3297625
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).
🤖 OpenAI Model Solves 80-Year-Old Math Problem
An OpenAI model solved a math problem posed by Paul Erdős in 1946. The task was to find how to place n points on a plane so that the maximum number of pairs share the same fixed distance. For decades, the best known solution was a square grid arrangement.
The AI found a way to arrange points that produces significantly more pairs at a unit distance than any grid. It discovered an infinite family of such arrangements, using algebraic number theory methods never applied to this geometry problem before.
This was not a specialized math program but a general-purpose model. It was not trained or guided specifically to solve this problem, bridging two distant math fields in a way humans had not seen.
📊@tech
🚀 SpaceX launches Starship Version 3
On May 22, SpaceX flew the 12th test of its Starship Version 3 from Starbase, Texas. This was the first launch of the fully redesigned third version and the first Starship flight since October.
At liftoff, one of 33 Raptor engines on the Super Heavy booster shut down, causing the booster to fail its return and crash into the Gulf of Mexico. The upper stage, Ship 39, lost one of six engines but still reached space.
Ship 39 deployed 22 satellites, including 20 Starlink mockups and two real satellites with cameras. It reentered, survived plasma passage, performed structural maneuvers, and made a controlled splashdown in the Indian Ocean.
📊@tech
🔊 Vollebak launches jacket with 180 speakers
Vollebak’s jacket has 180 speakers facing inward to stimulate the brain and aid meditation on the move.
It aims to relieve anxiety through sound therapy using specific sounds and frequencies.
📊@tech
📊 FigureAI’s Robot Nearly Beats Human in 10-Hour Sorting
FigureAI ran a 10-hour experiment where a robot competed against a human in sorting packages. The human took legally mandated breaks, while the robot worked nonstop and autonomously.
The task required quick reactions, fine motor skills, and some reasoning. The human sorted 12,924 packages, the robot 12,732, with an average speed difference of just 0.04 seconds.
The robot’s endurance nearly matched the human’s speed, but the human still won by a small margin. FigureAI’s creator said this was the last time a human would win.
📊@tech
Ⓜ️ Meta trains AI on employee actions amid layoffs
On April 30, Mark Zuckerberg told managers that training AI models on employee actions will sharply improve their capabilities. Meta uses data from computer sessions, code, mouse movements, clicks, and keystrokes to teach AI agents.
Meta believes its employees’ higher intelligence offers better training data than external annotators. This approach is part of the Model Capability Initiative, reported by Reuters on April 21.
At the same time, Meta announced layoffs of 8,000 employees and reassigned 7,000 more to AI roles. The company is betting on AI trained from internal work to gain an edge.
📊@tech
📱 Apple turns imperfect chips into a bestselling MacBook
The Wall Street Journal reports that Apple uses A18 Pro chips with one disabled graphics core in the new MacBook Neo.
These are chips that did not fully meet the specifications for the iPhone 16 Pro, where Apple reserves the highest-performing versions.
Demand for the lower-cost MacBook was so strong that Apple reportedly ran out of these partially disabled chips and began using fully functional chips with one core intentionally turned off.
What looks like manufacturing waste can become a product category
📊 @tech
🤖 Claude Code prints token usage receipts
Developer Chris Hutchinson connected a thermal printer to Claude Code to print real receipts after each session. The receipts show token usage by model, input and output counts, total cost, and a QR code.
Claude Code is an AI platform where token consumption can add up quickly. The receipts use the SessionEnd hook and support Epson printers, HTML, or ASCII art in terminals.
📊@tech
🔍 Google unveils Gemini Omni video model
Google introduced Gemini Omni, a model that creates videos from any input: images, sound, video, or text. It can even turn a simple hand drawing into a full scene.
Gemini Omni understands physics and combines this with Gemini's real-world knowledge and reasoning. It also allows video editing by voice commands, adjusting actions, angles, or lighting.
Each edit builds on the previous one, keeping characters consistent and preserving the scene's physics. The model maintains context for the entire video instead of editing frames independently.
📊@tech
🔍 Google Unveils Gemini 3.5 and New AI Tools
Google introduced Gemini 3.5 Flash, their fastest and smartest model yet, outperforming Gemini 3.1 Pro in coding and long tasks. The AI now monitors global events and sends alerts on urgent news or releases.
Gemini also gained a real-time design generator called Stitch, a market analysis AI for finding best prices, and a mini-app generator in search to create simple web apps. The Gemini app got a major redesign with article-style responses.
These updates are available to all users, aiming to streamline workflows and information access.
📊 @tech
⚡️ JUST IN: Andrej Karpathy, cofounder of OpenAI, joins Anthropic.
📊 @tech
🍏 Apple turns Siri into chat bot in iOS 27
In iOS 27, Siri becomes a standalone chat bot app with dialogue history, accessible via Siri or a swipe down from the center screen.
Chats auto-delete by default after 30 days, a year, or never, unlike competitors that require manual incognito modes. Apple also limits what Siri remembers and for how long.
Siri will function like ChatGPT or Claude, offering conversational AI with privacy controls built in.
Apple is also updating Genmoji, its AI emoji generator, to suggest emojis based on photos and common phrases, improving on last year’s clunky launch.
📊@tech
A chart shows AI use is still limited. Red marks users of agents, yellow those paying for chatbots, green those who tried free plans. The rest never used AI.
This suggests AI technology is in an early adoption phase. Even now, there is not enough compute power for inference.
📊 @tech
🖥 Microsoft launches Agentic AI Developer exam
Microsoft introduced the Agentic AI Developer certification on GitHub under code GH-600. The exam tests skills in orchestrating AI agents and costs $165.
The certification includes a detailed study guide, recommended prep materials, and a mini practice test in a sandbox environment. Passing grants official recognition as a skilled AI orchestrator.
📊@tech
📲 Engineers Use USB Plugs to Keep Laptops Awake
Engineers often carry laptops slightly open to avoid interrupting agent work. This habit is so common that developers created USB plugs that simulate an external monitor connection.
These USB plugs let laptops stay on while closed by preventing sleep mode. This workaround keeps work uninterrupted without the laptop lid being open.
📊@tech
🤖 OpenAI Adds Codex to ChatGPT Mobile
OpenAI integrated Codex into the ChatGPT mobile app, enabling coding on smartphones. The app stays active while Codex runs.
Codex lets users manage agent sessions remotely, avoiding downtime away from a computer. This is free for all ChatGPT users.
📊@tech
▶️ Higgsfield’s AI Sci-Fi Debuts at Cannes
Startup Higgsfield created a 95-minute Sci-Fi film called Hell Grind using AI. It is the first AI-generated feature shown at Cannes.
The film’s budget was about $500,000, with nearly 80% spent on computing power for generation.
📊@tech
📊 Orbit Robotics unveils four-armed space robot
Canadian company Orbit Robotics introduced HELIOS, a humanoid robot with four arms designed for microgravity environments. The extra arms help it move, stabilize, and perform tasks inside space stations without needing legs.
HELIOS uses a cable-driven actuator system with motors near its shoulders to reduce moving mass. It aims to cut down routine maintenance time, which currently takes astronauts about 35% of their work hours, including up to 50 hours unloading cargo ships.
By handling repetitive tasks like maintenance and cargo management, HELIOS could free astronauts to focus more on science and research aboard the station.
📊@tech
🤖 OpenAI taught Codex to work on a locked Mac from your phone
OpenAI updated Codex so users can now send tasks from a smartphone and have the agent execute them on a Mac even when the computer is locked and the screen is off.
The feature works through the new Computer Use plugin. Codex temporarily unlocks the system in the background, performs the task, and locks the Mac again. If someone touches the keyboard or mouse, the session immediately stops.
Until now, AI agents generally required an active desktop session to function.
📊@tech
🔋 Google AI Studio can now generate Android apps from prompts
Google added native Android app generation to AI Studio. Users can describe an idea in plain language, and the system writes the app in Kotlin.
The generated app can be tested directly in a browser emulator and installed on a real Android device.
The gap between idea and APK keeps shrinking.
📊 @tech
🛍️ Alibaba launches Qwen3.7-Max AI
Alibaba released Qwen3.7-Max, a free AI model outperforming Kimi-K2.6 and Claude Opus 4.6 in Max mode on all benchmarks.
It excels at coding and agent tasks, runs autonomously for tens of hours, and supports 1000+ tools with a 1 million token context window.
📊@tech
📊 Figma launches built-in design AI agent
Figma introduced its own AI agent that works directly inside the main app window, removing the need for third-party AI tools.
The agent quickly learns the current design system and navigates between frames. It can edit layouts, find references, or create alternative design versions.
📊@tech
🔍 Google launches audio smart glasses this fall
Google will release audio smart glasses this fall under its Android XR platform, developed with Samsung and Qualcomm. The first model is audio-only, with a display version to follow.
The audio glasses deliver sound via private speakers and respond to "Hey Google" or touch. They identify objects, provide navigation, and handle calls and messages without a phone.
Designed with Gentle Monster and Warby Parker the frames offer style and comfort. Features include photo/video capture by voice, real-time translation with tone matching, and multitasking via Gemini Intelligence
They support both Android and iOS devices, broadening their compatibility.
📊@tech
🔍 Google launches generative search interface
Google is rolling out a free generative search interface that builds custom visual tools and simulations tailored to each query. It analyzes the question, designs a response layout, fetches data, and generates interactive code on the fly.
The search creates small apps for every question, offering more than just text answers. Soon, multiple AI agents will be integrated to handle different tasks directly within search, starting with information agents.
These agents scan the entire web, including blogs, news, social media, and real-time data on finance, shopping, and sports. This aims to deliver constantly updated results specific to user queries.
📊@tech
🧑💻 GitHub Hack Exposes 4,000 Repos
Hackers accessed about 4,000 internal GitHub repositories and offered the source code for $50,000.
They exploited a flaw in a VS Code extension to breach the platform.
📊@tech
🤖 OpenClaw spends $1.3M monthly on Codex tokens
The creator of OpenClaw spends $1.3 million per month on Codex tokens, using 603 billion tokens and making 7.6 million requests in 30 days.
This covers 100 agents running constantly to check every PR and issue.
OpenClaw agents find security holes, close old bugs, deduplicate issues, check performance, ban spammers, and even create temporary environments.
Some agents listen to team meetings and start PRs during feature discussions.
📊@tech
🤖 Cursor launches Composer 2.5 AI model
Cursor released Composer 2.5, their most powerful model yet.
It writes code at the level of Claude Opus 4.7 but is much more efficient.
Composer 2.5 improves handling of long tasks and was trained on its own errors, boosting answer accuracy.
📊@tech
🗣️ X users criticized an “AI Monet” that was actually painted by Monet
A blogger posted an image on X and said it was generated in the style of Claude Monet. He asked people to explain why it looked worse than the real thing.
Hundreds of users responded with detailed critiques. They called it soulless, complained about the composition, and pointed to supposed flaws in depth and color.
The catch: the image was an actual Monet painting from his Water Lilies series.
📊 @tech
This device could predict incoming phone calls (2000s)
📊 @tech
🤖 Anthropic agents port Bun to Rust in 10 days
A merge PR with 1 million lines of code for the Bun project was completed in just over a week by agents. Bun, owned by Anthropic since late last year, is a runtime for JavaScript and TypeScript originally written in Zig.
The creator Jarred Sumner started an experimental rewrite of Bun in Rust using Claude about 10 days ago. The new Rust code is already on track to merge into the main build, passing 99.8% of tests in canary.
📊@tech
🤖 Claude Mythos hacks macOS with new exploit
Claude Mythos linked two previously unknown bugs to fully compromise macOS memory and gain complete system access.
The exploit targets one of the toughest operating systems for hackers, breaking into protected parts of macOS.
Researchers were so surprised they took the vulnerability directly to Apple's office, where it is now under review without a patch yet.
📊@tech