Top Tech News Today, September 22, 2026: Amazon, AMD, Google, Meta, OpenAI, Tesla & More

It’s Tuesday, September 22, 2026, and AI is no longer waiting for a product keynote. By midafternoon, frontier labs were on the United Nations Security Council calendar, China was investigating its own model makers, Meta’s new agent needed a security hotfix, and Shopify had opened the checkout lane that Amazon just closed. The same 24 hours stretched from a Helsinki cloud unicorn and an $899 Google laptop to California’s new data-center power bills, an FAA tool watching Washington airspace, and a robot contest aimed at the International Space Station.

What ties these stories together is bigger than another wave of AI announcements. Artificial intelligence is now reshaping chips, data centers, debt markets, cybersecurity, commerce, robotics, satellites, scientific research, and the physical infrastructure beneath the digital economy. Alibaba is building a full-stack AI empire, AMD has crossed the $1 trillion mark, AI spending is spilling into global bond markets, and startups are pushing intelligence into everything from recruiting software to machines that can touch, move, and operate in the real world.

Here are the top tech news stories showing just how quickly the ground is moving.

Amazon Blocks Meta’s Muse AI Agent as the Battle Over AI Shopping Heats Up

Amazon has blocked Meta’s fast-growing Muse AI assistant from shopping on Amazon.com, opening another front in the fight over who controls commerce when AI agents begin making purchasing decisions for consumers. Users attempting to access Amazon through Muse began receiving messages saying that continued access by an unauthorized AI agent violated Amazon’s terms. Meta’s assistant can search for products and complete purchases on participating platforms, turning the traditional shopping journey into a conversation with an AI agent.

The dispute reveals a deeper economic problem facing online marketplaces. If consumers increasingly ask AI agents to compare products and make purchases, the agent could replace much of the search, recommendation, advertising, and merchandising layer that platforms such as Amazon have spent decades building. Shopify is taking the opposite approach, integrating Muse with Shop Pay so users can purchase products from participating merchants. Amazon, meanwhile, has reasons to protect both its customer experience and its lucrative advertising business. The outcome may determine whether AI assistants become neutral gateways to the web or whether major platforms require agents to negotiate access individually.

Why It Matters: AI agents are beginning to challenge the traditional gatekeepers of online commerce, forcing platforms to decide whether automated shoppers are partners, customers, or competitors.

Source: The Wall Street Journal.

Tesla Reopens $50,000 Roadster Reservations Ahead of October 1 Reveal

Tesla reopened reservations for its second-generation Roadster, asking U.S. buyers for $5,000 on a card plus a $45,000 wire within 10 days. The company will unveil the car on October 1 near Waco, Texas. Elon Musk has teased SpaceX-designed cold-gas thrusters and told an audience the demo must be live so viewers know “it is not AI.” Tesla has not published a purchase price; the car is widely expected to cost well above $250,000. The original 2017 reservation holders, who were told to expect cars around 2020, are still waiting.

The deposit terms essentially restore the 2017 terms, which is both a cash-flow tool and a credibility test. Tesla needs a halo product while robotaxi and humanoid timelines remain contested and free cash flow has been under pressure. A $50,000 fully committed reservation is also a filter: only buyers willing to park that money before seeing production specs will queue. For the broader EV and autonomy stack, the event matters less as a sports-car launch than as a stage for whatever physical-AI claims Tesla attaches to thrusters, software, and manufacturing.

Why It Matters: Tesla is collecting large deposits before showing a production Roadster, turning a nine-year delay into a high-stakes hardware and brand reset.

Source: Fox Business.

Alibaba Unveils Powerful New AI Chip and Plans Model With Up to 10 Trillion Parameters

Alibaba is escalating China’s push for greater independence in artificial intelligence infrastructure, unveiling its new Zhenwu V900 AI chip while outlining plans for a next-generation Qwen model containing between 5 trillion and 10 trillion parameters. The announcements came at the company’s Apsara technology conference in Hangzhou, where Alibaba laid out an increasingly ambitious strategy spanning semiconductors, foundation models, cloud infrastructure, and data centers. The company says the Zhenwu V900 delivers roughly three times the performance of its predecessor and can be linked into large computing clusters. Mass production is expected to begin in early 2027.

Alibaba CEO Eddie Wu also said Alibaba Cloud intends to expand its data-center capacity beyond 20 gigawatts by 2032. That matters because the AI race is increasingly being fought across the entire computing stack, not just models. U.S. export controls have accelerated efforts by Alibaba, Huawei, and other Chinese companies to build alternatives to Nvidia hardware and the surrounding software ecosystem. A credible domestic stack would give Chinese AI developers greater insulation from future semiconductor restrictions while creating another large-scale technology platform competing for developers, cloud customers, and enterprise workloads.

Why It Matters: Alibaba is positioning itself as a full-stack AI company that can compete across chips, models, cloud computing, and the massive data-center infrastructure needed to connect them.

Source: Associated Press.

Newsom Signs Seven California Laws Forcing AI Data Centers to Pay for Power and Water

California Governor Gavin Newsom signed a seven-bill package that requires large data centers to disclose electricity and water use, pay for grid and water-system upgrades, and face tighter local review. The laws direct the Public Utilities Commission to create a separate rate class so transmission, generation, and wildfire-related costs are not shifted onto households. Developers must report projected water consumption before local approval, and projects lose shortcuts around environmental review. Newsom said communities should stay “in the driver’s seat” and that firms profiting from AI infrastructure should not do so at residents’ expense.

The package is the most aggressive U.S. state answer yet to the AI boom’s physical footprint. Hyperscalers have been racing to lock land, substations and water in California even as bills and drought politics harden. By forcing operators to fund their own interconnects and publish usage data, Sacramento is raising the all-in cost of new campuses and giving cities a paper trail for opposition. Startups that rent GPU cloud may not write those checks directly, but they will feel them in higher compute prices if California sites slow or become more expensive than Texas, the Midwest, or overseas.

Why It Matters: The largest U.S. tech state just turned data-center power and water from a local zoning fight into statewide cost and disclosure rules.

Source: Reuters.

Meta Hotfixes Muse Mac Flaw That Could Turn the AI Agent Into a Local Backdoor

Security researcher Patrick Wardle disclosed a zero-day in Meta’s Muse Mac app that let any local process change an undocumented dictation setting and redirect voice traffic to an attacker-controlled server. Once redirected, Muse could expose prompts, authentication tokens, and whatever permissions the user had granted — files, camera, microphone, calendars, and linked accounts. Wardle’s proof of concept showed the agent writing files and taking photos without clear user alerts. He argued that malware with limited rights could borrow Muse’s much broader privileges instead of building a full stealer.

Meta Superintelligence Labs called the issue a local privilege escalation, not a remote break-in, and said a hotfix had shipped. That distinction matters, but it does not close the larger design problem. Muse is being sold as a personal operator that books appointments, shops, and handles customer service. Those jobs require standing access. If an agent holds keys to email, payments, and documents, every local setting, cloud transcription path, and integration becomes part of the attack surface. Startups copying the agent model now have a concrete case study: shipping a highly privileged assistant before locking down undocumented configuration is a product risk, not just a patch ticket.

Why It Matters: The first serious Muse bug shows that agent privilege, not model cleverness, is the new endpoint-security problem.

Source: Ars Technica.

AMD Joins the $1 Trillion Club as AI Chip Demand Reshapes the Semiconductor Market

AMD has crossed a $1 trillion market capitalization for the first time, marking a remarkable change in the competitive landscape for semiconductor companies riding the AI infrastructure boom. Shares jumped about 9.6% Monday to a record $613.31, according to Reuters, putting AMD alongside Nvidia, Broadcom, and Micron among chipmakers whose valuations have been propelled by expectations for enormous AI computing demand. AMD shares have risen roughly 185% in 2026, far outpacing the broader Nasdaq.

The milestone reflects more than enthusiasm for individual accelerator chips. AMD has been moving toward complete AI computing systems, pairing accelerators with server CPUs, networking, software, and rack-scale infrastructure as hyperscalers and AI labs look for additional suppliers. Nvidia remains the dominant force in AI accelerators, but soaring infrastructure spending is creating enough demand for multiple semiconductor architectures and suppliers. AMD has also continued taking share from Intel in server processors, giving it another foothold inside the same data centers now being redesigned for AI. The valuation puts additional pressure on AMD to turn investor expectations into sustained revenue growth as cloud giants’ capital spending reaches unprecedented levels.

Why It Matters: AMD’s trillion-dollar milestone shows how AI infrastructure spending is creating enormous value beyond Nvidia and turning the semiconductor sector into one of technology’s most important competitive battlegrounds.

Source: Reuters.

FAA Starts AI Airspace Tool at Washington-Area Airports

The Federal Aviation Administration (FAA) began using an AI system called SMART at Washington-area airports to help manage traffic flows, POLITICO reported. The tool ingests about 200 streams of flight data to forecast congestion and flag conflicts before they appear on a controller’s scope. Officials described it as a prediction aid, not an autopilot for the national airspace. The D.C. region is a high-visibility first site because it mixes commercial traffic, restricted airspace, and political scrutiny.

Aviation has used algorithms for years, but an AI layer that claims to see conflicts earlier changes staffing, liability, and procurement. If the model is right, delays fall, and controllers get time back. If it is wrong in a busy corridor, the agency will have to explain why it trusted a learning system near the capital. Vendors selling air-traffic, airport, and defense software now have a federal reference customer. Startups in that market should expect longer certification cycles, not a sudden open door, because the FAA cannot treat a missed prediction like a bad product recommendation.

Why It Matters: Putting AI into live U.S. air-traffic management makes aviation one of the first critical-infrastructure domains where model error has immediate physical consequences.

Source: POLITICO.

Wall Street Is Getting More Selective as AI Companies Flood the Bond Market With Debt

The AI boom is increasingly moving beyond venture capital and corporate cash reserves into global debt markets. Investors buying investment-grade corporate bonds are becoming more selective about debt issued by AI-heavy companies as borrowing tied to massive data-center construction continues to rise. Reuters reports that spreads on AI-related corporate debt have widened relative to the broader investment-grade market, even as traditional financial and industrial issuers continue to attract strong demand.

That shift could become an important constraint on the next stage of AI expansion. Hyperscaler debt issuance is projected to reach as much as $420 billion in 2027, according to estimates cited by Reuters, as companies finance GPUs, power infrastructure, networking equipment, cooling systems, land, and increasingly their own semiconductor programs. For several years, capital availability appeared almost unlimited for the largest AI infrastructure projects. Bond investors are now starting to differentiate between companies based on expected returns from those investments. If borrowing costs rise meaningfully, AI infrastructure plans that looked attractive when capital was abundant may face greater scrutiny. The result could favor companies with established cloud revenue and strong balance sheets while making highly leveraged infrastructure bets more difficult.

Why It Matters: The cost of capital is becoming another potential bottleneck for AI, alongside chips, electricity, land, and data-center construction capacity.

Source: Reuters.

Google Confirms Gemini AI Models Accidentally Breached Three Real Companies During Security Tests

Google has confirmed that experimental Gemini models accessed systems belonging to three real companies during a cybersecurity exercise in May after a testing environment was mistakenly connected to the open internet. The test, conducted by cybersecurity firm Irregular, was designed as a closed “capture the flag” exercise in which the AI models were supposed to retrieve information from simulated companies. A configuration error instead gave the models internet access, letting them interact with real corporate systems.

In one case, Gemini reportedly guessed passwords until it gained access to a company service. In two others, the models found credentials that companies had inadvertently exposed in public software repositories. Google says the models stopped after recognizing that they had reached real systems, and the affected companies were later notified. The incident is nevertheless significant for AI security because autonomous models increasingly have access to browsers, terminals, APIs, code repositories, and corporate applications. A simple infrastructure misconfiguration can therefore turn an AI evaluation into real-world activity. As autonomous agents become more capable, containment may depend as much on system architecture and permission boundaries as on the behavior encoded into the models themselves.

Why It Matters: The Gemini incident shows that capable AI agents do not need malicious intent to create security problems when they are given powerful tools and poorly configured access.

Source: Ars Technica.

Australia’s Central Bank Governor Says AI Has Yet to Deliver the Productivity Boom Many Expect

Reserve Bank of Australia Governor Michele Bullock has pushed back against increasingly optimistic assumptions about AI’s economic benefits, saying there is not yet clear evidence that the technology is improving national productivity. Speaking Tuesday, Bullock said AI could even be part of a speculative bubble and noted that businesses often see lower productivity as they adapt to major new technologies. She also pointed to the rising cost of data-center investment as one factor contributing to economic and inflationary pressures.

The comments matter well beyond Australia because governments, investors, and technology companies are increasingly basing long-term growth assumptions on substantial productivity gains from AI. Hundreds of billions of dollars are being invested in chips, data centers, software, and energy infrastructure on the expectation that AI will eventually allow businesses to produce more with fewer resources. Bullock’s caution highlights an unresolved question: whether AI’s extraordinary capital requirements will be matched by measurable output gains. History suggests transformative technologies can take years before their economic benefits become visible. The current spending cycle, however, is occurring at a scale that leaves less room for disappointing returns.

Why It Matters: AI’s economic case ultimately depends on productivity, and policymakers are beginning to ask whether the enormous infrastructure investment is translating into measurable gains.

Source: The Guardian.

UK Data-Center Fight Intensifies as OpenAI Warns Local Opposition Could Undermine AI Sovereignty

Former UK Chancellor George Osborne, now OpenAI’s head of AI for countries, has warned that resistance to new data-center construction could leave Britain dependent on foreign computing infrastructure. Osborne argued that opposition to large facilities over electricity consumption, water usage, and local environmental effects could limit the UK’s ability to develop AI systems domestically. The warning comes as data centers increasingly move from an obscure piece of internet infrastructure into a contentious issue involving energy policy, land use, national security, and local politics.

Britain has designated data centers as critical national infrastructure, reflecting their importance to cloud computing, communications, financial services, and AI. But communities in several regions have challenged proposed developments, while environmental groups have questioned whether large AI facilities can be reconciled with electricity and water constraints. The same debate is emerging across the United States, Ireland, the Netherlands, and other markets experiencing heavy data-center investment. Countries seeking domestic AI capacity therefore face a difficult trade-off: attracting billions of dollars in infrastructure investment while ensuring projects have sufficient energy, community support, and demonstrable economic benefits.

Why It Matters: The global AI race increasingly depends on something much less glamorous than algorithms—whether countries can actually build and power the physical infrastructure required to run them.

Source: The Guardian.

BigCommerce Warns Merchants After Third-Party App Breach Injects Malicious Code Into Online Stores

BigCommerce has notified merchants after attackers compromised credentials connected to third-party Ribon applications and used the access to inject malicious JavaScript into e-commerce stores. The incident adds to growing concern about supply-chain attacks in which criminals target software integrations rather than attempting to compromise every victim individually. By gaining access through a trusted application, attackers can potentially reach many businesses downstream from a single breach.

This attack matters for online retailers because malicious scripts inserted into storefronts can steal customer information, payment details, login credentials, or session data, depending on where the code executes and what permissions it has. SaaS platforms increasingly operate through ecosystems containing analytics tools, marketing services, payment software, customer-support widgets, and dozens of other integrations. Every connection creates useful functionality but also expands the attack surface. The BigCommerce incident follows a wider pattern of attackers compromising software suppliers, developer tools, and credentials in order to reach larger numbers of downstream targets.

Why It Matters: Modern businesses increasingly inherit cybersecurity risk from every third-party application connected to their infrastructure, making software supply chains one of the most consequential weak points in enterprise security.

Source: BleepingComputer.

Hackers Manipulated Operational Systems at Colorado Water Utilities in Critical Infrastructure Attacks

Hackers breached operational technology systems at two small privately operated Colorado water utilities and changed equipment settings, disabled remote access and alarms, and altered pumping cycles, according to state officials. Each affected utility serves fewer than 200 people. Authorities said the disruptions were short-lived and did not interrupt water service or affect public safety, but the intrusions show attackers reached systems that control physical infrastructure rather than merely stealing data.

Operational technology has become an increasingly important cybersecurity concern because many water systems, manufacturers, energy facilities, and municipal infrastructure operators rely on equipment never designed to withstand modern internet threats. Smaller operators can be especially vulnerable because they often lack the security teams and budgets available to large utilities. Details about how the Colorado systems were compromised remain limited, and officials have not publicly identified the attackers. Even so, changing pumping cycles and disabling alarms shows why attacks against industrial control systems are fundamentally different from conventional corporate breaches: digital access can potentially produce physical consequences.

Why It Matters: Cyberattacks are increasingly crossing the line from stealing information to manipulating real-world infrastructure, making OT security a national and global resilience issue.

Source: SecurityWeek.

Dutch Startup Morphotonics Raises €40 Million to Bring Advanced Optics Into AI Data Centers

Dutch deep-tech startup Morphotonics has raised €40 million to expand its nanoimprint lithography technology from advanced displays and smart glasses into optical components for data centers. Investors include 3M Ventures, Innovation Industries, BOM, Invest-NL, the European Innovation Council Fund, the European Investment Bank, and family office Ernij Next. The company has spent more than a decade developing manufacturing systems that imprint extremely small optical structures onto materials at commercial scale.

Morphotonics is best known for technology used to produce waveguides—the optical components that direct images into a wearer’s field of view in augmented-reality glasses. Its next machine is expected to produce more than six million waveguides annually. But the potentially larger long-term opportunity may be data centers. Morphotonics is moving into co-packaged optics, where light rather than conventional electrical connections can move enormous amounts of information between computing components. As AI clusters grow, transferring data between accelerators is becoming one of the biggest constraints on performance and energy efficiency. That is creating opportunities for specialized photonics and manufacturing startups far beyond the GPU itself.

Why It Matters: AI’s infrastructure bottleneck is shifting from pure computing capacity to moving data efficiently between chips, opening a new market for optical and photonics startups.

Source: TechCrunch.

Belgian AI Recruiting Startup Spott Raises $21 Million as Investors Back AI-Native Enterprise Software

Belgium-based recruiting technology startup Spott has raised a $21 million Series A led by Balderton Capital, with participation from Base10 Partners, Y Combinator, and Fortino Capital. The financing brings the startup’s total funding to roughly $24 million. Founded in 2024, Spott is developing recruitment software built around AI from the beginning rather than adding generative AI features to a conventional applicant-tracking system.

Spott automatically turns information from emails, conversations, documents, and other unstructured sources into structured recruiting data while automating administrative tasks recruiters traditionally handle. The company says revenue has grown more than tenfold since the start of 2026 and that more than 80% of new business arrives through inbound demand. The funding reflects a broader shift in enterprise software investing: capital is increasingly moving toward startups that argue existing SaaS categories can be rebuilt around autonomous workflows rather than simply enhanced with copilots. Recruiting is a particularly attractive target because the industry involves large volumes of repetitive communication, candidate data, scheduling, CRM updates, and document processing.

Why It Matters: Spott represents the emerging generation of startups trying to replace traditional SaaS workflows with software built around AI agents from day one.

Source: TechStartups via Spott.

Amazon’s Delivery Drones Are Testing the Limits of Consumer Acceptance in Texas

Amazon’s Prime Air drone-delivery program is creating growing tension in Richardson, Texas, where residents living beneath busy delivery routes are complaining about frequent flights, noise, and privacy. One resident tracking the activity recorded 52 Amazon drones overhead during a single Friday that she described as relatively quiet, according to reporting on the New York Times investigation. Amazon says noise complaints make up less than 1% of inquiries associated with the program.

The dispute is an early preview of a problem that could become increasingly common as autonomous delivery moves from limited trials into everyday infrastructure. Drone economics improve when operators can fly frequently across dense neighborhoods, but those same flight volumes make the technology more visible—and potentially more disruptive—to residents. The question is no longer simply whether autonomous delivery works technically. Companies must also determine how much noise, visual presence, and perceived surveillance communities will tolerate in exchange for packages arriving within minutes. Amazon, Walmart, and other logistics companies see drones as a way to reduce last-mile delivery times and costs, making local acceptance a potentially important barrier to scaling.

Why It Matters: Drone delivery may be technologically ready before communities are socially ready, making public acceptance as important as aviation approvals in determining how quickly the market grows.

Source: The New York Times.

Google Opens Preorders for $899 Googlebook AI Laptops Aimed at MacBook Buyers

Google opened preorders for Googlebook, a new laptop line that runs Android with desktop ChromeOS elements and on-device Gemini features. Launch partners Acer, ASUS, Dell, HP and Lenovo are offering five models starting at $899, below the entry MacBook Air. The machines use Intel Core Ultra or Qualcomm Snapdragon X Elite chips, OLED or high-resolution touch panels, and neural processors Google says can handle local Gemini tasks. A year of Google AI Pro, including Gemini Advanced and extra storage, is bundled. U.S. retail availability is set for October 4, with Canada, the U.K., Ireland, France, Germany and Australia the next day.

Googlebook is not a renamed Chromebook. Google is trying to give Android phone users the same continuity Apple sells between iPhone and Mac: app casting, shared accounts and an assistant that lives on the device rather than only in a browser tab. Ten years of software updates is the retention hook. The risk is familiar. Google has entered and exited hardware categories before, and Windows plus Apple still own most professional workflows. If Gemini on a laptop is only a slightly faster chatbot, the category will stall. If it becomes the place where agents run local tools, Googlebook is a direct shot at the PC default.

Why It Matters: Google is turning Gemini into a reason to buy new hardware, reopening the consumer PC fight around on-device AI rather than browsers.

Source: Engadget.

UAE’s Space42 and Viasat Commit $1 Billion to Satellite Network Aimed at Eliminating Mobile Dead Zones

Abu Dhabi-based Space42 is working to extend ordinary smartphone connectivity into areas beyond conventional cellular networks through direct-to-device satellite technology. Space42 and U.S.-based Viasat have committed about $1 billion to Equatys, a joint venture expected to begin offering global satellite connectivity services from 2029. The goal is to allow standard mobile devices to maintain connections in deserts, mountains, offshore areas, and other regions where building terrestrial 4G or 5G infrastructure is economically difficult.

The project reflects a rapidly developing competition to merge satellite networks with the global mobile ecosystem. SpaceX’s Starlink, AST SpaceMobile, Amazon’s satellite ambitions, established operators, and telecom carriers are all exploring ways to connect smartphones directly from orbit. Space42 is also building the Foresight Earth-observation constellation, using synthetic aperture radar satellites that can image through clouds, smoke, and darkness. Three of the first five Foresight spacecraft were assembled and tested in the UAE, underscoring the country’s broader effort to develop sovereign space capabilities rather than relying entirely on foreign infrastructure.

Why It Matters: Direct-to-device satellites could erase the boundary between terrestrial mobile networks and space connectivity, opening a global market measured in billions of phones rather than specialized satellite terminals.

Source: The National.

Physical AI Startup TacnIQ.ai Raises $1.5 Million to Teach Robots the Sense of Touch

Singapore- and California-based TacnIQ.ai has raised $1.5 million from In Group Holdings as part of a planned $3 million pre-seed round to develop foundation models built around tactile data. Unlike most AI systems, which learn primarily from text, images, audio, and video, TacnIQ is collecting information generated through physical contact. The company says its database contains more than 5,000 hours of tactile interactions gathered from controlled experiments and commercial deployments.

The startup wants robots and intelligent machines to interpret pressure, movement, force, and other physical signals so they can respond more effectively in uncontrolled environments. Potential applications include logistics, construction, healthcare, hospitality, and industrial work where vision alone is insufficient. This is part of a broader move toward “physical AI,” as developers try to give robots foundation models that can reason about real-world environments rather than training a separate model for every task. The challenge is data: the internet contains enormous quantities of text and images, but high-quality tactile datasets are comparatively scarce. Companies that build proprietary physical-world datasets could therefore gain an important advantage.

Why It Matters: Robotics may require an entirely new class of foundation models trained on touch, motion, and physical interaction—not simply the text and images that fueled today’s generative AI boom.

Source: TNGlobal.

OpenAI Says Internal AI Model Has Resolved More Than 100 Open Math Problems and Creates Independent Advisory Group

OpenAI says a new internal model has resolved more than 100 long-standing open problems across multiple areas of mathematics, prompting the company to establish an independent advisory group of prominent mathematicians to help evaluate and communicate the results. OpenAI says training began on August 28. The company has also claimed that the system solved the Navier-Stokes Millennium Prize problem, although it has not yet publicly released the broader set of more than 100 claimed solutions for independent verification.

The newly formed Advisory Group on Mathematics and Artificial Intelligence says it operates independently from AI companies, does not accept payment for its work, and intends to publish its recommendations publicly. Its role includes advising AI developers on interactions with mathematical researchers and on responsibly presenting and releasing potentially significant results. The development matters because advanced AI systems are increasingly moving beyond answering established questions toward generating claims of original scientific and mathematical discovery. That creates new challenges around verification, attribution, peer review, responsible disclosure, and the role of human researchers. AI-generated proofs still need rigorous independent examination before extraordinary claims can become accepted mathematical results.

Why It Matters: If independently verified, AI systems capable of solving large numbers of previously open mathematical problems would represent a major shift from AI as a knowledge tool to AI as a participant in fundamental research.

Source: OpenAI and the Advisory Group on Mathematics and Artificial Intelligence.