Innovation for Everyone?
AI doesn't just replace work. It speeds up everything else.
AI is setting off the fastest wave of innovation in history, but not for everyone yet.
In brief- For the first time, the machine comes up with ideas: this year an AI solved two math problems that had been open since 1970.
- Building an app now takes little more than an idea, but of 1,600 Lovable apps tested, one in ten exposed its users' data.
- Robots take vast sums: one startup, Apptronik, raised more than $935 million, and only a few players can keep up.
- It all runs on AI owned by a handful of companies. I think every country will want its own, and that energy will matter most.
Two math questions had been waiting for an answer since 1970.
This spring, an AI from Google DeepMind, AlphaProof Nexus, solved them, along with seven other problems left behind by the mathematician Paul Erdős. A computer checked every step.1 Since then, the announcements have kept coming. In October, OpenAI said it had proved a conjecture from 1991, and on November 20, Levent Alpöge claimed a proof of the Riemann hypothesis, worked out with Claude. Neither has been reviewed by mathematicians yet.2
In September, a drug discovered with generative AI entered the last stage of human trials, for idiopathic pulmonary fibrosis: a world first, according to its maker, Insilico Medicine.3 In October, a personalized mRNA vaccine for melanoma, added to the usual treatment, cut the risk of relapse or death by 42%.4 People who could no longer use their hands now control a computer with their thoughts, thanks to Neuralink. By January, the company had already implanted 21 patients.5
Not long ago, any one of these would have been the science story of the year. Every one of them came out this year. I think we’re entering the fastest wave of innovation humanity has ever seen.
The last piece ended on a promise: if machines take on much of the work, human creativity will finally get the room it deserves. But is this wave really within everyone’s reach? In the digital world, yes. In the physical world, not yet. And either way, the question remains: who owns the computing power, and who controls these AI models?
Sharing faster, learning faster
For most of human history, an idea traveled as far as a voice could carry. Writing, about 5,200 years ago in Mesopotamia, let ideas outlive the people who had them.6 The printing press, around 1455, spread them to thousands of readers.7 The telegraph, in 1844, sent them from one city to another in an instant.8 The web, invented in 1989, put much of the world’s knowledge one click away.9
Each innovation put more people to work on the same problems, faster and faster.
Every time we learned to share faster, we learned faster.
AI is the next leap, and it’s a different kind of leap. ChatGPT arrived in November 2022, thirty-three years after the web.10 AI reads, writes code and, as those Erdős problems show, can find a proof no one had ever found.
For the first time, the machine doesn’t just pass ideas on. It comes up with them.
And the wildest part is that the next accelerator of knowledge is already at the prototype stage, inside a few patients’ heads: a chip wired to the brain could one day move ideas at the speed of thought rather than the speed of fingers or voice.
Everyone builds
Building an app used to take a team. Now you just hand the idea to an AI.
You describe what you want in a few sentences: an online store, a blog, a phone app. After a few back-and-forths, an AI builds a first version that works. On Lovable, one of these tools, users were starting about a million new projects a week in June.11 Most will go nowhere, and that’s fine: trying now costs almost nothing.
Making things used to take skill and time, which meant money. In the digital world, AI is making it far cheaper and faster. What becomes scarce is what no machine can supply for you: an idea that meets a real need, and the trust of the people it’s meant for.
But when everyone can build, everyone can also build badly. The flip side is a flood: millions of apps and pieces of content fighting for the same attention, much of it mediocre, much of it fake. Mass-producing false information has never been easier.
Security is going the same way. An app built in one evening by someone who isn’t a programmer can hide flaws: invisible mistakes that let a hacker get in and steal customer data. In March 2025, a researcher tested more than 1,600 apps built with Lovable: one in ten let anyone read its users’ data: names, emails, sometimes financial details. The platform has since added checks.12 Our data already had value. It will have even more, and we keep handing it to apps we know nothing about.
Amid all that noise, I expect many more small businesses to succeed: the ones that manage to get heard and, above all, trusted.
The physical world follows, on one condition
Robots are another matter: they need factories and vast sums. One humanoid-robot startup, Apptronik, has raised more than $935 million in a single funding round, spread over 2025 and 2026.13 Only a few players can keep up, but what they build is moving fast.
At BMW’s plant in Spartanburg, in the United States, humanoid robots from the startup Figure spent ten months of 2025 on the line: more than 90,000 parts moved, more than 30,000 cars built.14 In China, one robot ran a half-marathon faster than the human world record,15 and another ran 100 meters faster than Usain Bolt.16 The machines’ brains were already racing ahead. Now their bodies are catching up.
In Texas, robots have already 3D-printed the walls of some 100 houses, almost all of them sold.17 On the Moon, a robotic lander from the American company Firefly ran NASA instruments in 2025,18 and NASA wants its robots to start assembling the future Moon base before astronauts even arrive.19 Figure has also tested its robot in 30 homes around San Francisco, where it managed to tidy living rooms, fold towels and make beds.20 In time, robots like this should be able to do all the housework. Figure’s rival 1X, which also makes humanoid robots, is already taking preorders for NEO, a $20,000 home robot, with first US deliveries announced for 2026 and an expert who can guide it remotely through tasks it doesn’t know.21 Before they catch on widely, prices will have to fall, makers will have to learn to build them at scale, they will have to be safe, and people will have to accept letting them into their homes. That will take time, but the direction is set.
In software, anyone can play. In the physical world, capital still deals the cards.
On someone else’s land
Small businesses build the apps, but the giants own the land: the AIs that write them belong to a handful of companies, almost all American or Chinese, which set the prices and the rules. The shop owner who put up an online store in one evening depends on them entirely. For a country, it’s a matter of sovereignty: can it let its hospitals, its banks or its army depend on an AI whose servers and decisions sit somewhere else?
I think every country will end up wanting to own its AI, its cloud and its computing power, just as it wants its own energy or its own army. That’s why so-called open models, which you can run on your own machines, are drawing interest: often a little less powerful, they leave you less dependent on an outside service, though you still have to buy the hardware. A funny turn of events: after twenty years of sending everything to the cloud, we want our tools back at home.
And behind these tools sit the chips, the metals to make them, and above all energy. I think energy will be the most important variable, because without it a country can’t run its AI.
What I hope for
I won’t promise a cure for cancer. But for the first time, I can picture one in my lifetime. I think the coming years will amaze even the people building these tools.
The question is who this wave will benefit. These tools also serve scammers, hackers, and states whose interests are not ours. The same tools can heal or harm: the AI that helps design a drug could help design a biological weapon. Between those who protect and those who attack, the cat-and-mouse game is only beginning. And as soon as a rival speeds up, nobody can afford to slow down. That is the race nobody can stop, the one the next piece takes up.
Sources
- Google DeepMind and co-authors, “Advancing Mathematics Research with AI-Driven Formal Proof Search” (AlphaProof Nexus: 9 of 353 formalized open Erdős problems, 44 open OEIS conjectures, proofs checked in Lean), arXiv:2605.22763, May 2026, revised 8 Jun 2026 — https://arxiv.org/abs/2605.22763 · Aarhus University, “Gergely Bérczi and DeepMind”, 19 Jun 2026 (two questions posed in 1970) — https://math.au.dk/en/currently/news/news-item/artikel/gergely-deepmind
- OpenAI, “Building scientific understanding with academic communities”, 26 Oct 2026 (local smoothing conjecture, Sogge 1991; proof and Lean formalization released, expert review invited) — https://openai.com/index/building-scientific-understanding-with-academic-communities/ · Levent Alpöge, announcement of a proof of the Riemann hypothesis (“I believe I have, working with Claude, discovered a proof”), 20 Nov 2026 — https://alpo.ge/remark.pdf · Lean and Mathlib communities, “Lean verification of the Riemann Hypothesis”, 21 Nov 2026 (the computer-checked proof certified correct) — https://leanprover-community.github.io/blog/posts/riemann-hypothesis/
- Insilico Medicine, first patient dosed in GENESIS-IPF-3, phase 3 trial of rentosertib for idiopathic pulmonary fibrosis, 9 Sep 2026 (“world’s first” is the company’s claim) — https://www.prnewswire.com/news-releases/insilico-medicine-doses-first-patient-in-genesis-ipf-3-the-worlds-first-phase-iii-trial-of-a-generative-ai-driven-innovative-drug-302873749.html
- INTerpath-001, intismeran autogene (Moderna/Merck) + Keytruda vs Keytruda alone, resected melanoma, 1,137 patients: hazard ratio 0.58 for recurrence or death, ESMO 2026; 2-year recurrence-free survival 82.2% vs 70.0%; investigational — https://www.cancernetwork.com/view/interpath-001-data-detail-survival-subgroup-and-safety-findings-in-melanoma · Merck, Q3 2026 results, 29 Oct 2026 — https://www.businesswire.com/news/home/20261029138599/en/
- Neuralink, “Two years of telepathy”, 28 Jan 2026 (21 participants worldwide) — https://neuralink.com/updates/two-years-of-telepathy/
- British Museum, Mesopotamia resources (cuneiform, about 3200 BC) — https://www.britishmuseum.org/sites/default/files/2019-09/Visit_Mesopotamia_KS2b.pdf
- Library of Congress, “The Gutenberg Bible” (“probably completed late in 1455 at Mainz”) — https://www.loc.gov/exhibits/bibles/the-gutenberg-bible.html
- Library of Congress, Samuel F. B. Morse Papers, “Invention of the Telegraph” (24 May 1844) — https://www.loc.gov/static/collections/samuel-morse-papers/articles-and-essays/invention-of-the-telegraph/
- CERN, “The World Wide Web turns 35” (proposal, March 1989) — https://home.cern/world-wide-web-35/
- OpenAI, “Introducing ChatGPT”, 30 Nov 2022 — https://openai.com/index/chatgpt/
- TechCrunch, “Lovable says it has hit $500M in annualized revenue with 1 million new projects a week”, 9 Jun 2026 (company figures) — https://techcrunch.com/2026/06/09/lovable-says-it-has-hit-500m-in-annualized-revenue-with-1-million-new-projects-a-week/
- Matt Palmer (a Replit employee, a Lovable competitor), “CVE-2025-48757”, 29 May 2025 (1,645 apps tested on 21 Mar 2025, 170 exposed data; Lovable has since added a pre-publication security scan and disputes responsibility) — https://mattpalmer.io/posts/2025/05/CVE-2025-48757/ · Semafor, 29 May 2025 — https://www.semafor.com/article/05/29/2025/the-hottest-new-vibe-coding-startup-lovable-is-a-sitting-duck-for-hackers
- Apptronik, “Apptronik closes over $935 million Series A”, 11 Feb 2026 (company source) — https://apptronik.com/news-collection/apptronik-closes-over-935-million-series-a
- BMW Group, humanoid robots in production (Figure 02 at Spartanburg: 10 months in 2025, more than 90,000 parts, more than 30,000 BMW X3) — https://newsroom.bmwgroup.com/emea/bmw-group-to-deploy-humanoid-robots-in-production-in-germany-for-the-first-time_20260824100831028
- AP via PBS, “Humanoid robot wins Beijing half-marathon, defeating the human world record”, 19 Apr 2026 (50:26 vs 57:20; how autonomous the winning run was is disputed) — https://www.pbs.org/newshour/world/humanoid-robot-wins-beijing-half-marathon-defeating-the-human-world-record
- AP via OPB, “A robot sprinter shatters a 100-meter record again”, 27 Aug 2026 (Tiangong Ultra, 8.64 s; Bolt 9.58 s; robot and human rules differ) — https://web.archive.org/web/20260915113947/https://www.opb.org/article/2026/08/27/a-robot-sprinter-shatters-a-100-meter-record-again-as-sparks-fly-at-chinas-robot-competition/
- ICON, “Wolf Ranch” (100 homes in Georgetown, Texas, walls printed by ICON’s robots, completed 2025; company source) — https://www.iconbuild.com/projects/wolf-ranch · The Star, 24 Nov 2025 (98 of the 100 homes sold) — https://www.jpph.gov.my/v3/wp-content/uploads/2025/12/42_The_Star_24.11.2025.pdf
- NASA, “Touchdown! Carrying NASA science, Firefly’s Blue Ghost lands on Moon”, 2 Mar 2025 — https://www.nasa.gov/news-release/touchdown-carrying-nasa-science-fireflys-blue-ghost-lands-on-moon/
- NASA, “NASA provides updates on Moon Base cargo landers, tech demonstrations”, 4 Aug 2026 (a plan: “Before astronauts arrive, robotic missions will deploy critical scientific instruments”) — https://www.nasa.gov/missions/moon-base/nasa-provides-updates-on-moon-base-cargo-landers-tech-demonstrations/
- Figure, “Helix 2.5: zero-shot generalization to 30 homes”, 17 Sep 2026 (trials: “We evaluated three tasks across 30 unseen Bay Area homes”) — https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization
- 1X, NEO order page, accessed 5 Nov 2026 (“$20,000”, “US Deliveries start 2026”, “an Expert from 1X can remotely supervise its actions”) — https://www.1x.tech/order
The Race Nobody Can Stop
Nobody can slow the AI race. It is putting capabilities that once took the resources of a state within reach of small groups, and concentrating power in a few companies.
Read on →