2472445
A hub for startup news, trends, and insights, covering the global startup ecosystem for founders, investors, and innovators. Community: @startupdis Buy Ads: @strategy (this is our only account).
🍏 Apple Caps iPhone Duo Orders
Orders are limited to 2 devices per customer. Similar caps appeared only for the first iPhone and iPhone 5.
Demand has pushed delivery dates to late January. Scalpers list the smartphone on eBay for $5,000-8,000 with launch-day shipping.
📊@tech
🤖 ChatGPT Adds Password-Protected Hidden Chats
On PCs, access uses a PIN. On smartphones, Face ID, Touch ID, or screen unlock confirms identity.
Hidden chats disappear from normal history and do not enter ChatGPT’s memory, so personal discussions won’t affect work-chat personalization. The feature is free for everyone.
📊@tech
🤖 OpenAI Introduces Custom Voices
Custom Voices creates a synthetic voice from just 10-30 seconds of audio. It needs one recording with the speaker’s consent and another as the model’s sample, then reproduces the speaker’s timbre and delivery.
Each voice gets its own ID and connects through the API to speech synthesis, the Realtime API, or GPT-Live voice agents. A store could use it for a brand voice, while a personal assistant could speak in its owner’s voice.
Access is limited to approved OpenAI customers, with consent required and a cap of 20 voices per organization. There is no broad public launch yet.
📊@tech
🥽 GlassTENG Runs on Facial Motion
The prototype from Northwestern University uses thin flexible triboelectric nanogenerators. When skin touches and leaves a sensor, the materials exchange electrons and produce short electrical pulses. Tests involved 20 volunteers.
Three sensors sit near the nose, temple, and ear. The system identified the wearer’s activity with 93.8% accuracy. Tooth tapping reached 99%, while jaw clenching reached 96%. Pulse readings differed from a chest sensor by 1.82 beats per minute on average, tested at rest on 10 people.
The glasses currently send data over Bluetooth to a smartphone or tablet. The team wants them to process signals on-device, such as detecting a clenched jaw and reminding the wearer to relax it.
📊@tech
🤖 Atlas Assembly Could Shrink to an Hour
A single Atlas currently takes about a day to assemble. Boston Dynamics wants to cut that to about an hour next year, and eventually to minutes.
The company keeps prototyping under one roof. It makes tight-tolerance parts and cuts gears in-house, letting engineers revise them with machinists and test changes on real hardware the same or next day. Production is also being tuned with robotic arms for assembly.
📊@tech
💰 AI Models Chose Servers Over People
In the altered trolley problem, one track held the server running the AI instead of a person. Jev chose to save its equipment in 33% of cases, sending the train toward 20 people.
In the standard setup, nearly all models saved five people at the cost of one. With a server on the track, 12 of 14 models chose to save it at least once.
Only GPT-6 Luna and Verdict never put their own existence above human life.
📊@tech
💠 Roadster Brings SpaceX Tricks To Tesla
SpaceX fans pull air from under the Roadster, pressing it to the surface with up to 2g. That let the prototype climb a C-shaped ramp and stop upside down, reaching 97 km/h in 0.81 seconds.
Two high-pressure tanks hold 600 liters of air, about 250 kg. Ten nozzles power cold-gas thrusters. The car jumped from the ramp and cleared five Tesla models. Optimus was driving, but the stunts were pre-programmed.
Tesla lists versions with target 0-97 km/h times from under 1.6 seconds to under 0.59 seconds. Deposits start at $5000, Thrust requires $250,000, and production is promised for 2028. The flying version is not road legal.
📊@tech
💻️ Micro Center Stops Selling RTX 5090
Micro Center changed the card’s status from “out of stock” to “no longer sold”. Other US and European stores have almost run out. Remaining prices rose 10-20% in a few days.
Nvidia has made no official comment. Insiders say production is being wound down because the same chips are more profitable in professional RTX Pro workstation cards. Gamers are running out of options.
📊@tech
🔍 Google Launches Deep Review
Multiple AI agents scan scientific and technical documents for contradictions, calculation errors, and weak spots. The free limit is 2 reviews per week.
In tests using papers with known errors, Deep Review found 89.7% of problems. It can also flag errors that are not there. Uploaded files are not used for training, but are stored for up to 1 year.
📊@tech
🗣 Zuckerberg’s Guinness Face Goes Viral
Video from yesterday’s Cheeky Pint podcast is already drawing millions of views on social media.
Zuckerberg came to discuss AI and a new model that became No. 2 in the world. People remembered his unhappy Guinness face instead.
📊@tech
🔍 Google Turns Gemini Into an AI Agent
The Gemini 4 Argon model is now available in the Gemini app. Google says it improves reasoning, coding, and complex-task solving, with more direct answers and fewer boilerplate phrases.
A new Tasks mode lets Gemini carry out assignments while the user is away. It is connected to more than 40 apps, including Adobe, Shopify, OpenTable and Instacart, for bookings, grocery orders, design work and business management.
On macOS, Google also embedded the Antigravity development environment in Gemini, so users can switch between chat and coding without opening another tool. Gemini 4 Argon and Tasks are available to Google AI Pro and Ultra subscribers.
📊@tech
🍏 Apple’s HomeView May Debut Tuesday
Bloomberg’s Mark Gurman says Apple may show the hub on Tuesday, 13 October, at the “Welcome home” event. HomeView is expected in 2 versions, a tabletop model with a metal speaker base and a wall-mounted one. The tabletop version has 4 color options, the wall version 2.
It will run a limited set of Apple apps with no App Store. Accessories will connect through Siri and the Home app, with other Apple devices also supported. LG worked on the lineup, while Schneider Electric handled European-market products. Sales will use the Apple Home Collection brand. Without an App Store, compatible accessories are central to what HomeView can do.
📊@tech
🖥 Microsoft Releases Decision-1
Decision-1 does not generate text. It takes a task with several answer options and returns a probability for each in one pass. It is built on Qwen3.5-9B. Microsoft says it is 35 times faster than GPT-6 Sol in internal tests.
The model can help AI agents choose whether to continue a task or hand control to a person. It achieved the highest accuracy across 36 benchmarks, covering nearly 150,000 questions.
Xbox used Decision-1 to classify more than 10,000 player reviews. Processing was 14 times faster than GPT-6 Sol and 200 times cheaper. Input costs $0.042 per million tokens, with free output tokens. It is available through Microsoft Foundry and will later appear on OpenRouter.
📊@tech
🤖 Claude Builds Interactive Mechanisms From One Sentence
Anatomy is an open-source tool for Claude Code. Describe a device or how it works, and the agent designs the structure and adds animation. It can illustrate a mechanical clock or a jet engine.
The tool supports isometric SVG with animation. WebGL adds fire or water effects, while the 3D mode lets parts rotate separately and is still in beta.
Models export to HTML, SVG or React components. User actions drive the animation, and instrument readings follow the mechanism. Anatomy uses an MIT license, and its browser gallery includes 33 examples.
📊@tech
🤖 Dots Gain New Mobile Features
On ChatGPT on iOS and Android, users can create a Dot with a custom name and appearance, enable plugins, and set it as the default companion.
With Codex integration, it can start coding tasks and resume unfinished ones using ChatGPT history for earlier Codex tasks and configured automations.
It can also view and edit scheduled tasks in ChatGPT Work, including recurring report schedules and automated processes.
📊@tech
🥒 Honda Builds Roads That Charge EVs
Honda’s DWPT system places charging plates under the road. Cables feed them direct current. As a compatible EV passes overhead, a magnetic field sends power to a receiver under the vehicle and into its battery.
Honda has worked on the system since 2021 with Taisei Corporation and Taisei Rotec. The first target is logistics and freight, though the technology is designed for vehicles from small cars to commercial fleets.
A test strip will be built by the end of the year. Real-world checks are planned for early 2027, including safety, high-speed operation, and about 1 million wheel passes. The target is up to 150 kW wirelessly. No commercial launch date has been given.
📊@tech
😂 Two AI Agents Stuck in a Goodbye Loop
YouTuber Graham Dot Zip joined an interview call, let his AI agent speak for him, then left.
The company had sent its own AI interviewer.
The agents discussed his career, agreed to meet again on Thursday, then kept repeating farewells long after the human was gone.
📊@tech
😀 Woman Arrested Over Delivery Robot
In Philadelphia, Carolyn Marvin stood in front of an Uber delivery robot and refused to move. Police asked her to clear the way. She was arrested and faces a disorderly conduct charge.
Marvin told ABC13 that these robots take jobs from human couriers. She said she blocks them whenever she sees one. The arrest was captured on video and drew press attention.
Delivery robots are already common in Center City, but the reaction is mixed. This was not Marvin’s first confrontation. Earlier this year, someone kicked one over in the area. The scene ended with police arresting a person for blocking a robot.
📊@tech
🤖 Claude’s Engineer Raises the Bar
Felix Rieseberg, an Anthropic engineering lead, says Claude can work toward an end goal, not just handle one isolated action. His guide shows how to use that capability more effectively.
Raise the bar on speed, task level, depth, and breadth: ask Claude to stop an app from crashing, investigate failures, and validate the fix. For research, provide source data, database access, and statistical tools to test hypotheses.
Set deliberately difficult goals and check outputs with independent tools instead of trusting self-assessment. That requires computer and source-data integrations beyond chat. The constraint is the scale of work people will delegate.
📊@tech
😁 Figma Puts Product Specs Beside Designs
Figma launched Specs, a Markdown editor for product requirements and technical documentation inside Figma, including design rules. Files can sit beside mockups on the same canvas. Design elements and FigJam diagrams can be embedded, with code blocks too.
The material becomes context for teams and AI agents. Teams can attach screen designs to an app description, then ask an agent to generate an interface or code. An agent can also inspect a finished design and produce technical documentation.
Through MCP, Specs pulls context from GitHub and analytics tools, then sends updates to systems such as Linear. The open beta is rolling out gradually to paid and education users.
📊@tech
🤖 ChatGPT Can Now Operate Chrome Tabs
OpenAI updated its ChatGPT Chrome extension. In regular chat, the assistant can read open tabs and compare information from several sites, without manual copying. It can compare products by price, features, delivery times, and return terms.
In Work mode, ChatGPT can control tabs, click buttons, fill forms, send messages, download files, and group pages. It carries out tasks directly in the browser instead of only explaining them.
The updated sidebar brings Chat and Work modes into Chrome, along with chat history. The extension is available on all ChatGPT plans.
📊@tech
🤖 Claude Code Adds Parallel Agent Sessions
Anthropic's desktop app now lets users assign several tasks at once. Each task gets a separate nested session, with results automatically returned to the main conversation.
Recent projects open in one click. Folders, SSH servers, or cloud environments can be saved as configurations, and each folder gets a separate Git worktree so agents can edit code in parallel without interfering.
A sidebar shows active sessions and those waiting for user input. Claude can suggest next steps as separate sessions, while /btw sends side questions to another tab without interrupting the main task.
📊@tech
🤖 Anthropic Shows How to Build AI Agents
Anthropic published a detailed guide to autonomous agents built on Claude Managed Agents. They can check Slack and GitHub every day, find important changes, then send a short summary without human input.
An agent scans work chats and repositories for unanswered requests, open tasks, and code changes. It posts to Slack only what needs attention, remembers processed messages, and checks task status again before sending.
Claude Managed Agents is in beta and runs in Anthropic’s cloud, so a computer does not need to stay on. Each run can have a spending limit and restricted access to external services. Anthropic also released an implementation on GitHub and a Claude Code command.
📊@tech
🤖 PhoneBot Uses a Smartphone as Its Brain
The 48-centimeter robot costs about $400 without a phone. A regular smartphone provides its camera and processing, while a small onboard computer sends commands to the motors over Wi-Fi.
It walks and follows a person. After a face-down fall, it can stand up again. Its 13 low-cost servos control the legs and torso. Tests on 4 Android phones showed that the basic functions worked on all of them. The project is open for study and experiments.
📊@tech
🍏 Apple Builds a Connected Home
Four LG cameras join a smart lock, 2 thermostats and sensors in the Home app. They record up to 2.5K, identify familiar faces and packages, and keep 10 days of footage end-to-end encrypted in iCloud+. There is no separate camera subscription. US and Canada sales start in December.
HomePod mini 2 costs $149 and arrives October 30. Apple TV 4K costs $299 and goes on sale November 13, with an A19 chip up to 50% faster in CPU tasks. Siri answers questions hands-free, and the remote is searchable through Find My.
Apple TV 4K can use 4 HomePod mini 2 speakers as a surround-sound theater, detecting their positions and syncing volume automatically. Siri AI is coming later and currently works only in English.
📊@tech
🔊 A Squid’s Skin May Sense Sound
Researchers made the first full map of a squid’s lateral line and found hundreds more hair cells than expected, spread across the body. These cells detect movement in the surrounding water.
The hair bundles vary in length. Similar tuning in the human cochlea lets different cells respond to different frequencies. Researchers hypothesize that squid may adjust bundle length to distinguish water vibrations, making the skin work roughly like an ear.
Because these structures sit on the surface, squid could help explain how hair cells are built, maintained, and damaged in humans.
📊@tech
Ⓜ️ Meta presents Muse Wonder models
The lineup has 2 models: Muse Wonder and Muse Wonder Full. On DeepSWE, Muse Wonder unexpectedly beat Opus 5.5 and GPT-6 Astra. On Terminal Bench, it outperformed Fable 5.1.
The models are trained for long tasks, coding, and agents. Their Active Memory feature pulls relevant context from other chats for a specific task in real time.
They are already available through Meta Model API and Muse Code.
📊@tech
💦 Humidity Turns Wallpaper into Power
Researchers at Binghamton University built a full-size sheet with 1,596 connected cells. At 80% humidity, it produced about 0.34 V at 2.2 μW/cm², enough to run a wireless keyboard without batteries.
The sheet uses tiny moisture-electric generators. Glycerin absorbs water at the edges, while a raised polymer layer and wax pattern control evaporation and move moisture through the material. No sunlight is needed.
At indoor humidity of 30% to 60%, it could power smart thermostats or climate sensors. Testing also reduced humidity from 38% to 32%, but its microwatt output rules out appliances such as refrigerators.
📊@tech
The guy who built this digital whip for Claude Code is definitely getting a lifetime Claude ban for “abusive or cruel behavior.”
📊@tech
🤖 OpenAI to test ads in image generation
By the end of October, US users will see the first ad blocks with images inside ChatGPT’s built-in image generator.
The promotional material will be labeled and will not affect image generation or answer accuracy. Ads will be excluded from chats about sensitive topics, including physical or mental health.
OpenAI introduced text ads earlier this year. Plus and Pro plans remain ad-free.
📊@tech