Tech news picked on 2026-10-18
- Previous ChatGPT answers were mainly text, code, and static images. Now the model shows charts, input forms, buttons, and calculators directly in the chat window when they fit the question, and users do not have to choose a feature mode.
- According to the OpenAI announcement, GPT-6 Sol has been applied to the Plus, Pro, Business, and Enterprise plans starting October 7, 2026, and GPT-6 Luna has been applied to the Free and Go plans starting October 8. The announcement also states that weekly users exceed 1.2 billion.
- The announcement gave examples such as an explanation of bicycle parts, a mortgage calculator where the repayment amount changes with a slider, a tax return preparation assistant, and a tool that draws charts from natural language requests. The source does not include any 3-party verification of these examples.
- According to the original report, Meta removed the 'train' wording from the list of website collection purposes in its new business account terms and narrowed use to advertising and business products.
- Legal experts said the phrase 'developing business products' is still broad enough to include training. The revised terms also contain no language saying Meta will respect robots.txt (a file that tells websites whether search or collection bots may access them).
- A Meta spokesperson said robots.txt settings remain effective as before and that the permissions for news sites have not changed. This is Meta's claim, and the original report did not independently verify it.
- The new terms take effect on October 30, and according to the original report, business accounts have no option to reject the new terms before they take effect.
Google Workspace Gmail expands voice questions about email content to paid users
- Gmail Live lets users ask questions by voice in the mobile app and hears answers spoken aloud, based on the contents of their inbox. The original article gives examples such as order status, the next training deadline, the latest email from a manager, and checking a boarding gate.
- Each answer shows a source card displaying the emails it was based on, and a voice selection feature has been added. According to the original, it opens gradually in the Rapid Release and Scheduled Release domains starting on October 15, 2026, and over up to 15 days.
- Gmail AI Inbox automatically extracts to-dos, deliverables, and deadlines from emails and shows a summary of key topics. According to the original, it opens gradually for Rapid Release domains starting on October 12, and for Scheduled Release domains starting on October 26.
- A next-step feature has also been released that lets users, after a Google Meet meeting, choose on the screen to write a follow-up email or export tasks to Google Tasks. The original says this feature is based on Gemini and Workspace Studio.
Turning ping command output into structured data with parser combinators
- A parser combinator is a way to turn text into structured data, such as objects. It reads a long string by chaining small parsing functions in order.
- Each parser cuts off the front part of the string, returns a value, and passes the remaining string to the next parser. The source shows an example where a number parser returns 10 and the remainder ' 20 30' from '10 20 30'.
- The author presents an example of directly parsing ping command output sent to a Cloudflare server using TypeScript. According to the author, doing the same task with regular expressions becomes harder to read as nesting increases.
- The author says regular expressions get messy for complex tasks and considers tree-sitter, a large syntax parsing library used for syntax highlighting in code editors, to be excessive. The author argues this approach fits the middle ground. This is the author's opinion, and the source contains no performance measurement results.
Robotaxi rules differ by city, adding limits on where and how vehicles can operate
- Federal autonomous driving legislation has been stalled in Congress for several years, so cities and states are writing their own road operation rules, the source reports.
- Minneapolis is considering an ordinance that would require a human monitor inside vehicles, which the source says would effectively block driverless operation. The San Diego City Council passed a resolution asking that Waymo be kept from entering the airport.
- Austin firefighters worry that Tesla Cybercabs, which have no steering wheel or pedals, may be hard to move when stopped, and they plan to ask autonomous operators to have emergency towing procedures, the source reports.
- Santa Monica has restricted charging facilities for 24 hours based on noise and traffic ordinances, the source reports. In California, a bill has been introduced that would give cities the power to limit the number of vehicles and their operating hours.
Teal, a statically typed dialect of Lua that records data types in code
- Teal is a language that extends Lua syntax, allowing variables and functions to be annotated with types such as arrays, maps, and records. The original says it also supports interfaces, union types, and generics.
- The original states that Teal aims to fill a role in the Lua world similar to what TypeScript does in the JavaScript world. At the same time, it says it will keep Lua's simplicity, portability, and embeddability.
- The tl compiler turns .tl files into regular .lua files. The compiler can be installed with the LuaRocks package manager, or obtained as prebuilt binaries for Linux and Windows mentioned in the original.
- For large projects made of multiple files, the Teal build tool Cyan is recommended. Editor integrations such as vscode-teal for VS Code and teal-language-server for NeoVim and others are also introduced.
- The project was started by Hisham Muhammad and is developed on GitHub. Its license is MIT, the same as Lua. The original body contains no performance figures or real-world use cases.
DeepSeek API adds web search and page fetching
- The search API finds web pages from query terms and returns titles, URLs, and summaries as structured results. According to the source, it uses DeepSeek's own search engine.
- The fetch API takes one URL and converts the page body into Markdown text. The source notes that results may vary depending on the availability of the target site and its access policies.
- According to the source's price list, search and fetch each cost 0.75 dollars per 1,000 successful requests. Requests that return empty results, fail, or time out are not charged.
- According to the source, server-side tool calls made when using the API with DeepSeek Harness, Claude Code, or Codex are currently free, and per-request billing starts at midnight Beijing time on November 1, 2026. The account-level per-second request limits are 5 requests each for search and fetch.
Porsche Macan Turbo EV, an electric SUV that combines everyday driving with sports car performance
- A reporter tested the 2026 Porsche Macan Turbo EV, an electric SUV, for a week, using it for driving children to soccer practice and grocery shopping. The original article says its exterior is ordinary and did not attract attention from people nearby.
- Acceleration was described as fast and cornering as good. The original article explains that with the 4-wheel steering option, the air suspension lowers the body in Sport Plus mode.
- The article states that the highway driving range is about 285 miles. The fact that charging from 19% to 80% at a 350-kilowatt charger took 20 minutes is a result of conditions the reporter measured directly.
- The tested vehicle's price, including options and shipping, was 149,610 dollars, and the base starting price is 112,700 dollars. The reporter felt that some options, such as the passenger display and ambient lighting decoration, were unnecessary.
- When the reporter asked the navigation to find a nearby fast charger while the battery was 22%, it guided them to use two stations in turn, which the reporter pointed out was confusing. The reporter then went directly to a 350-kilowatt charger.
Sphinx, a Git tool that checks understanding with questions before AI-written code is pushed
- Sphinx is introduced as a tool that steps in just before a developer pushes code to a remote repository and asks what the changes made by an AI coding tool actually do. A Git hook refers to such an automatic execution point.
- The original example is a situation where an event ID is saved before a charge() call while processing a payment webhook. It raises the question that a customer may not be charged when a retry follows a failed payment. This is an illustrative example and not an actual usage record.
- The original states that the questions are written by a separate reviewer that has never seen the user's conversation history, reading the diff, meaning the difference in changed code. If the answer is wrong, it asks again with a different scenario. If the answer is right, the push continues.
- According to the original, it can be set to stop only on protected branches, the check can be skipped with the sphinx yolo command, and skipped runs are recorded. The answers are not passed to the coding agent, and .env files and keys are not opened.
- The original says it works with Codex, Claude Code, Cursor, OpenCode, Pi, and plain Git. Installation requires Node 22.19 or later and macOS, Linux, or WSL. The claim that it was built to work with minimal tokens is the makers' design claim.
Boss Noise-Canceling Wired Earbuds Review: Plug Into USB-C and Use Right Away
- These earbuds work as soon as you plug them into a USB-C port. No Bluetooth pairing, charging, or app setup is needed, and they connect to recent phones and computers without a dedicated dongle.
- The noise canceling gets its power from the USB-C port instead of a battery. The price is 99, and according to the reviewer's experience, they reduce low-frequency noise such as air conditioners or low trucks well, but block human voices weakly.
- Reviewer John Higgins wrote that the wired connection gives low latency and supports 24-bit/48kHz high-resolution audio. The sound quality assessment is the author's subjective listening experience.
- Drawbacks the review pointed out: the cable catches easily on clothing or objects, wind noise mixes in easily, and call quality is poor. The Verge's own score is 8 points.
Open-source 3D design program SolidCraft: how the model recalculates when a dimension changes
- SolidCraft is an open-source 3D CAD program written in pure Rust. CAD refers to a tool for designing the three-dimensional shape of product parts on a computer.
- In the original demo, a Raspberry Pi 4 case is created from an empty design file, and when the height parameter is changed from 24mm to 32mm, the whole model is recalculated. This is a demonstration video provided by the development team.
- The original lists installers for Windows, macOS, and Linux, as well as a web version that runs without installation. It states that the current installers are not digitally signed, so Windows may show a SmartScreen warning.
- The original states that 75 of 76 public reference parts match analytical or FreeCAD calculated values for volume, area, center of mass, and bounding box. This result is announced by the development team, and the original describes the project as early stage but usable for simple parts.
Lean proof tool adds a feature to recheck results with multiple checkers
- Lean is a proof tool that uses a machine to check the correctness of mathematical proofs and programs. According to this article, the lake check command has been added to Lean v4.35.0, so you can export your work and recheck it with the official checker leanchecker. The whole process runs inside a sandbox by default.
- Using the lake check --paranoid option, the same work is rechecked with four additional checkers: nanoda, lean4lean, con-leche, and con-ron. A memory-hardened version of the official checker also runs, so memory errors show up as a check halt instead of a silent wrong answer.
- The article says con-leche is the first kernel written in Lean and proven consistent. The proof relies on a set-theory model, and the approach is described as reducing implementation-mistake type bugs. On the full Mathlib export, the author reports a measurement of 0.69 times the instruction count of the official kernel, using 8 threads in about 100 seconds.
- con-ron is a kernel that the AI ported from con-leche to Rust, then proved to match the original checking algorithm using the Aeneas tool. Aeneas is a tool that verifies Rust code with Lean. The article explains that using both kernels together gives stronger protection against errors in the Lean compiler or runtime.
Actor Mark Wahlberg takes on the high-end longevity market with a low-cost membership club
- Actor Mark Wahlberg told a TechCrunch interview that he is opening a membership club of 32,000 square feet called Municipal Gym near Las Vegas in December. The source describes it as a place that combines exercise, blood tests, and recovery facilities in one location.
- The source says the facility will include blood tests, epigenetic testing (a test that looks at how genes switch on and off), recovery equipment, compression therapy, cold plunges, and a cafe.
- Wahlberg, citing 99 as an example, said the club offers 3 blood tests and 1 hours of doctor consultations. These are his statements, and the source does not give the period or membership terms.
- The source lists as competing products Fountain Life at 21,500 per year (as reported), The Estate membership at 35,000 (in the planning stage), and an additional Equinox offering at 36,000 (due to launch).
- Pele is a lightweight open-source library that converts diagrams such as text-written flowcharts into SVG image files. SVG is an image format that stays sharp when enlarged.
- Colors and fonts are set with CSS variables. According to the source, changing colors takes effect immediately without redrawing the diagram.
- The source states that Pele has 0 dependencies and a bundle size of 230kB when gzip-compressed. In the same table, Mermaid.js is presented as 1.6MB with 23 dependencies. These are comparison figures published by the Pele side.
- The source presents rendering time as Pele 0.05 to 0.3ms and Mermaid.js 6 to 82ms, and claims it is up to 500 times faster. The measurement was made by the Pele development team, and the measurement environment is not given in the source. Pele was developed for the note-taking app Obsidian, and it reportedly supports all diagram types built into Mermaid 12.1.0.
Replit switches the default AI model for its agent from Anthropic to OpenAI's ChatGPT
- Replit is a cloud development tool that lets users build and deploy apps in the browser. CEO Amjad Masad said on stage at the Disrupt event that Replit had switched the default model for Replit Agent to OpenAI's ChatGPT.
- Masad cited cost, partnerships, and performance as reasons. The performance claim is based on benchmarks that Replit published itself, and the original article does not present independent verification or the measurement conditions.
- The original article reports that users can still choose other models such as Claude, but most customers tend to keep the default. It also adds that this change could significantly reduce Replit's spending on Anthropic.
- According to the original article, Replit was until recently one of Anthropic's top 5 customers and once ran its platform on Claude models. However, one internal source said the partnership is still ongoing, and a blog post from September also says Claude models are used in Replit Agent's production environment.
Anthropic says in an internal report that it cut junior hiring and increased meeting time
- In a report Anthropic published in October 2026, it said the share of junior-level employees within its technology departments fell from 21% in early 2025 to 14%. The share of junior technical hiring dropped from 27% to 9%.
- In the same report, the share of technology teams in total headcount fell from 61% to 56%. The company cited the growth of its business functions, along with the fact that non-developers can now build software directly with coding tools.
- Based on calendar data, the company said meeting time per 1 employee rose 20% compared with 1 years earlier, and rose 16% for technical staff. The company explained that in-person collaboration is needed to build trust and accountability.
- In a survey of employees the company conducted in April 2026, 27% said their productivity had increased thanks to Claude, and 20% said their work had shifted from doing tasks themselves to managing Claude. The company claimed it found no signs of quality decline even though code output had grown.
- An electric large SUV that can carry a family, yet delivers acceleration and cornering close to a sports car. The source describes it as the large SUV for which Mercedes-AMG developed the entire design in-house for the first time.
- There are three trims: GT 53, GT 55, and the top GT 63. The source says the top model uses the same 3 motors as the previously revealed GT 4-door coupe to produce 1,153 hp. The 1,150 hp in the title is a shorthand notation.
- With a 800-volt structure, the 128kWh battery trim supports up to 710kW, and the GT 53, which uses the 106kWh battery, supports 600kW charging. The source says it can travel up to 450km under the WLTP standard with a charge from 10 to 80 percent in 12 minutes, or 10 minutes.
- The estimated maximum range of the 128kWh trim is 435 miles. It also includes a suspension with active hydraulic elements and a sound feature that plays back a real AMG engine sound. US launch is scheduled for the following summer, and the price has not yet been announced.
Meta removes contract clause that raised concerns about web scraping and AI training
- Meta revised the terms of service for business accounts and added language saying that content from domains, web pages, and URLs connected to business accounts may be used to provide, improve, and research Meta products, including AI training.
- Legal experts interviewed by Nieman Lab said the language could be read as permission for Meta to scrape news publisher websites with business accounts and use them to train AI models.
- A Meta spokesperson said the terms change did not alter Meta's web crawling practices. According to The Verge, the language was later removed.
- In a statement, Meta said the unified crawling terms apply only to advertising and business products, and that it clarified this in the terms. This is Meta's own explanation and has not been independently verified.
DeepSeek releases the V4.1 Flash minor update model with improved reasoning and coding performance
- DeepSeek-V4.1-Flash-1017 is a minor update to the language model that reads text, answers questions, and writes code. The source says its performance improved in reasoning, coding, agent tasks that handle multiple steps on their own, and visual benchmarks related to images.
- The development team claims it can cut task completion time and output length, so the same work costs less. This is the team's claim, and the source contains no measurements from independent organizations.
- In the benchmark table released by the team, Terminal-Bench 4.0 scored 42.3 points and DeepSWE v1.1 scored 77.0 points, higher than the 31.2 and 74.2 points of the previous V4.1 Flash. The source says the two tests were measured in the Claude Code and Mini-SWE-V2 evaluation harnesses, respectively.
- This release does not include a Jinja chat template (a format that organizes conversation structure). Instead it includes encoding.py, a Python reference implementation. For production use, the deepseek-recipe library made in Rust is also provided. The recommended settings are temperature 1.0, top_p 1.0, and a context window of 1M tokens.
Meta deletes terms language that seemed to allow AI training on news site content
- According to the report, Meta removed language from its commercial terms page stating that it could use content from websites and URLs for product development and AI training. The language was read as potentially applying to many sites, including news sites.
- Nieman Lab and Cornell University professor James Grimmelmann and others said the language could be read as granting a right to use content for free beyond robots.txt blocking instructions. robots.txt is a configuration file that websites use to block or allow access by search robots.
- Meta VP Andy Stone and spokesperson Daniel Roberts said the way its web crawlers are used has not changed, and that the crawling terms apply only to advertising and business products. This is Meta's claim.
- Roberts said content owners can indicate their preferences through robots.txt, and that this update does not give any new rights beyond those preferences. The report does not include independent verification of how this explanation applies in practice.
- The new Mug 3 lets you change the temperature with the button on the bottom of the mug, without opening the app. Each press raises the temperature by 5 degrees. The mug cycles through five preset stages, starting at 125 degrees Fahrenheit and going up to a maximum of 145 degrees Fahrenheit.
- The original author said they used this button most at night in bed during testing. However, they also noted a drawback: because the button is on the bottom, you have to tilt the mug while holding a hot drink and press it firmly.
- In the app, you can connect to the mug over Bluetooth, choose presets for each drink, or set a temperature yourself. Ember's recommended settings are 136 degrees Fahrenheit for black coffee, 132 degrees Fahrenheit for green tea, and 142 degrees Fahrenheit for lemon water.
- The original article reports that the mug can keep its temperature for up to 3 hours on a single charge. The sliding lid is designed to reduce spills, and it is dishwasher safe on the top rack but cannot be used in a microwave. Prices are 10 oz for 129 and 14 oz for 149.