Tech news picked on 2026-10-19
Small model further trained to reduce AI-style writing, can run on a laptop
- Qwen3-4B model was further trained to reduce writing habits that look AI-generated. According to the creator, it uses fewer dashes, fewer 'not X but Y' sentences, and fewer lists grouped in threes, and it can be used without a separate system prompt.
- The creator's own rule-based detector scored it at 1000 per word, dropping from 121 for the original model to 12.5. This score is based on the creator's regex criteria, and no blind test with readers has been done yet.
- The creator said the training data included 161 answers generated on a MacBook Air. The creator also reported that it got all 30 answers right in a quiz of 30 factual questions, and this is the creator's own measurement.
- The original post says it can be run in Ollama by downloading the Q4_K_M quantized file (about 2.5GB). The original weights and the LoRA adapter are also published.
Wyze's new scale moves the display to the handle, making readings easier at eye level
- Wyze unveiled the Ultra BodyScan Plus, which moves a 4.3-inch color display onto a foldable handle to show measurements close to eye level, according to the source.
- The price is 149, about 30 more than the previous Ultra BodyScan (119.98). According to the source, the earlier model measured 13 items, including weight, BMI, body water and muscle mass, and this one expands that to 50 or more, including visceral fat, protein and basal metabolic rate.
- It provides analysis of arms, legs and torso using BIA (bioelectrical impedance analysis, a method that passes a weak current through the body to estimate body composition). The source says it uses an ITO (indium tin oxide) coated glass surface so measurements work regardless of foot position. These figures and accuracy claims are Wyze's own, and the source contains no independent verification.
- According to the source, data can be sent to Apple Health, Google Health and Android Health Connect without a subscription, either by Bluetooth through a smartphone or directly to the cloud over home Wi-Fi. The source says competitor Withings requires new users to pay an annual subscription of 99.50.
- Up to 8 users can be registered, and guests can take one-time measurements without limit. There is also a mode for weighing children, pets or yourself while holding items, but detailed health metrics are not provided.
- The original article reports that Apple revised its App Review guidelines to ban apps that use AI to create offensive content. Nude-synthesis apps are apps that create images of people in photos with their clothes removed.
- According to the original article, Apple bans apps that appear to be ordinary apps during review and then reveal nude-synthesis features after approval. This tactic is called bait-and-switch. The policy also says apps that add features after submission that were not present at submission will be removed.
- An Apple blog post says apps that handle user-uploaded content must filter out offensive material and provide a reporting function and a way to block abusive accounts. The original article does not confirm how strictly this will be enforced.
- According to the original article, Apple blocked more than 150,000 apps from release during review last year and removed 190,000 3thousand developer accounts over fraud and abuse. The city of San Francisco ordered Apple and Google to remove nude-synthesis apps in July.
Mistral releases Timestral, a model that forecasts future values from past numbers
- Timestral is a family of models that looks at numbers recorded in time order, such as equipment sensor readings or daily sales, and predicts future values. It comes in four sizes, from 5M to 209M parameters, and instead of a single predicted value it shows a range of uncertainty using several quantiles.
- Mistral says its largest 209M model finished 1 place on three public benchmarks: GIFT-Eval, TIME, and fev-bench. This is the developer's claim. The comparison was made among zero-shot models that are used without additional training, using a single checkpoint for each.
- According to the source, on a 8-thread server CPU, forecasting one day ahead from 3 series recorded at 15-minute intervals took 10ms for the 5M model and 140ms for the 209M model. These speed figures are also conditions measured and reported by Mistral, and the source explains that inference can run without a GPU.
Maeving RM2 electric motorcycle: style and quiet ride are good, but finish and price disappoint
- This motorcycle runs on an electric motor inside the rear wheel instead of an engine and chain. Reviewer Tim Stevens named its quietness, with almost no drivetrain noise, as its biggest advantage.
- The rated top speed is 70mph. The fastest figure the reviewer recorded in the strongest mode was about 63mph, leaning forward pushed it up to 67mph, and 70mph was reached only downhill. The measurement was based on the reviewer's own riding feel.
- The official range is 80 miles per charge. The reviewer thought the figure was only likely in the default mode (limited to 45mph), and wrote that in sport mode the average was about 50 miles per charge.
- The battery is a removable 2.73 kilowatt-hour (kWh) pack, and there are 2 packs. Each pack weighs 36 pounds. The bike's charging port uses a proprietary standard, so you must buy a standard J1772 adapter separately (440 dollars), and this adapter does not increase charging speed. Charging 80% with the basic charger took about 4 hours.
More people are using AI coding tools to fix firmware on music gear themselves
- Vibe coding is a way of describing the function you want in words and having AI generate the code. According to the article, this method has greatly increased the number of hobby developers who modify firmware, the internal programs of instruments such as synthesizers.
- The article's reporter observed that adding a single velocity-sensitive feature to the Korg Volca used to take weeks to months, but now hobby developers turn low-cost devices into grooveboxes in a few days.
- Several custom firmwares have appeared for the M-Vave FM-1, a DX-7 compatible synthesizer priced at about 70. The article reports that some firmware pages on GitHub list Claude as a contributor or disclose AI use, while Baudgirl only says AI is suspected to have been used.
- Elektron warned on its forum that it bans links to custom firmware and that devices can be bricked. By contrast, M-Vave has released the open-source firmware FM1 Open and a development kit.
How the overhead projector went from military technology to a classroom tool
- The overhead projector (OHP) is a device that shines light through transparent film with writing and drawings on it, projecting them onto a screen. The source says this technology was mainly used for military purposes before it entered classrooms.
- The source says this is the final episode of the fifth season of the podcast Version History. Host David Pierce, along with Jake Kastrenakes of The Verge and Stanford University professor Antero Garcia, discusses the development of the OHP and how it spread through classrooms.
- The panelists are introduced as discussing why the OHP disappeared and whether PowerPoint, computer screens, and smart boards are better than the OHP. These are questions the hosts raise; the source does not give their conclusions.
- At the bottom of the source are additional reading links, including a history of the magic lantern, the history of innovation at 3M, and an NPR article on American science education after Sputnik.
Running your own intermediate server for AI model calls lets you see usage records
- The intermediate server that requests pass through when calling an AI model is called an LLM proxy. The author, Jared Siegel, says that if you run this intermediate server yourself, you can check which calls went out, when, and to which model.
- According to the original claims, one proxy can group multiple providers and models, record which key called which model, and let you use separate virtual API keys for each task. These are the author's stated advantages, and the original gives no performance figures.
- The author says that when a cache miss occurs during coding work (a situation where a stored result cannot be used and must be recalculated), he can find the cause by hand within about 20 seconds. He has not automated this yet, but he adds that the proxy has an API, so automation is possible.
- The author does not recommend any specific proxy product. Instead, he advises installing one under your own control and modifying it to fit your needs, and he describes adding a lookup function and an MCP server connection to his own setup.
- Even is an Indian startup founded in 2021 that aims to change a model in which hospital profits grow as tests or admissions increase. The company builds its own hospitals, pays doctors fixed salaries, and said it can pay bonuses based on patient recovery outcomes rather than the number of tests or surgeries.
- In a Series B round led by Khosla Ventures, it raised $50 million, and according to the company, its valuation is $300 million. Cumulative investment exceeds $120 million.
- The company said that the unplanned readmission rate at its own hospitals is below 1%, lower than the industry benchmark of 8 to 15%, and that patient length of stay is 40% shorter. These are figures Even presented, and the source contains no verification from an independent body.
- According to filings submitted to regulators, its net loss for fiscal year 2025 was about 902 million (about $9.5 million), and revenue was about 272 million (about $2.86 million). Reports from the Indian startup outlet Entrackr said the company laid off about 350 people in August.
Linux 7.3 kernel officially released with record-scale changes
- The Linux kernel is the core software that connects computer hardware and programs. According to the source, Linus Torvalds officially released version 7.3, and the merge window for version 7.4 starts the next day.
- The source says this release set records in every metric. Before the merge window, 18,765 commits came in from 3,021 developers, and 807 of them were participating in kernel development for the first time.
- Main changes include extensions to binfmt_misc, group scheduling work aimed at improving performance on large systems, a new filesystem failfs that makes all operations fail, the fchroot() system call, and global per-CPU data for BPF programs.
- It also includes completing sched_ext sub-scheduler features, tracking virtual machine work sets, alternate data stream support for the ntfs3 filesystem, and SMB server improvements. For more details, see two LWN merge window summaries and KernelNewbies' guide page for 7.3.
superlite, an automatic build tool for Claude Code: hand over a spec and receive tested results
- This tool is an extension added to the AI coding tool Claude Code. When a user hands over a spec, it reduces approval questions, multiple AI agents split and build the code, and then testing and review proceed. It is a project created by author Iheatu Chris Wogu.
- According to the original author's benchmark, on the same project evaluated with 177 hidden tests, the average of 3 runs of the default superpowers setup was 33.6 minutes, and the latest superlite run of 1 took 16.1 minutes. The author conducted the measurement, and both approaches reportedly passed 177 of 177.
- Work is split by file and interface, and independent tasks run at the same time. According to the original, the latest run had up to 18 agents working simultaneously. Tests run in the background so progress does not stop.
- The original explains that rules are enforced by hooks, a code mechanism, rather than by prompt wording. It reportedly blocks the coordinating agent from directly editing project code, and records automatic decisions so they can be reviewed later.
Anthropic announces plan to contribute 5% of company shares to a public benefit fund
- Anthropic announced it will contribute 5% of its company shares to a 'public benefit fund (public benefit fund)'. According to the original report, the fund will be set up as a fully owned subsidiary inside the company.
- The company also said it may contribute an additional 2% each time the company's value rises by 50%. This is the company's own statement.
- Examples given for the fund's use include large donations to traditional nonprofit organizations and direct support for specific groups. The original report does not give specific support criteria.
- The company said it has already invested $200 million in studying how AI will affect the job market. It said it will publish more details on the fund's operating structure and decision-making process within the next few months.
Cyberspace trademark filing and William Gibson's objection, as the author recounts
- Author Erik Gullichsen says he built a virtual reality prototype at the Autodesk research lab in 1989. The setup connected a head-mounted display, a position tracker, and a data glove that reads finger movements to a 386 computer, and the source says the system ran at roughly 6~10 frames per second.
- According to the source, Autodesk filed to register 'cyberspace' as a trademark, and William Gibson sent a letter through his lawyer objecting to it. The author recalls that he did not respond to this letter.
- According to the author, the claim that 'Gibson tried to trademark Erik Gullichsen's name' was spread as a prank by Timothy Leary. The source explains that Paulo and Rudy Rucker later added to the story and spread it, and that it was eventually taken as fact.
- At an event in Barcelona in 1990, the source reports that Gibson quoted the idea that a street finds its own uses for things and said, 'Autodesk is no street.'
ServiceNow Unveils AI Agent Permission Controls and Employee Automation Tools
- An AI agent is an AI program that works across several business systems on behalf of a person. ServiceNow announced AI Passport, a digitally signed certificate that lists an agent's identity, permitted actions, accessible systems, and responsible owner. The source says this permission check takes place at the moment work is executed.
- In a demonstration set up by ServiceNow itself, a 3 customer service agent was configured to process only refunds under 25 dollars. Requests for 18 dollars were processed automatically, while requests for 40 dollars went beyond the human review threshold and were rejected at the execution stage, according to the source.
- The update to the employee-facing EmployeeWorks adds features that let staff build helper tools for research, scheduling, and HR tasks without coding, and automatically run repetitive work such as sales preparation in the background. According to the source, helper tools, routines, ToolSmith, and the Conversations API are in limited release.
- According to the source, AI Passport, shadow AI detection, and AI cost management will be available later this quarter, while operational resilience features for financial services are available now. NVIDIA participated as a design and launch partner for AI Passport.
Encryption to prepare for quantum computer threats begins to apply to Google Meet calls
- Google says it is applying post-quantum cryptography (PQC, a cryptographic technology that protects data in ways that future quantum computers would find hard to break) to Meet calls. The claims about the protection effect are Google's own.
- According to the source, the rollout starts with laptop web clients and meeting room hardware, and mobile support is coming soon. It says users will see no visible change.
- In the same week's update, after a Meet meeting ends, next-step items appear on screen to help draft follow-up emails or export to-do lists. The source says this is based on Gemini and Workspace Studio.
- According to the source, Gmail's AI inbox is rolling out to paid Workspace users, and Gmail Live voice search is rolling out to users of the Android and iOS mobile apps.
Apple Opens Pre-Orders for Foldable iPhone Duo on October 16
- Folded, it fits in a pocket. Unfolded, it offers a large screen for watching video, gaming, and using several apps at once. Apple said pre-orders will be taken on apple.com and in the Apple Store app starting at 5 a.m. Pacific Time on October 16.
- Apple claims the inner screen, when unfolded, is 50 percent larger than the iPhone 18 Pro Max screen and that it is the thinnest iPhone. The outer screen, when folded, is 5.4 inches.
- According to Apple, even when folded it provides 90 percent of the screen area of the iPhone 18 Pro. Screen size, thickness, and battery details are all Apple's announcements, and the source contains no measurements from independent organizations.
- Trading in an iPhone 13 or newer through Apple Trade In reportedly earns an instant credit of 175 to 885. Apple Upgrade Lease starts at 57.99 per month based on 24 months, and this price excludes tax and trade-in credit.
Waymo opens robotaxi service to everyone in San Antonio, San Diego, and Tampa with no waitlist
- The source says Waymo has opened its robotaxi service to everyone in San Antonio, San Diego, and Tampa. A robotaxi is a taxi that carries passengers using self-driving technology without a driver.
- The removal of the waitlist means riders no longer need to wait after signing up, and the source identifies this as the key change. The source does not give details on wait times or the exact service areas.
- The source says Waymo received approval to take part in an autonomous vehicle pilot program in Toronto, Canada. The approval is a fact the source reports, but the source does not say whether actual service has started in Toronto.
- Dynamic packages are a way to bundle components or extensions separately so they can be installed and replaced. The author sysrpl announced that full dynamic package support for the Free Pascal compiler and the Lazarus IDE has been released for the first time. This is the author's announcement.
- The original says components can be installed without rebuilding the entire IDE, and that installation finishes within a few seconds. This time is the author's description, and the measurement conditions are not given in the original.
- For distribution, a program can be packaged as a single executable. It can also be split into multiple dynamic packages to create a plugin structure in which parts are replaced while the program is running.
- Limitations: unit names inside a package cannot be the same as the package name, and if unit names overlap between different packages, they cannot be installed together. The package source file automatically gets a Package suffix.
Lenovo Legion Go S SteamOS model discounted to 999.99 at Best Buy
- The Legion Go S is a handheld that runs the same SteamOS software as the Steam Deck OLED. The source says its trigger buttons are designed to press in two stages.
- According to the source, the screen is a 8-inch IPS panel that supports 1080p and up to 120Hz, while the Steam Deck OLED has 800p and 90Hz. The memory 32GB and storage 1TB are double those of the Steam Deck OLED, according to the source. The processor is the Ryzen Z1 Extreme, which the source describes as slightly faster.
- The discounted price the source confirms at Best Buy is 999.99. The source presents this price against the 1579.99 list price model, and introduces the Steam Deck OLED at 789.
- The source author says the Asus ROG Xbox Ally X20, priced at 1299.99, has higher performance thanks to the Z2 Extreme chip and OLED screen. The author's subjective view is that the Legion Go S has lower performance but is not lacking.
Bill Seeks to Make Game Purchases Ownership, Introduced in US House
- Representative Rich Torres of New York 15 district introduced the 'Power to the Players Act', which would require the Federal Trade Commission (FTC) to write rules so that game buyers own the games rather than holding a license to use them.
- According to the article, many games require a server connection, so if a company shuts down its servers, the game cannot be run. The article uses the NBA 2K25 as an example, reporting that 2K removed the game from sale and will shut down its servers on 31 December.
- The bill also requires developers to offer an offline mode before turning off servers. The article says a similar bill, the 'Protect Gaming Act' introduced by Pennsylvania State Senator John Kane, would require one of three things: an offline mode, an online mode that does not depend on servers, or a refund of at least 25 percent.
- The article cites Sony's plan to phase out physical discs by 2028 as background for the debate over ownership and resale.
Linear co-author on the people behind software and the limits of adding features
- Karri Saarinen, writing for Linear, argues that good products carry the experience, interests, and way of working of the people who made them. This means users can feel that point of view in a product even without knowing its makers.
- The original post says Linear values intuition over data or procedure. Here, intuition is not arbitrary taste but judgment that comes from talking with customers and learning through mistakes.
- According to the original post, a single feature can work well on its own and still make the whole product worse. The author notes an impression that many recent products are shifting toward adding features rather than rethinking core problems.
- The author speculates that a pattern where individuals make their own changes together with agents could fuel this trend. The original post offers no research or figures to support this.
- Harness² is a framework that leaves the AI model itself unchanged and instead adjusts the settings around the model (system prompts, memory, tool guidance, skills, and so on) for each task, helping agents perform better. The repository is published under the google-research GitHub organization, and the README states that it is not an official Google product.
- During the task phase, it creates revised settings, applies them to the same task, compares the differences in results, and combines the new settings. According to the README, this process runs only on differences between executions, without correct-answer feedback.
- The README reports that in a single-step run across three specialized benchmarks and four combinations of models and settings, GLM-5 and Gemini 3.5 Flash improved over their existing settings. The largest gains are 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points in LAB partial pass rate.
- Running it assumes a Linux and Python 3.12 environment. Benchmarks that use models incur API costs, and a local demo that runs without model credentials is also provided.
14 Milk frother: a home experience making coffee cream topping
- A milk frother is a small handheld tool with a handle. You dip it into a cup of milk, hold the power button, and spin it to make foam. Verge reporter Elizabeth Lopatto named this product the best purchase of the year and said the price was about 14.
- According to the reporter, with new batteries it takes about 30 seconds to 1 minutes, and when the batteries weaken the time gets longer, so replacement is needed. This is the reporter's own experience.
- According to the article, the frother is also used for beating eggs, mixing protein powder or creatine, and making hot chocolate. The reporter used it most for making cream topping to put on coffee.
- The cream topping is said to originate from the Einspänner, a drink from Vienna, Austria. Later it became popular in Korea in several variations, and in the United States it spread on the trend made familiar by Starbucks cold foam drinks, the reporter explained.
µRepRap, an open project to build a micrometer-scale 3D printer from common materials
- µRepRap aims for 1 micrometer precision (1 of 1000 parts of 1mm) with an 3D printer method, and it is an open project whose goal is to build the printer from easily obtainable materials.
- The developer claims that the recommended build, PIKA V0.02, shows repeatable positioning precision of about 1 µm and accuracy of 10 µm or better in a cube region of about 8mm. This is the developer's own measurement.
- According to the developer's report, PIKA dropped UV resin droplets of 4 µm in size using a fine probe that it etched itself, and it cured resin lines of 9 µm using an optical fiber probe.
- The source notes that thin UV resin does not cure well in air, and that moving other substances requires a solvent that does not evaporate during the work.