A prime example is literature reconciliation: unifying, cross-indexing, and verifying the entire body of mathematical publications is suddenly feasible.

A vast collection of longstanding puzzles in theoretical physics and sister fields is poised for rapid resolution because the fundamental techniques required to solve them are already scattered across existing papers. The field may not require a once-in-a-century genius: "Maybe physics isn't waiting for a new Einstein, it's just waiting for someone to read all the papers!" Because software access is widely distributed rather than restricted to mega-corporations, independent researchers around the world can participate directly, kicking off a temporary golden age where virtually anyone can make genuine scientific discoveries. Eventually, society will reach the ultimate fork in the road: "Well, then we will either all die, or never work again, depending on whom you ask." Until then, the takeaway from the historical record is plain: "don’t listen to the experts. Follow me for more uninformed opinions."

---

[06:07](https://www.youtube.com/watch?v=xdwFncdu8t0&t=367s) Sponsor: Outskill and OpenAI's GPT-6 Astra Crash Course

With AI accelerating and OpenAI's newly surfaced GPT-6 Astra bringing the field closer to artificial general intelligence (AGI), keeping up with practical toolsets has become essential. Today's sponsor, Outskill, is offering a free three-hour crash course designed to teach participants hands-on workflow development for GPT-6 Astra. As an official partner of OpenAI, Outskill obtained advance access to test the model thoroughly before its public rollout.

The course demonstrates how to build tailored personal and professional systems, such as building targeted job-hunting workflows based on individual resumes or creating automated real-time monitors to track specific stock holdings. Enrollees also receive the Astra Power Playbook, complete with plug-and-play prompt formulas and preconfigured workflows. Viewers can register using the link in the video description or by scanning the on-screen QR code.

## Notable quotes > [01:56](https://www.youtube.com/watch?v=xdwFncdu8t0&t=116s) the new part is that they admit it. — Sabine Hossenfelder

> [03:24](https://www.youtube.com/watch?v=xdwFncdu8t0&t=204s) the important skill of sounding very confident while wasting my time. — Sabine Hossenfelder

> [05:05](https://www.youtube.com/watch?v=xdwFncdu8t0&t=305s) Maybe physics isn't waiting for a new Einstein, it's just waiting for someone to read all the papers! — Sabine Hossenfelder

> [05:05](https://www.youtube.com/watch?v=xdwFncdu8t0&t=305s) Well, then we will either all die, or never work again, depending on whom you ask. — Sabine Hossenfelder

> [05:05](https://www.youtube.com/watch?v=xdwFncdu8t0&t=305s) don’t listen to the experts. Follow me for more uninformed opinions. — Sabine Hossenfelder

## Mentioned Sabine Hossenfelder, OpenAI, GPT 6 Astra, Outskill, Millennium Prize Problems, LiveCodeBench Pro, ChatGPT, International Mathematical Olympiad (IMO), World Economic Forum (WEF), Leopold Aschenbrenner, Albert Einstein

## Video description 👉 Grab your free seat to the GPT 6 Astra Crash Course: https://links.outskill.com/SAHOOCT2 🔐 100% Discount for the first 1000 people 💥 Master GPT 6 Astra, Build Powerful AI Agent Workflows & Unlock the Astra Power Playbook — All for Free! 🎁 Bonuses worth $5100+ if you join and attend

For years, experts have tried to predict how fast artificial intelligence would progress, and they've been wrong again and again. Benchmarks they expected in a decade fell in a year. But in some areas, they got it wrong in the opposite direction. Let's take a look.

#artificialintelligence #technology #sciencenews

👕T-shirts, mugs, posters and more: ➜ https://sabines-store.dashery.com/ 💌 Support me on Donorbox ➜ https://donorbox.org/swtg 👉 Transcript with links to references on Patreon ➜ https://www.patreon.com/Sabine 📝 Transcripts and written news on Substack ➜ https://sciencewtg.substack.com/ 📩 Free weekly science newsletter ➜ https://sabinehossenfelder.com/newsletter/ 👂 Audio only podcast ➜ https://open.spotify.com/show/0MkNfXlKnMPEUMEeKQYmYC 🔗 Join this channel to get access to perks ➜ https://www.youtube.com/channel/UC1yNl2E66ZzKApQdRuTQ4tw/join 📚 Buy my book ➜ https://amzn.to/3HSAWJW

[Source](https://www.youtube.com/watch?v=xdwFncdu8t0)

中文简报

AI提前数年突破数学与编程目标，实体应用却陷入停滞 由于纯软件不存在摩擦，算法打破了专家的时间表预期；但受制于物理限制和行业瓶颈，更广泛经济领域的自动化进程陷入停滞。 AI专家一再低估了纯数字和符号领域的进展。由于学术界预测者未能预料到私人资本的推进速度，AI在竞技编程、数学证明和漏洞检测等领域远超预期地打破了跨年度基准测试。相比之下，预测者严重高估了AI对体力劳动和基础设施的近期影响；在这些领域，发电、硬件建设和制度阻力等现实摩擦使AI的普及速度显著低于预期。由于纯软件表现出零摩擦特性，不久的将来科学文献综合领域将迎来一场快速革命，使分散在各地的研究人员能够在面临更广泛的社会十字路口之前，解决长期存在的理论难题。 纯软件领域的AI基准预测已提前数年被打破，其中包括AI在2026年9月前实质性地协助解决了一道千禧年大奖难题，尽管专家调查认为到2027年底实现这一点的概率仅为10%。 (0:00) 编程和推理能力远超预期，顶级模型在2026年5月于LiveCodeBench Pro上达到了53.8%的成绩（而年底预测值仅为14%），并比预期提前五年达到了国际数学奥林匹克竞赛的金牌水平。 (0:00) 在物理世界中，AI的普及程度远低于预测：世界经济论坛此前估计到2025年全球劳动力将有50%的任务实现自动化，但现实中的替代率仅为22%。 (3:24) 科学发现的近期未来很可能会由AI推动，在不涉及物理硬件摩擦的各个理论领域中，通过AI对碎片化的学术文献进行交叉索引和综合。 (5:05)

Reading summary

Sabine Hossenfelder argues that experts consistently underestimate AI progress in purely digital domains while overestimating its disruption of the physical economy, failing to recognize that frictionless software development moves far faster than material infrastructure and institutional labor. Digital Benchmarks Surpassed Years Ahead of Schedule (0:00) Specialists have repeatedly failed to gauge the speed of algorithmic improvements in pure software environments. In 2025, a survey of over 200 AI researchers yielded a median probability of only 10% that an AI system would solve or significantly aid in solving a Millennium Prize Problem by late 2027; the milestone was achieved in September 2026. On the LiveCodeBench Pro coding benchmark, experts expected top scores to climb from 7.7% to 14% by late 2026, but models reached 53.8% by May 2026—nearly four times the forecast over half a year early. Similarly, when specialists surveyed in spring 2026 estimated that the AI task-completion time horizon would reach roughly 3.4 hours by year-end, a newly released model logged 3 hours and 6 minutes before the polling closed. Comparable forecasting failures appeared across mathematics and cybersecurity. In 2022, forecasters projected that AI would not match gold-medal performance at the International Mathematical Olympiad until 2030, a threshold reached in 2025. In 2023, cybersecurity professionals claimed autonomous tools capable of identifying and exploiting zero-day software vulnerabilities remained years away, yet such systems demonstrated those capabilities in 2024. Academic Pacing Versus Commercial Capital Velocity (1:56) The persistent underestimation of software velocity stems largely from the institutional background of forecasters. In autumn 2023, a poll of over 2,000 AI researchers yielded an average estimate that independent, top-tier journal-worthy mathematical theorem proving by AI remained 22 years away. Within three years, preprint servers in physics and mathematics were flooded with AI-authored submissions that submitters acknowledged they could not fully understand, shifting publishing debates toward legal and scholarly attribution. Hossenfelder attributes this gap to credentialed researchers anchoring their expectations to academic timelines, where projects advance slowly. They routinely fail to anticipate the sheer speed of commercial markets when driven by hundreds of billions of dollars in private capital. Physical Constraints Hinder Real-World Labor Automation (3:24) While forecasters understated pure software capabilities, they made the reverse error regarding the material economy. Predictions that commercial trucking would be rapidly replaced by autonomous driving have failed to materialize, and human drivers remain largely undisplaced. In 2020, the World Economic Forum estimated that nearly 50% of global workforce tasks would be automated by 2025, but an audit in 2025 measured actual task substitution at approximately 22%. Hossenfelder points to Leopold Aschenbrenner's 2027 'intelligence explosion' thesis as an extreme example of ignoring material logistics. Aschenbrenner postulated generating hundreds of gigawatts of power by drilling natural gas to power AI supercomputers, which would design robots to construct the gas power plants and drill more wells. Hossenfelder describes this recursive feedback loop as circular logic, emphasizing that building physical laboratories, permitting power plants, expanding electrical grids, and retooling supply chains face material and regulatory friction requiring decades, not months, to resolve. Literature Reconciliation as the Next Scientific Engine (5:05) Because pure software operates without physical friction, Hossenfelder contends that near-term breakthroughs will focus on reconciling and synthesizing published scientific literature. Theoretical physics and mathematics are constrained less by a lack of underlying principles than by human cognitive limits in absorbing vast archives of research. Systematically reading, cross-indexing, and verifying all existing papers could resolve longstanding problems using techniques that already exist in disconnected studies. Because access to software models is widely distributed rather than restricted to large corporations, independent researchers globally can engage in this synthesis, enabling a period of decentralized scientific discovery before society reaches broader structural inflection points. Hossenfelder concludes that evaluating technological progress requires separating frictionless digital environments from the physical economy. Pure software capabilities will continue to outpace conservative expert timelines, whereas real-world automation remains constrained by infrastructural, regulatory, and industrial friction.

Source quotes

[1:56] Sabine Hossenfelder: the new part is that they admit it. [3:24] Sabine Hossenfelder: the important skill of sounding very confident while wasting my time. [5:05] Sabine Hossenfelder: Maybe physics isn't waiting for a new Einstein, it's just waiting for someone to read all the papers! [5:05] Sabine Hossenfelder: Well, then we will either all die, or never work again, depending on whom you ask. [5:05] Sabine Hossenfelder: don’t listen to the experts. Follow me for more uninformed opinions.

Original transcript

[0:00] A decade ago, most of us would have been stunned  by the idea that a computer program can write [0:06] an entire book, let alone a scientific paper.  But come September 2026 and the headlines read [0:12] “Using AI, a professor wrote 200 papers  this year” and it doesn’t even surprise [0:17] anyone anymore. We have gotten so used to AI  so quickly that it’s easy to forget just how [0:22] much it’s outperformed expectations. At least  some of them. Today I have a quick rundown of [0:28] these wrong predictions, and some thoughts  on what we can learn from this. Make sure [0:34] you stick around till the end. I've got a  great offer from today's sponsor, Outskill. [0:39] Last year, a group of AI researchers asked more  than 200 experts how likely it was that artificial [0:47] intelligence would solve or substantially help  solve one of the Millennium Prize Problems by [0:53] the end of 2027. The median answer was 10%.  AI hit this benchmark in September 2026. [1:01] The same group of experts was asked how well  the best AI would perform during 2026 on an [1:07] extremely difficult programming test called  LiveCodeBench Pro. At the time, the best score [1:12] was 7.7%. The median expert prediction was that  this would rise to 14% by the end of 2026. By May, [1:21] AI had reached 53.8%. So the actual result  was almost 4 times the experts’ prediction, [1:28] and it arrived months before their deadline. [1:31] In April and May 2026, experts were asked about  something called the AI task-completion time [1:37] horizon. Roughly speaking, this measures how  long a coding job would take a human expert [1:41] when an AI can still complete it reliably.  The experts predicted that by the end of 2026, [1:47] AI would reach about 3.4 hours.  While the survey was still running, [1:52] a new model had already  reached 3 hours and 6 minutes. [1:56] This isn’t a recent thing, it’s  been going on for a few years now. [2:03] In autumn 2023, more than 2000 AI researchers  were asked how long various advances would [2:10] take. They estimated it’d be 22 years until AI  could prove mathematical theorems publishable in [2:17] leading mathematics journals. Today’s reality  is that the maths and physics preprint server [2:23] has announced they are being flooded with  papers that are, by admission of the author, [2:28] AI-written and that even the authors do not  understand. The new part isn’t that they don’t [2:33] understand their paper, the new part is that  they admit it. And the question is not any [2:37] longer whether the papers are publishable,  but who is responsible for the content. [2:42] In 2023 cybersecurity experts said that  AI wouldn’t be able to autonomously find [2:49] and exploit vulnerabilities for years  to come. AI hit the mark the next year. [2:54] In 2022, experts estimated that AI  would reach gold level skills in the [2:58] international Mathematical Olympiad  by 2030. AI reached it in 2025. [3:02] So you see the pattern there, experts have  for many years underestimated AI progress. [3:08] My personal speculation for why this is is that  many of the so-called experts work in academia. [3:13] They’re used to things taking forever. They  don’t understand that market economies operate [3:18] at a completely different pace, especially when  bolstered by hundreds of billions of dollars. [3:24] But there’s a flipside. If you look at what  experts predicted would happen to jobs, [3:29] they were largely wrong. For example, one of  the jobs that was supposed to be automated [3:34] first were truckers. This hasn’t even really  begun to happen. In 2020, the World Economic [3:40] Forum expected nearly half of work to be  automated by 2025. In a follow-up from 2025, [3:46] they found that the actual rate was more like 22%.  And I tried to get ChatGPT to write this script, [3:51] but it still can’t tell relevant from  irrelevant information. It has, however, [3:55] mastered the important skill of sounding  very confident while wasting my time. [4:00] The most telling failure though is  Leopold Aschenbrenner’s prediction [4:05] of the impending intelligence explosion by  2027. This, if you remember was supposed to [4:11] happen by building a few 100 GigaWatt power  plants by drilling for a little more gas. [4:17] And that in return was supposed to be done  by robots who were built by the AI run on [4:23] the gas-powered supercomputers that were  built by the robots. You see the problem. [4:28] That the Silicon Valley bros  let this guy handle $30 billion [4:32] tells you how far gone they are. Because really  the picture here isn’t so difficult to see. How [4:37] fast you can change reality depends dramatically  on how many physical changes you need to perform, [4:44] and how many institutional hurdles stand in the  way. Maths, coding, and scientific publications [4:50] are a straight entry with near zero resistance.  But building a lab, building autonomous vehicles, [4:56] building robots, changing workflows in  existing companies, building power plants, [5:00] mining resources, etc. That’s a completely  different thing. It will take decades. [5:05] And from that we can also extrapolate  what’s going to happen in the near future. [5:10] Rapid progress in anything that  can be done by software alone, [5:15] especially if that’s tasks which previously  were impossible, such as reconciling all [5:21] mathematics literature, which has now become  possible. There are many theoretical problems [5:25] in physics and related disciplines that will be  up for grabs soon. Because they are solvable. [5:30] And the techniques for solving them are  already in the literature. Maybe physics [5:35] isn’t waiting for a new Einstein, it’s just  waiting for someone to read all the papers! [5:40] The most exciting thing is that this is not  exclusively something that AI companies can [5:46] do. Everyone can contribute. And that, I think, is  the future we’re headed into. The glorious phase [5:52] in which everyone is a scientist, for a few years.  And then… Well, then we will either all die, [5:57] or never work again, depending on whom  you ask. I guess the lesson here is: [6:02] don’t listen to the experts. Follow  me for more uninformed opinions.. [6:07] Artificial intelligence is developing rapidly, and  OpenAI's GPT-6 Astra is closer to AGI than ever. [6:14] Are you keeping up? Are you sure? Today's sponsor,  Outskill, will bring you up to speed. Outskill's [6:21] offering a GPT-6 Astra crash course. It's 3 hours  long, and it'll teach you how to build your own [6:28] Astra workflow hands-on. Outskill is an official  OpenAI partner, so their team has exclusive access [6:35] before anyone else. They've tested it upside down  and had time to figure out how to get the most [6:41] out of it. In the workshop, you'll learn,  for example, how you can use Astra to find [6:46] jobs that are tailored to your skills, or how to  alert you of information concerning your stocks. [6:51] When you sign up, you'll also get the Astra power  playbook with prompt formulas and workflows ready [6:58] for you to plug and play. The link to sign-up's  in the description, or you can scan the QR code on [7:05] screen. I've signed up for it because it's an easy  way to stay up-to-date. So, go ahead, sign up for [7:11] this, and learn what less than 1% on this planet  know. Thanks for watching. See you tomorrow.