Tech, in brief

OpenAI said on Monday that one of its internal AI models has disproved the Exponential Time Hypothesis, an idea about computing that most experts believed true. [1][2]

The hypothesis is about how hard it is to tell whether a long list of yes-or-no conditions can all be true at once. [1] Hundreds of papers lean on it to show that problems in network maths, biology and databases are hard to solve quickly. [1]

OpenAI's paper gives a method that counts the answers to such puzzles faster than the hypothesis said was possible, OpenAI says. [2]

OpenAI says the model found it on 16 October, and Lean, a program that checks every step of a proof by machine, passed it the next day. [1] The write-up and the Lean proof are public on GitHub, a code-sharing site. [1]

Russell Impagliazzo and Ramamohan Paturi of UC San Diego, who helped propose the hypothesis in 1999, and Ryan Williams of MIT will check it. [2] They wrote that taking part 'should not be taken as certification of the claim'. [1]

Williams says the three signed no secrecy agreement and are not being paid. [1] They asked OpenAI to release all its raw data and to pay for a conference to examine the claim, and OpenAI agreed. [1]

OpenAI also claims a maths result: a proof of the local smoothing conjecture, set by Christopher Sogge in 1991, in every dimension. [1] That would also settle the Kakeya conjecture, which Hong Wang solved in three dimensions and won a Fields Medal for a few months ago. [1][2]

Its third claim is in physics: a proof that theories describing quarks and gluons, tiny particles inside atoms, exist in four dimensions. [1] Part of it touches one of the Clay Millennium Prize problems, famous unsolved questions. [1] A group led by David Gross, a Nobel physicist, has kept that material back to write it up itself. [1][3]

Alex Iosevich, a mathematician at the University of Rochester, called the maths result a 'historic development'. [3] He also said he felt 'a twinge of sadness'. [3]

A day earlier AGMAI, a panel that advises AI firms on maths, said labs should not hand results to chosen experts first. [2] It named unequal access and borrowed reputation as the problems. [2] OpenAI admits its choice of experts 'creates asymmetry of information'. [3]

More than 1,000 mathematicians have joined a group that pledges not to work for AI companies, after OpenAI posted 722 maths papers at once on 6 October. [4][1]

'If the claim is correct, it will change an important part of theoretical computer science,' the three checkers wrote. [1]

Why it matters. The next test is the open check led by Impagliazzo, Paturi and Williams, with the conference OpenAI agreed to fund, and no date has been set for its findings. [1]

OpenAI released GPT-6.1 Terra on Monday, a model in the middle of its range, below Sol and Astra and above Luna. [5][6] Developers pay $1 per million tokens, small pieces of words, going in and $5 per million coming out, half the price of the Terra it replaces. [6] OpenAI says Terra matched GPT-6 Astra, its strongest model, on one test while costing a tenth as much per token. [7] It also says Terra did tasks for about a ninth of the cost of Anthropic's Claude Opus 5.5, in its own tests, which nobody outside has checked. [6]

Why it matters. Terra is built for high-volume office work such as reading invoices into accounting software. [5][6] A data scientist quoted by The Register says prices keep falling because the firms that run AI models compete with each other. [5]

Texas sued OpenAI and its boss, Sam Altman, on Friday, saying the company sells ChatGPT as safe while its safeguards fail, including for children. [8] The case uses the state's law against misleading trade and tries to make Altman personally liable, which is unusual. [8] Texas wants fines of up to $10,000 per violation and a ban on children using ChatGPT. [8] It also wants a court-appointed expert to approve new models. [8] OpenAI said it works with experts to improve its safeguards and wants rules for the whole industry. [8]

Why it matters. Ken Paxton, the Texas attorney general who brought the case, is in a close race for the US Senate. [8] He has sued or settled with Roblox, TikTok and Meta in the last year, and Meta paid Texas $1.4 billion in 2024 over face recognition. [8]

Apple said on Monday it knows of 'an issue impacting the inner display' on a small number of iPhone Duos, its first folding phone. [9][11] Owners whose screen fails can get a replacement through AppleCare, Apple's repair service. [10] Since the phone went on sale on Friday, owners had posted pictures of bright lines and cracks along the fold. [12] Apple did not say what causes it or how many phones are hit. [10][11] Outside the warranty, an inner-screen repair costs $999, and the phone starts at $1,999. [10]

Why it matters. Apple told Bloomberg that affected phones show the fault soon after first use. [10] The BBC notes that some pictures of broken phones online may have been made with AI, so the real number is hard to judge. [11] A phone that CNET tested has folded more than 55,000 times so far without breaking, Engadget reports. [13]

More than 300 of Boston Dynamics' roughly 1,400 staff signed a letter asking bosses to stop a sale of its Spot robot dogs to ICE, the US immigration agency. [14] Boston Dynamics is owned by Hyundai, the Korean carmaker. [14] A $1.3 million US contract from 25 September pays Radeco, a Connecticut firm, for four Spots it has fitted to detect gas, radiation and explosives in tunnels. [14] The company stands by the deal and says its terms ban using Spot as a weapon. [14]

Why it matters. The letter points to a 2025 raid that held 475 mostly South Korean workers at a Hyundai factory in Georgia. [14] Staff fear the deal will make hiring harder while the company plans 1,250 new jobs by 2033. [14]

The Dutch government wants to cut 25,000 full-time jobs, about 16% of its workforce, by 2035. [15] It aims to save 900 million euros by 2027, rising to 3.5 billion euros by 2035. [15] Police and defence are left out. [15] The cabinet says it will set goals for using AI to work more productively in early 2027. [15] So how AI is meant to replace the work is not yet clear. [15]

Why it matters. Eric van der Burg, the state secretary in charge, told parliament he wants a 'simpler, smart and smaller' government. [15] The first savings target falls in 2027, the same year the AI plan is due. [15]

A Danish ministry's test of swapping Microsoft Office for Collabora, a British open-source office suite, ended in under six months, Computerworld Denmark reported. [16] Open-source means its code is free for anyone to use and change. The test began in 2025 with 38 staff, under a minister who wanted less reliance on American technology. [16] Staff complained about formatting, blank pages and tracking changes. [16]

Why it matters. France and the Netherlands are also moving away from US software. [16] A Copenhagen Business School professor says the Danish ministry never properly studied the test, so other public offices cannot learn from it. [16]

Volkswagen has a licence to test Level 3 self-driving, where the car drives itself in some conditions with a driver ready to take over, on public roads in Hefei, China. [17][18] Xpeng, a Chinese carmaker, may now run Level 4 robotaxis across Guangzhou, at first for riders with invitation codes. [17] China stopped issuing new robotaxi licences in April after an incident with a Baidu robotaxi in Wuhan, and started again in July, Bloomberg reported. [17]

Why it matters. It is a test permit, so VW cannot sell Level 3 cars yet, and it aims to have the technology ready in 2027. [18] Xpeng wants to carry riders with no safety driver in 2027 and is talking to European ride-hailing firms. [17]

Chinese research ships spend more time near undersea internet cables in the Pacific than ships from any other country, a report from CSIS, a US research group, found. [19] It compared 200 research ships from 47 countries. [19] Between July and October 2024 one Chinese ship spent time near three cables linking the US with Australia, Indonesia, Singapore and Guam. [19] The ships moved back and forth in a pattern used to map the sea floor. [19]

Why it matters. The report says such maps could help sabotage or submarine missions, though the ships are not part of China's navy. [19] A report last week counted 500 cases of Russian ships watching cables over five years. [19]

A Chinese team says it sent 3.2 petabits of data a second down one optical fibre cable in real time, a record for this kind of test. [20] A petabit is a million billion bits, and the team compares it to downloading more than 20,000 4K films a second. [20] The same team reached 2.5 petabits earlier this year, so the speed rose almost 30% in a year. [20] The cable has 24 cores, separate glass threads that each carry light. [20]

Why it matters. The test used hardware built to commercial standards, not lab-only equipment, so it is closer to a cable that could be laid. [20] Lab tests on similar fibre have reached 5.29 petabits a second. [20]

Lip-Bu Tan, Intel's chief executive, has spoken with Lisa Su, who runs AMD, about working together, Tom's Hardware reports. [21] The two companies have competed for decades to make the main processors in PCs and data centres. [21] At a talk on 24 October, Tan said Intel's 18A chip-making process has reached volume production. [21] He said its next process, 14A, should be ready for volume production in early 2027. [21]

Why it matters. Tan wants other companies to have their chips made in Intel's factories, but he named no new customers. [21] Neither company has announced any deal from the talks. [21]

Micron, a US maker of memory chips, set out how it will keep shrinking DRAM, the working memory inside computers. [22] Its next process should enter volume production in the second half of 2027, with about a quarter more memory in the same space. [22] After that it plans wafer bonding, joining chip layers together, then High-NA EUV, the newest machines for printing the finest chip patterns. [22] 'DRAM is nearing atomic limits,' its technology chief, Scott DeBoer, said. [22]

Why it matters. Micron is also designing new kinds of memory to sit directly on top of processors for AI work. [22]

Hobbyists found a way to remove a limit Nvidia puts on some AI maths in its RTX 20, 30 and 40 gaming graphics cards, Tom's Hardware reports. [23] Nvidia's own specifications run one kind of this maths at half speed on these cards, but not on its pro cards. [23] With the limit removed, an RTX 4090's peak rate doubled, from 165.6 to 331.6 trillion operations a second. [23] The trick uses a flaw in Nvidia's own security firmware. [23]

Why it matters. Real tasks gain much less than double, and chatbots reply no faster. [23] The projects warn that the unlock can damage the card. [23]

IBM announced IBM Z Atlas, a coming AI tool for its Z mainframes, the large computers it sells for heavy business work. [24] The tool will map how the parts of a customer's system connect and test changes before they go live. [24] IBM says untested changes are the leading cause of unplanned outages on these machines. [24] People, not the AI, keep the final say, IBM says. [24]

Why it matters. IBM's chief executive said last week that its newest Z17 mainframes have outsold every earlier generation by $1 billion. [24] Rivals pitch AI for moving mainframe programs onto other machines, while IBM wants AI to keep customers on its own. [24]

Garrett Langley, chief executive of Flock, whose cameras read car number plates for police, told police chiefs on Sunday that protests against his company are not real, 404 Media reports from leaked audio. [25] He claimed, without evidence, that most online criticism of Flock came from Russian bots. [25] He said online chatter about Flock is down about 95% since July. [25]

Why it matters. He repeated a disproven theory that Ukraine's drone strikes on Russian data centres cut bot activity. [25] He also said his wife had advised him never to apologise. [25]

Apple announced a new iPad mini with a larger 8.5-inch OLED screen, where each tiny dot makes its own light, and a faster A20 Pro chip. [26] It starts at $699, or $849 with a mobile connection. [26] Apple says it runs AI tasks up to 4.5 times faster than the last model. [26] Orders open on 28 October and it goes on sale on 6 November. [26]

Why it matters. It arrives two days before Apple's 'Pro and Tell' event on Wednesday. [26] It runs the new Siri that Apple released for iPhones the same day. [27]

Apple released iOS 27.2 on Monday, the first big update to its newest iPhone software. [27] Siri can now recall earlier conversations and look at voicemails, Safari history and photos in messages. [27] Users can choose which apps it may read. [27] The test version of Siri's AI now works in French, Japanese, Korean, Portuguese and Spanish. [27] FaceTime calls can now use the front and back cameras at once. [27]

Why it matters. Deaf users can now add a sign-language interpreter or a live captioner to a video call. [27] The same update lets the iPhone Duo use an Apple Pencil. [27][10]

Instagram is testing long, widescreen videos it calls Features, of up to 90 minutes. [28][29] The test starts small in the US and with some creators in Brazil. [28] Adam Mosseri, who runs Instagram, said he had been against long videos 'for a long time'. [29] He worried that long viewing means fewer photos and Reels sent between friends. [28]

Why it matters. The move is aimed at YouTube, which has become a big hit on TV screens. [30] Unlike YouTube, Instagram does not yet share ad money with creators, and says it is too early to name a payment plan. [29]

T-Mobile, a large US phone network, launched T-Mix plans that bundle phone service with picks such as Netflix, YouTube Premium or Xbox Game Pass. [31] The cheapest plan, at $55 a month, gives one pick, which can be changed up to four times a year. [31] It streams video at 480p, a low quality, while the $75 and $95 plans stream up to 4K. [31] Some picks can only be swapped once a quarter or less often. [31]

Why it matters. Video quality is one of the things that separates the cheap plan from the dearer ones. [31] Picks can also be traded for money off a new phone, CNET noted. [31]

Turo, a site where people rent out their own cars, will show which Teslas have Full Self-Driving (Supervised) switched on, from 28 October. [32][33] The software handles most driving but still needs a driver watching. [32] Hosts get one free month, then pay Tesla's usual $99 a month. [33] Tesla says about 1.7 million customers have paid for it, out of more than 10 million cars it has sold. [32]

Why it matters. Electrek notes that Full Self-Driving subscriptions count toward Elon Musk's pay package. [33] Turo says thousands of Teslas with the software on are already listed. [33]

Bedrock Robotics, a two-year-old start-up, raised $600 million at a value of $4.1 billion. [34] It sells a kit that makes existing excavators, the big digging machines, drive themselves. [34] Its machines now work on real sites, including a water treatment plant in Nevada. [34] Its chief executive, Boris Sofman, once ran the self-driving lorry project at Waymo, Google's sister company. [34]

Why it matters. It takes five years or more to master an excavator, Bedrock says. [34] A builders' trade group expects more than 40% of US construction workers to retire in the next five years. [34]

Entire, a coding platform from Thomas Dohmke, the former head of GitHub, opened to all developers on Monday. [35] It is built for the flood of AI-written code that, ZDNet says, has caused growing outages at GitHub. [35] Entire replaces the usual step where people read each change with 'Trails', where AI agents score every change. [35] Developers can merge code when the checks pass 'without needing to look at the code itself', the company says. [35]

Why it matters. Entire claims its code hosting is up to 89 times faster than GitHub's. [35] The design lets a team merge code that no person has read. [35]

Cadence, one of the two biggest makers of software for designing chips, unveiled Maestro, AI models trained for chip-design work. [36] It claims up to 30% better results and up to five times fewer tokens than a general AI model. [36] Some customers get early access late this year and can train the models on their own designs. [36]

Why it matters. Its rival Synopsys built its model to run on OpenAI's computers without using customer data, so the two now differ on where a customer's designs go. [36] A Cadence executive said the cost of running AI agents had 'really started to explode'. [36]

The RealReal, a site that resells designer clothes, made it harder for AI agents to log in, fill carts and buy. [37] The Verge had shown that Meta's Muse agent could re-add an item to a cart each time the shop's 20-minute hold ran out. [37] That blocked other buyers for hours. [37] Agents can still find listings and send people to them, the company's technology chief said. [37]

Why it matters. Meta says shopping is the most common use of Muse, with purchases over $200 on average. [37] Amazon has also blocked agents from shopping on its site. [37]

Fanatics, which sells official sports clothing, built a 'Shop the Look' feature with Codex, OpenAI's coding tool. [38] It suggests matching team gear that is in stock on each product page. [38] Fanatics sells more than 1.5 million different items across hundreds of websites. [38] It says the feature raised how much shoppers engage with products by 10%. [38]

Why it matters. Team merchandise is still Fanatics' biggest source of money. [38] It makes more than 45 million items a year and sells in more than 180 countries. [38]

Torsten Slok, chief economist at Apollo, a large investment firm, says AI has not yet raised profits for US companies outside the tech sector. [39] Revenue per worker is rising at the seven biggest tech firms, such as Apple, Nvidia and Tesla, but not across the rest of the S&P 500, a list of 500 large US companies. [39] Apollo also sees growth in AI spending slowing, as firms switch to cheaper models when prices rise. [39]

Why it matters. Some critics answer that computers also took a decade to show up in productivity figures. [39] If firms stay unconvinced, the data centres being built to serve them would earn less, The Register notes. [39]