3297626
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).
🌍 World Labs makes its 3D world-generation model “Marble” available to everyone
World Labs, the spatial-AI startup co-founded by Fei-Fei Li, has opened access to Marble, a generative model that can create full three-dimensional worlds from text, images, video, or existing 3D layouts.
🔸 Marble can generate persistent, editable 3D environments and export them as meshes, Gaussian splats, or video for use in games, film, and simulation.
🔸 The model accepts multimodal inputs from a single photo to multi-view video and includes an editor called “Chisel” for sketching rough 3D structures before applying style and detail.
🔸 World Labs positions Marble as a step toward richer “spatial intelligence,” enabling designers, creators, and robotics teams to build virtual worlds far faster than traditional workflows allow.
Generative AI is moving beyond text and images into full spatial understanding, turning world-building into a prompt-driven process.
🎙 Fireflies founders faked their AI for a year, now it’s a $1B unicorn
In 2016, the co-founders of Fireflies, now a unicorn meeting intelligence startup, promised clients an AI that could automatically join and transcribe calls. There was just one catch, the AI didn’t exist yet.
🔸 For the first year, the two founders personally joined every client call, manually tracking who spoke and what was said.
🔸 The service cost $100 per month, covering about 100 meetings, while customers believed an autonomous system was behind it.
🔸 A year later, they built the real automation, eventually raising $19M and reaching a $1B valuation.
Today it’s a legendary MVP story, but it raises the timeless startup question: when does “fake it till you make it” cross from hustle to scandal?
🎵 An AI-generated song tops Billboard for the first time ever
For the first time in history, a fully AI-created track has reached #1 on the U.S. Billboard chart. The song, “Walk My Walk” by the fictional artist Breaking Rust, has already surpassed 1 million streams on Spotify.
🔸 The track was entirely generated by AI, with no human performer behind it.
🔸 It spread virally across streaming platforms and social media before users realized it wasn’t real.
🔸 Spotify, now flooded with AI-made tracks, has no plans to curb the influx, signaling a new era for digital music.
If algorithms can now make hits, what happens to the concept of a “musician” itself?
Examples of Nano Banana 2 in action. There are rumors that it is scheduled to be released early this week.
📊 Powered by Crypto Insider
📰 Bad news: AI now writes more articles than humans
A new analysis shows that, for the first time, AI-generated articles have overtaken human-written ones on the web. The gap is still small but growing fast.
🔸 Researchers sampled 65,000 web articles (2020–2025) from the CommonCrawl dataset and analyzed them with Surfer AI Detector.
🔸 If over half of an article’s text was flagged as AI-written, it counted as machine-made.
🔸 Cross-checks by publication date helped avoid false positives older pre-ChatGPT articles, for example, were assumed to be human.
🔸 The tipping point came in November 2024, when AI-made content first surpassed human output; after a brief dip, AI now dominates again.
Not all “AI-written” pieces are fully automated, many show signs of human editing or partial rewriting and such content still appears less often in top search results.
Still, the web’s content balance has flipped. The internet’s new authors may not be people, just machines that learned to sound like them.
🧠 PewDiePie builds a $20,000 supercomputer to run local AI models
One of the world’s biggest streamers, PewDiePie, has taken a sharp turn from YouTube entertainment to AI independence. After more than a decade of gaming videos and vlogs, he’s now built a personal supercomputer worth around $20,000 to run local large language models (LLMs) entirely offline.
🔸 The setup includes 10 GPUs: 8 custom RTX 4090s (48 GB each) and 2 RTX 4000 Ada cards, powerful enough to handle models like Llama 70B, gpt-oss-120B, and Qwen 245B or up to 64 smaller models simultaneously.
🔸 He even created his own local AI interface and a “council of models” that debate his questions in real time.
🔸 PewDiePie says his goal is simple: keep full control of his data and avoid AI corporations like OpenAI, criticizing them for storing user chats including deleted ones.
From Let’s Plays to local LLMs, PewDiePie’s pivot marks a new era where creators aren’t just using AI, they’re owning it.
🧠 Anthropic reinvents AI agents with code-driven MCP execution
Anthropic has rolled out a major update to how AI agents operate introducing code execution through MCP (Model Context Protocol), a shift that could redefine agent efficiency.
🔸 Traditionally, agents relied on multiple tool calls through MCP, consuming large amounts of compute credits and quickly filling the context window.
🔸 With the new approach, agents now write and execute code that calls the necessary tools directly through MCP, cutting token usage by 98.7%.
🔸 Anthropic’s guide includes code samples, step-by-step implementation details, and integration methods for existing agent frameworks.
🔸 The result: faster, cheaper, and more autonomous AI agents capable of handling multi-step tasks with minimal overhead.
This update isn’t just optimization, it’s a new programming paradigm for building scalable, self-sufficient AI systems.
🧠 The first team of AI scientists is now conducting research on its own
A new breakthrough has arrived: the first publicly available team of AI scientists can now carry out full-scale scientific research from forming a hypothesis to publishing a finished article without any human involvement.
🔸 You can enter any problem or dataset, and the AI autonomously defines the research goal and begins investigation.
🔸 These AI scientists generate hypotheses, verify novelty, plan experiments, run code, and analyze results.
🔸 The system produces a complete, publication-ready scientific paper as its final output.
🔸 Papers have already been released across 13 disciplines, including quantum physics, chemistry, and neuroscience.
🔸 Independent experts reviewed the results and confirmed that the AI’s work matches the quality of real peer-reviewed research.
If this scales, we may be witnessing the start of an era where AI doesn’t just assist science it actively advances it.
📹 REAL Video Enhancer launches, upgrade any video by tens of times
A new tool called REAL Video Enhancer has just dropped, promising professional-grade video restoration and upscaling all processed locally on your device.
🔸 Enhances and sharpens footage by tens of times, improving clarity and resolution.
🔸 Compresses files without any visible quality loss, ideal for large projects.
🔸 Removes noise, distortion, and artifacts from low-quality or old recordings.
🔸 Maintains full detail and texture integrity during enhancement.
🔸 Works entirely offline, keeping all content private.
🔸 And it’s completely free to use.
For creators and editors, REAL Video Enhancer could turn grainy clips into studio-quality footage without needing cloud access or paid software.
🩷 Tinder will soon use AI to read your camera roll and match you better
Tinder is testing a new AI feature called “Chemistry” that analyzes your camera roll with your permission to understand your lifestyle and interests more deeply than profile bios or prompts ever could.
🔸 The system uses computer vision to spot patterns in your photos (like travel, pets, or food) and combines them with short quiz answers to build a personalized match profile.
🔸 The pilot is currently rolling out in New Zealand and Australia, with plans to make it a major feature in Tinder’s 2026 redesign.
🔸 Match Group expects a short-term $14 million revenue dip as it experiments with the rollout, betting that deeper personalization will drive long-term engagement.
If it works, Tinder’s next matchmaker might not be an algorithm guessing from swipes but an AI quietly studying your photo gallery.
🎆 China breaks world record with nearly 16,000-drone sky show
In Liuyang, known as the “fireworks capital of the world,” organisers launched 15,947 drones in one synchronized display, setting two Guinness World Records for “largest number of drones controlled by a single computer” and “largest number of drones launching fireworks”.
🔸 Each drone flew a precisely controlled path using RTK positioning and mesh-network communications, forming shapes like towers, blossoms and a glowing “Sky Tree”.
🔸 The event replaced traditional chemical fireworks with digital choreography, signalling a shift from classic pyrotechnics to high-tech light shows.
🔸 While visually spectacular, the show also leveraged technologies used in autonomous systems mmwith implications beyond entertainment, including mapping or coordinated drone deployments.
🔸 The organisers acknowledged the risks: drone swarms require flawless software and precision, and past drone shows in Liuyang have malfunctioned, causing drones to fall toward crowds.
This display isn’t just a stunning light show, it’s a demonstration of how large-scale drone orchestration is maturing.
🎥 YouTube turns likes into mini-animations to boost engagement
YouTube has updated its like button, it now shows custom animations depending on the video’s theme. A car video gets a spinning wheel, a football clip a bouncing ball, and a travel vlog a plane taking off.
🔸 The like itself hasn’t changed, you still can’t “multi-like” or react with emojis, only one click but the visual effect now varies across content types.
🔸 Dislikes remain static with no animation added.
🔸 The goal is to make liking more fun and “convert better,” nudging users to click more often.
More likes mean richer engagement data for algorithms, longer average watch time (since people are less likely to quit videos they’ve liked), and deeper platform attachment through frequent interaction.
A small UI tweak but a clever psychological move to keep viewers watching longer.
💸 Fewer checks, bigger bets: early rounds now dominated by single investors
Peter Walker, Head of Insights at Carta, shared data showing a sharp shift in early-stage funding, lead investors are now taking nearly 100% of Seed and Series A rounds. The rest, a symbolic 1%, goes to angels.
🔸 The logic is simple: lead investors secure better terms today and tighter control tomorrow, while founders get a cleaner cap table and faster closes.
🔸 For follow-on rounds, a fully led seed signals conviction “if Fund X took the whole thing, they must’ve seen something.”
🔸 Angels and syndicates are getting edged out, but founders benefit from speed, clarity, and fewer competing opinions.
🔸 The trade-off: less investor diversity means fewer networks and viewpoints early on.
So what’s better — many small believers or one big backer?
📊 Powered by Crypto Insider
📈 AI is enriching investors while compressing wages
Artificial intelligence is accelerating wealth for investors even as it diminishes the value of human labor, creating a historic concentration of capital in tech stocks.
🔸 The seven largest technology companies now have a combined market capitalization exceeding $20 trillion, an all-time high.
🔸 These companies’ shares account for roughly 35% of the S&P 500, marking a level of concentration unseen even during the dot-com bubble in 2000.
🔸 While AI boosts productivity and profit margins, it simultaneously reduces the economic leverage of human workers, widening inequality.
🔸 Investors are reaping unprecedented gains, but the broader workforce sees limited upside from the same technological boom.
The rise of AI highlights a fundamental shift: the greatest beneficiaries of technological progress are increasingly shareholders, not employees.
✅ Hugging Face drops a free 200-page textbook on building neural networks from scratch
Hugging Face just released a comprehensive open textbook that teaches how to create and train AI models from the ground up, completely free.
🔸 The 200-page guide covers the full training pipeline, from core math and data preparation to backpropagation and model optimization.
🔸 It’s written in accessible language and packed with diagrams, making deep learning fundamentals understandable even for beginners.
🔸 The release fits Hugging Face’s mission of open AI education and transparency, lowering the barrier for anyone to learn how modern neural networks actually work.
In an industry dominated by closed systems and paid courses, Hugging Face just turned AI education itself into open-source infrastructure.
⚙️ You can now steer ChatGPT mid-response
OpenAI just added a small but powerful feature, you can now update your prompt while ChatGPT is responding, instead of stopping and starting over.
🔸 Simply hit “Update” to tweak your query on the fly, add missing details, correct instructions, or shift direction (“Use Java instead of Python,” “Make it less formal,” etc.).
🔸 Works even in Deep Research mode, where responses take longer and restarting would waste time.
🔸 It feels minor, but it transforms how you interact with long-running or complex tasks, making conversations more iterative, flexible, and human-like.
A quiet update, but one that makes ChatGPT feel less like a black box and more like a real-time creative partner.
🚀 OpenAI rolls out GPT-5.1 with “Instant” and “Thinking” modes
OpenAI has released GPT-5.1, a major update introducing two new models designed for different user needs, one for fast, natural conversation and another for deep reasoning.
🔸 GPT-5.1 Instant brings back the warmth and humor users missed from GPT-4o, offering quick, engaging replies that feel more human.
🔸 GPT-5.1 Thinking is optimized for complex reasoning, it adapts its response time to question difficulty, thinking longer on hard problems and speeding through easy ones.
🔸 The update responds directly to user feedback that GPT-5 felt “cold” and overly mechanical, restoring balance between personality and precision.
🔸 Older models remain in the legacy section for three months, while the new lineup rolls out first to paid users.
With GPT-5.1, OpenAI is clearly splitting its path, one branch for warmth and speed, the other for depth and intelligence.
🤖 Russia’s first humanoid robot AIdol collapses during live debut
In Moscow, developers unveiled AIdol, touted as Russia’s first anthropomorphic robot capable of walking, handling objects, and interacting with humans. The event, however, didn’t go as planned.
🔸 The robot entered the stage to the Rocky theme, accompanied by two handlers.
🔸 Moments later, AIdol fell mid-presentation, prompting staff to shield it from the audience while addressing the issue.
🔸 Developers said the incident stemmed from a test configuration error, noting the project remains in its trial stage.
Despite the stumble, the demo highlights how far and how fragile, the push toward humanoid robotics still is.
💻 Learn any IT profession with an AI course generator
Found a cool site called Getinclub that automatically builds personalized learning programs for any tech career.
🔸 Just enter a profession say, Data Analyst or Frontend Developer and the tool creates a full learning roadmap.
🔸 For every step, it gathers free courses, tutorials, articles, and video lessons from across the web.
🔸 Includes intermediate quizzes and tests to reinforce progress.
🔸 No ads, no subscriptions, just structured, open-access learning.
A smart way to turn “I want to learn IT” into an actionable, guided journey.
⚡️ Inside Tesla’s first Optimus production line
Tesla has quietly set up its first pilot production line for the humanoid robot Optimus. It’s a small-scale test run for what the company hopes will evolve into a full manufacturing system by 2026, when a larger line is planned.
🔸 The current line assembles and tests robots directly at Tesla’s factory and R&D labs, where Optimus units are already performing basic tasks.
🔸 The company’s target is ambitious, a $20,000 cost price per robot, putting it in the same range as a compact car.
🔸 This phase remains experimental, aimed at refining motion control, battery integration, and safety routines before ramping up.
🔸 Musk’s bet: mass-producing humanoids could one day be as routine as building cars — though testing them next to car lines raises the question, what could possibly go wrong?
Tesla’s robot factory may be small, but symbolically, it marks the moment the company moved from demo mode to industrial ambition.
🚀 Neural networks that can replace your entire business team
These new AI services can handle nearly everything, from writing and coding to design, scheduling, and client communication, helping founders save hundreds of thousands in operating costs.
🔸 Text & code: advanced models now draft full marketing campaigns, blogs, and even production-ready codebases.
🔸 Design & media: AI tools generate realistic product photos, animations, and promo videos in minutes no studio required.
🔸 Ops & support: smart agents manage calendars, emails, and even chat with customers 24/7 with human-level tone and context.
For entrepreneurs, these neural networks don’t just boost productivity, they’re a direct path to scaling lean, fast, and profitably.
🎓 Free AI & Python courses with certificates from Google, IBM, and Microsoft
Some of the biggest tech companies are offering free, certified courses to help you build real AI and coding skills no paywall, no subscription.
⚙️ Google — Machine Learning Crash Course
• 40+ hours of hands-on exercises, TensorFlow tutorials, and real-world data projects.
• Includes a verified certificate from Google.
🧠 IBM — AI Engineering Professional Certificate
• Covers NLP, ML, and Deep Learning with practical labs and model-building projects.
• Recognized pathway for IBM’s AI roles.
💻 Microsoft — Python for Beginners
• A full video series made by Microsoft engineers.
• Teaches Python step-by-step perfect for coding newcomers.
🤖 DeepLearning.AI — Generative AI with LLMs
• Learn how to build prompts, use GPT models, and apply LLMs in real scenarios.
• Co-created with top AI researchers.
📊 Kaggle Learn — Python & Machine Learning Tracks
• Short, interactive modules on Python, Pandas, ML, and AI foundations.
• Practice directly in your browser no setup required.
These courses don’t just teach,they certify your skills from the companies shaping the AI revolution.
💎 45 startups poised for breakout exits in the next decade
CB Insights has identified 45 private startups most likely to achieve major exits IPOs or acquisitions within 5–10 years, driven by shifts in AI infrastructure, strategic partnerships, and vertical specialization.
🔸 Infrastructure over applications: As AI agents replace traditional SaaS, infrastructure companies are becoming the real profit centers. Many founders, ex-Google, Stripe, and Robinhood engineers are building the core systems that others depend on.
🔸 Partnerships drive scale: The strongest startups aren’t going solo. They’re integrating with corporations, creating ecosystems where partnerships are as critical as technology itself.
🔸 Vertical AI wins: With foundation models turning into commodities, value is moving toward startups with domain-specific data and expertise in sectors like health, law, and energy.
The biggest exits of the 2030s may come not from AI chat apps, but from the deep infrastructure and vertical players quietly powering the entire AI economy.
🧠 Anthropic launches “Skills”, modular extensions that let Claude learn on the fly
Anthropic has introduced Skills, a new way for users to extend Claude’s capabilities with custom instructions, scripts, data, and templates turning the model into a flexible, task-specific assistant without retraining it.
🔸 A Skill can contain rules (“if X then Y”), datasets, documents, or even company-specific resources like brand guidelines or tone templates.
🔸 Claude can automatically select the right Skill for a task or let users toggle between manual, semi-automatic, or fully automatic activation.
🔸 Anthropic also ships default Skill packs for common tasks like data analysis, summarization, and editing, available to all paid Claude users, including API and Claude Code.
🔸 Unlike OpenAI’s separate user-built GPTs, Skills work as modular layers on top of the main model, meaning multiple can run or switch dynamically in a single session.
By blending flexibility with simplicity, Anthropic’s Skills may redefine how users personalize AI not by building new bots, but by teaching one model to think in many modes.
🤖 Dobot’s affordable robot dog aims to make home robotics mainstream
Dobot has launched the Rover X1, a multi‑purpose robot dog priced around US $1,030, designed to bring advanced robotics into everyday households.
🔸 The Rover X1 combines precise sensors and AI-powered navigation with a consumer-friendly design, handling tasks from item transport and filming to home security and coding education.
🔸 Over 5,000 units were reserved on the first day, showing strong early demand for versatile, affordable home robots.
🔸 Unlike most single-function robots on the market, Rover X1’s multi-purpose capabilities aim to justify its presence in the home and expand the adoption of robotic assistants.
🔸 The launch aligns with China’s push for “home service robots” as a growth sector, signaling both policy support and market potential.
Versatile, accessible, and AI-enabled, the Rover X1 could shift robot dogs from novelty gadgets to practical household helpers if it proves reliable and easy to integrate into daily life.
🥤 Coca-Cola drops new AI-generated New Year ad made in a month, not a year
Coca-Cola has unveiled its latest New Year campaign, created entirely with the help of artificial intelligence marking the second year the brand has leaned on AI for its holiday storytelling.
🔸 The project involved a team of 100 people who generated and refined over 70,000 AI-created clips before assembling the final video.
🔸 Compared with traditional film production, the process was dramatically faster and cheaper.
🔸 What once took nearly a year of shooting and editing was completed in just one month thanks to generative tools.
For Coca-Cola, AI is becoming more than a creative experiment, it’s turning into a new production model for global campaigns.
🥗 One photo, full nutrition: AI app counts calories from any meal
A new app called CalPulse can now identify the full nutritional breakdown of food from a single photo, whether it’s a café dish, delivery meal, or restaurant plate.
🔸 Snap a photo, upload it to CalPulse, and the app instantly analyzes calories, proteins, fats, and carbs.
🔸 It also recognizes menu items from restaurants and delivery services, matching them to verified nutritional databases.
🔸 Designed for precision tracking, it helps users log macros without manually entering ingredients.
AI is making calorie counting effortless, just point, shoot, and eat smarter.
⚙️ Reduino lets you program microcontrollers entirely in Python
A new library, Reduino, makes Arduino-style projects accessible without learning C++, letting hobbyists and engineers code microcontrollers in native Python.
🔸 Simplified coding: Control LEDs, RGB strips, sensors, and other electronics using familiar Python syntax, loops, conditions, arrays, tuples, and even classes.
🔸 Automatic translation: Python code is converted behind the scenes into C++ system files, ensuring compatibility and stability.
🔸 Seamless deployment: The library can compile and upload code directly to devices, removing the traditional Arduino workflow friction.
🔸 No C++ required: Beginners can focus on logic and experimentation without learning Arduino’s programming language.
Reduino could reshape DIY electronics by making Python the universal language for microcontrollers, lowering the barrier for creators and innovators alike.
🚀 The first Nvidia H100 GPU just went to space
A small satellite, roughly the size of a mini-fridge, has carried the first Nvidia H100 into orbit, part of an early experiment by Starcloud, a startup aiming to build gigawatt-scale data centers in space.
🔸 The satellite includes solar panels, radiators, and a single H100 GPU, designed to test how high-performance chips operate beyond Earth’s atmosphere.
🔸 Starcloud’s long-term goal is to create massive orbital data centers with millions of GPUs running on direct solar power.
🔸 Operating in a sun-synchronous orbit means the data center is always in sunlight, removing the need for batteries or nighttime power plants, a huge advantage in energy efficiency and cost.
🔸 A proposed 5-gigawatt orbital facility would weigh around 20,000 tons and cover 16 square kilometers, nearly half the total mass humanity has ever launched into space.
The concept hinges on launch economics: Starcloud’s plan assumes orbital costs of about $30 per kilogram, achievable only with SpaceX’s Starship. Until then, the company is testing power, cooling, and GPU performance in orbit, a first step toward making space the next frontier for AI compute.
🔔 Elon Musk predicts the end of smartphones, AI displays will take over
Elon Musk believes smartphones are on their way out. Within the next 5–10 years, he says, they’ll be replaced by AI-powered displays that act as direct interfaces between local and cloud-based large language models.
🔸 Instead of apps or operating systems, these handheld displays will generate personalized interfaces and content in real time.
🔸 Communication, media, and productivity will all flow through AI, no icons, no downloads, no screens full of apps.
🔸 The idea aligns with Musk’s broader vision of human-AI integration, blurring the line between digital and biological interaction.
If this future materializes, it won’t just replace the smartphone, it could erase the entire concept of a “device,” making AI the true operating system of daily life.