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).
🔒 Hackers stole data about OpenAI API users
OpenAI announced that in early November, the analytics service Mixpanel, which it uses for product insights was compromised, leading to a leak of API user data.
🔸 Exposed information includes:
• Name linked to the API account
• Email address associated with the API account
• Approximate location based on browser data (city, region, country)
• Operating system and browser used to access the API
• Referring websites
• Organization or user identifiers tied to the API account
🔸 OpenAI said the incident affected Mixpanel data only; no API keys, passwords, or payment information were leaked.
This incident is a reminder that even data analytics tools can be weak points, users and organizations should monitor accounts and update security practices accordingly.
🎙️ OpenAI rolls out a redesigned Voice Mode, now built directly into the chat
OpenAI has updated Voice Mode so it no longer opens in a separate fullscreen window, it now lives right inside the chat interface.
🔸 You can speak while still seeing your full message history, attachments, and the model’s responses as they appear.
🔸 The flow is faster and more natural: voice input + chat context in one place.
🔸 Prefer the old layout? You can switch it back in Settings → Voice Mode → Separate mode.
🔸 The update is already rolling out, just update the app.
A small UX change, but one that makes voice interactions feel like a seamless extension of the chat, not a separate feature.
🖼 FLUX 2 launches, a cheaper rival to Nano Banana with 4K image generation
Black Forest Labs has released FLUX 2, a new image-generation model positioned as a lower-cost alternative to Nano Banana while still delivering high-end output.
🔸 Supports 4K-resolution image generation
🔸 Can complete sketches and refine partially drawn concepts
🔸 Accepts up to 10 reference images for style and content control
🔸 Available to try here: https://playground.bfl.ai/image/generate
🔸 Open-weights version is live on Hugging Face: https://huggingface.co/collections/black-forest-labs/flux2
A strong new entrant in the race for fast, affordable, high-fidelity image models.
💻 Claude Opus 4.5 launches and becomes the new programming champion
Anthropic has released Claude Opus 4.5, and it now leads the industry in coding tasks, the first model to hit 80% on SWE-Bench, surpassing all competitors.
🔸 API pricing: $5 / $25 per million tokens (input/output).
🔸 Developers can now choose how the model thinks: fast and cheap, or slow with maximum reasoning depth.
🔸 Claude Code is now available inside the desktop app, with long coding sessions no longer breaking, the model automatically summarizes older context to keep the thread going.
A clear signal Anthropic wants the coding-assistant crown going into 2025.
🔒 Andrey Karpata launches LLM Council: AI answers by committee
Andrey Karpata’s new project, LLM Council, takes the ChatGPT interface and adds a twist: your request is processed by multiple models at once, then refined collaboratively.
🔸 Requests go through OpenRouter and are answered by several LLMs simultaneously (currently GPT-5.1, Gemini 3 Pro, Sonnet 4.5, and Grok 4).
🔸 Each model sees the other’s anonymous answers, comments, and ranks them before sending feedback.
🔸 All commentary is fed to the “LLM chairman”, which synthesizes the final answer for the user.
🔸 Models often pick another LLM’s answer as superior, GPT-5 consistently praised, Claude frequently rated lowest.
🔸 The project hints at a new frontier in ensemble LLM design, where context framing and model interplay may outperform single-model outputs.
LLM Council suggests that AI reasoning may benefit as much from collaboration and critique as raw model power.
📱China’s new “vibe-coding” app LingGuang goes viral after crashing from demand
Ant Group’s new AI coding app LingGuang has exploded across China, so fast that one of its flagship tools temporarily crashed from traffic.
🔸 The app lets anyone build mini-apps in under 30 seconds just by typing a prompt, Ant calls it “flash programming.”
🔸 It also generates 3D models, animations, charts, and interactive tools, plus a real-time camera that can understand scenes.
🔸 LingGuang hit 1 million downloads in four days, then passed 2 million, becoming the #1 free app on China’s iOS App Store.
🔸 Demand was so high that its flash-program feature briefly went offline as users rushed to test the app.
🔸 The product is part of Ant Group’s push to expand AI consumer tools and compete with China’s rising wave of easy-build app platforms.
LingGuang’s viral growth shows how fast AI-native creation tools can jump from niche demos to mass-market hits in China’s consumer ecosystem.
✅ Nano Banana Pro FOR FREE, the most powerful image-generation model just landed on Higgsfield
🔸 Generates flawless infographics, blueprints, charts, maps, and technical diagrams 100% accuracy on structured visuals
🔸 Produces perfect screenshots, hyper-realistic photos, sketches, UI mockups, product renders, and more
🔸 Can solve problems directly on the image, equations, homework steps, calculations, annotations
🔸 Engineered for zero hallucinations, making it ideal for technical, scientific, and instructional graphics
A free model that nails both creative and structured visuals is a serious disruption to every paid imaging tool on the market.
📮 Maildrop launches one-click disposable email to block spam
A new service called Maildrop has launched, offering instant disposable inboxes that protect your real email from spam, leaks, and shady sign-ups.
🔸 With a single click, you generate a temporary mailbox and immediately start receiving verification codes, login links, or any other incoming emails.
🔸 No registration required, just open the site, copy the address, and use it anywhere you don’t want to expose your primary inbox.
🔸 For more serious use cases, you can also create a custom mailbox name, giving you a cleaner, more memorable burner address.
🔸 Ideal for testing services, avoiding marketing spam, and preventing your main email from being scraped or sold.
A simple, lightweight tool and exactly the kind of utility people will keep returning to as online clutter keeps getting worse.
✅ Perplexity launches Comet browser on Android
Perplexity’s Comet Browser is now live for Android users, bringing built-in AI that can answer questions about your open tabs, help with research, and even handle purchases. The team says an iOS version is on the way, set to roll out in the near future.
A notable step in Perplexity’s push toward fully AI-assisted browsing.
📊 Powered by Crypto Insider
⚙️ Learn Python, Docker, Git, DevOps & more, all in one interactive platform
A new all-in-one learning resource has popped up with fully interactive courses covering Python, Docker, Git, DevOps, and many other tech topics, perfect for beginners and upskillers.
🔸 Step-by-step explanations for every concept, function, and method.
🔸 Interactive lessons with lots of hands-on practice and real examples.
🔸 Covers everything from basic syntax to advanced workflows.
For anyone looking to build strong technical skills from scratch, this is a clean, beginner-friendly place to start.
⚡️ GPT-5.1-Codex-Max: Windows-native coding power with breakthrough longevity
OpenAI has released GPT-5.1-Codex-Max, a new coding model built to counter Gemini’s momentum and early impressions point to major capability jumps with a few unanswered questions.
🔸 Full Windows + PowerShell awareness marks a first for Codex, with native understanding of Windows paths, environment quirks, and a new Agent mode that can autonomously operate inside the terminal.
🔸 Long-horizon autonomy is the headline, with OpenAI claiming the model can work continuously for 24+ hours on multi-step tasks, rivaling Anthropic’s 30-hour benchmark and signaling that long-context scaling is far from saturated.
🔸 “Compaction” enables massive effective context, letting the model summarize older information and roll it into a fresh window, a kind of short-term/long-term memory hybrid that supports multi-cycle workflows.
🔸 New SOTA coding performance, hitting 77.9% on SWE-bench Verified and outperforming Gemini 3 and Claude Sonnet 4.5, while using ~30% fewer tokens thanks to more efficient reasoning paths.
GPT-5.1-Codex-Max looks like OpenAI’s strongest engineering model yet but real judgment will come once its autonomous Windows agent and API rollout hit broader production use.
📊 10 key takeaways on how AI is reshaping development, GitHub Octoverse 2025
🔸 Developer growth is booming: GitHub added 36M new developers in 2025; 80% adopted AI tools immediately.
🔸 TypeScript takes the crown: surpassing Python as the most popular language.
🔸 AI repos explode: over 1M AI-focused repositories now exist.
🔸 India leads: 5M new Indian developers joined GitHub this year.
🔸 Commits hit 1B: a 25% increase over 2024.
🔸 Faster vulnerability fixes: automation cuts fix time from 37 to 26 days.
🔸 Python dominates AI repos: still the go-to language for AI projects.
🔸 Six languages rule new repos: Python, JS, TS, Java, C++, C# cover ~80% of new projects.
🔸 Top coding error: Broken access control, AI often overlooks authentication checks.
🔸 Productivity up: more bugs fixed and pull requests merged thanks to AI-assisted development.
AI isn’t just assisting developers, it’s accelerating growth, speeding fixes, and shifting language popularity across the platform.
🚲 Citi Bike now charges for pedal assistance, welcome to the new “freemium mobility”
Citi Bike has rolled out a new paid feature: pedal assistance on its e-bikes. A small boost is free for the first 45 minutes, after which it costs $0.25 per minute. Full assistance? That’s $0.25 per minute from the start.
🔸 It’s basically “microtransactions for biking”, a tiny upgrade that suddenly isn’t so tiny when the meter starts running.
🔸 The pricing model mirrors classic app economics: a free taste, then pay-as-you-go for the real power.
🔸 And surprisingly, there’s still no extra subscription tier… which probably means they just haven’t gotten around to launching it yet.
Mobility keeps finding new ways to upsell even if the upsell is literally your own legs.
⚠️Cloudflare blames massive internet outage on a “latent bug”
According to TechCrunch, Cloudflare’s network suffered a wide-reaching outage affecting major services like ChatGPT, Claude, Spotify and X, caused by a previously undetected bug in a bot-mitigation service that crashed after a routine configuration change.
🔸 The fault arose when the latent bug triggered cascading failures across Cloudflare’s global network; the company confirmed the event was not an attack, but rather a fault in its internal infrastructure.
🔸 Cloudflare acknowledged that the bug went unnoticed during testing and pledged to publish a detailed breakdown of the incident and its remediation steps.
🔸 The incident underscores how dependent the internet is on a handful of infrastructure providers: Cloudflare serves around 20% of all websites, making even a single error capable of major disruption.
For startups and infrastructure teams, the outage is a stark reminder that resilience isn’t just about external attacks, latent internal bugs in core services can bring down the ecosystem.
🧹 A robot vacuum that doubles as a real-time metaverse scanner
A new “metaverse-grade” robot vacuum is making the rounds online not because it cleans well, but because it can reconstruct full 3D environments in real time using onboard AI.
🔸 The design stands out immediately: rugged, futuristic, and equipped with oversized wheels that let it climb over obstacles and even operate outdoors.
🔸 The real magic is in the head-mounted sensor array, a voxel-based neural pipeline running on NVIDIA Jetson Orin, generating photorealistic 3D maps as the robot moves.
🔸 With a bit more resolution, this is essentially a portable metaverse scanner, capturing full spatial meshes of homes, offices, and outdoor areas on the fly.
🔸 The entire software stack is built in Rust, includes day/night SLAM, dynamic obstacle detection, and low-latency path planning, letting the robot squeeze into tight corners and navigate complex spaces.
If robot vacuums become high-fidelity 3D mappers by default, the real metaverse hardware platform might quietly roll in from under our furniture.
💬 Jensen Huang on careers: “It’s easier to fall in love with what you do than to find what you love.”
NVIDIA’s CEO shared a surprisingly grounded philosophy on work and motivation:
🔸 “Find what you love” isn’t always realistic. Jensen says he didn’t start out passionate about any particular job, he simply learned to love the work in front of him.
🔸 He recalls loving every job he ever had: dishwasher, busboy, paper delivery, waiter and now decades at NVIDIA.
🔸 The mindset shift: fall in love with the craft, not the title. When you care about doing a job well, passion follows.
🔸 And once you love the work, excellence becomes easier, you naturally push harder, learn faster, and stand out.
A rare reminder from a tech icon: meaning comes from commitment, not from waiting for the “perfect” dream job to appear.
🚐 Uber and WeRide launch fully driverless robotaxis in Abu Dhabi, a first for the Middle East
Uber and WeRide have officially switched on Level 4 fully driverless rides in Abu Dhabi, marking the region’s first commercial robotaxi deployment with no safety driver onboard.
🔸 The service runs on Yas Island, where some UberX and Uber Comfort requests may be matched with a WeRide autonomous vehicle.
🔸 WeRide provides the full self-driving stack, while Uber handles fleet operations like cleaning, charging, and maintenance.
🔸 This launch is supported by a UAE-issued public-road permit, making it one of the few fully driverless ridehail services outside the US and China.
🔸 Uber and WeRide say this is only the beginning, with plans to expand robotaxi services to additional global cities over the next five years.
A major milestone signaling Uber’s shift from experimenting with autonomy to actually deploying it at scale.
🎓 Find Your Grind raises $5M to help students discover non-traditional career paths
Find Your Grind, an education platform focused on helping students explore careers beyond standard college tracks has secured $5M in new funding to expand its product and reach more schools.
🔸 The platform blends assessments, personalized recommendations, and real-world career discovery to help students identify paths that fit their interests and lifestyle.
🔸 Funding will go toward scaling partnerships with school districts and improving the platform’s adaptive learning engine.
🔸 The company says demand is rising as schools seek modern alternatives to outdated guidance counseling models.
🔸 Find Your Grind positions itself as a “future-of-work navigator” for Gen Z, pushing students toward emerging roles, not traditional templates.
A reminder that the next edtech wave isn’t test prep, it’s helping teens figure out who they want to become.
⚙️ GPU prices are about to climb: AI demand is driving costs up
Social networks are buzzing, and it’s true: graphics cards are set to get more expensive due to soaring memory demand from the AI market.
🔸 Memory costs are up 15–25% recently, forcing manufacturers to raise GPU prices.
🔸 Mid-range and high-performance GPUs could jump 15–20%, while budget models may see reduced production.
🔸 AMD has already warned partners of a 10% price increase on all GPUs, including gaming cards.
🔸 Nvidia hiked prices earlier this year (~15%) due to production costs and geopolitical factors, and another increase is likely after New Year.
if you’ve been planning a GPU purchase, now is the time to buy before the next round of price hikes.
🔥 Stay connected anywhere — without roaming.
Teleleo gives you full remote control of your SIM card, so your number works for you no matter where you live or travel.
✔️ With Teleleo platform you can:
● Receive and make calls without international charges
● Receive & send SMS remotely
● Forward SMS to email or messengers
● Plan SMS campaigns
● Set up auto-reply & auto dialer
● Manage multiple devices
● Use USSD commands remotely
● View full logs of calls and SMS
● Connect to the Internet via your own private VPN
● Use Wi-Fi mode
🎁 Your device unlocks 30 days of free full access. Try it first, decide later.
⚡️ Generate websites and apps in just a few clicks, the powerful AI tool Orchids is out now
🔸 Instantly recreates the design of any website and the output looks like a real human-made clone, not an AI imitation.
🔸 Builds full sites and apps up to 3× faster than competing generators.
🔸 Completely free to use at launch, making it one of the most accessible builders on the market.
The bar for AI-generated frontends just jumped and Orchids wants to become the go-to tool for instant product creation.
⚠️ Anti-AI-washing is coming, luxury brands will sell “human-made” as a premium feature
Coca-Cola’s fully AI-generated New Year’s ad sparked a wave of backlash not because it was bad, but because it broke a cultural symbol people expected to be handcrafted, nostalgic, and human. That reaction is the first hint of a counter-trend that always emerges when a technology goes mainstream: consumers start craving the opposite.
🔸 Just like “organic,” “handmade,” and “non-GMO” became marketing shields (even when the claims were mostly symbolic), brands will begin loudly distancing themselves from AI to position as premium.
🔸 Expect luxury labels to announce zero-AI creative pipelines, human-only craftsmanship, or no-AI customer interactions as a status signal even if the real operations still quietly use AI in the background.
🔸 Many companies will lightly tweak one process to qualify for a new label: the same way “handmade” often means one human touched it once, “anti-AI” may mean one non-essential task was done manually.
🔸 A smaller group of brands will go fully AI-free as a true differentiator, carving out a “human-made luxury” niche for customers tired of algorithmic sameness.
🔸 As AI becomes ubiquitous and cheap, non-AI becomes rare and in luxury, rarity is the whole business model.
The irony is perfect: the AI boom will turn the everyday human touch into a premium product.
🎧 Lumia launches a $250 smart earring, the smallest wearable yet
Lumia has unveiled a $250 smart earring that packs full biometric sensing into a module weighing under 1 gram, claiming the title of the smallest wearable on the market. Only one earring contains the sensor it sits behind the ear, pressed against an artery that carries blood to the head.
🔸 The device tracks blood-flow changes to infer stress, focus, fatigue, sleep/wake cycles, and overall cognitive load.
🔸 Temperature sensing and other standard wearable metrics are included, though more advanced claims (menstrual cycle, infection detection) remain unproven.
🔸 The key breakthrough is miniaturization: the module detaches for charging, while the earring itself stays on.
🔸 Real-world concerns remain: battery life is advertised as 5–7 days, but the tiny form factor raises doubts.
🔸 Blood-flow sensing behind the ear is promising but new, accuracy issues and signal glitches are possible.
🔸 The form factor naturally limits the audience; without redesigns, the product skews toward users who already wear earrings.
Lumia’s concept is bold: the best wearable is the one you forget you’re wearing, but ultra-mini devices create as many questions as they solve.
🤖 OpenAI drops group chats to steal back attention from Google
For the second day in a row, OpenAI is rolling out new features at full speed, this time unveiling group chats in ChatGPT, right as Google dominates headlines.
🔸 The feature is straightforward: you create a chat, invite others, talk as usual, and pull the model into the conversation whenever needed.
🔸 It’s handy for small project discussions or collaborative problem-solving, but feels like a niche use case for most workflows.
🔸 The update is already live for all paid ChatGPT users.
A timely release and a clear attempt to redirect some of the spotlight back to OpenAI.
🎨 Google lit up the Las Vegas Sphere with a Gemini-powered McLaren showcase.
Ahead of the F1 Grand Prix this weekend, Google and McLaren unveiled a massive Sphere takeover where Gemini redraws the McLaren race car in a series of shifting artistic styles.
A flawless example of Google landing the perfect collab at the perfect moment.
📊 Powered by Crypto Insider
🎓 100+ free neural network courses now on GitHub for AI enthusiasts
A massive collection of over 100 neural network courses has been gathered on GitHub, ranging from Stanford’s top AI classes to tutorials on fine-tuning AI agents.
🔸 Covers theory, coding exercises, and practical projects for all skill levels.
🔸 Includes both foundational courses and cutting-edge materials for large model training.
🔸 Fully free and open, making it easier than ever to upskill in AI development.
For anyone looking to learn AI from scratch or level up, this is now a one-stop repository.
🧪 Google’s Antigravity IDE: clever features meet rough edges
Google has released Antigravity, a free AI IDE on Mac, Windows, and Linux, giving users 50 requests every 5 hours. Early users are impressed but also frustrated.
🔸 Standard editing, autocomplete, and agent chat work as expected, but the overall experience feels raw, similar to Cursor.
🔸 The standout is Inbox mode, an asynchronous panel where you assign tasks to multiple agents, track progress, and run projects in parallel or queue.
🔸 Chrome integration lets agents launch apps, browse sites, test elements, take screenshots, and more, appearing more capable than Cursor.
🔸 Weak points include GitHub/git support, occasional macOS lag, and VPN logins that can stall indefinitely.
Antigravity is promising, but early adopters will need patience while Google irons out stability and integration issues.
⚡️ Grok 4.1 rolls out with major upgrades in empathy and accuracy
According to early testers, the new Grok 4.1 feels noticeably more emotionally aware and expressive, with sharper creative writing and responses that better read the mood of the conversation.
🔸 The model now delivers state-of-the-art performance on EQ-Bench, a test focused on soft-skill reasoning and emotional intelligence, though there’s still no direct comparison to GPT-5.1.
🔸 The biggest breakthrough: Grok 4.1 hallucinates three times less than the previous version, marking one of the most meaningful improvements in reliability to date.
🔸 On the competitive side, Grok’s “Thinking” variant reclaimed the top of the LMArena leaderboard, surpassing Gemini 2.5 Pro by 31 points.
The update is already live and for many users, the real upgrade isn’t just smarter answers, but far fewer wrong ones.
✨ 7000 ready-made n8n workflows to automate your job with AI
A massive collection of 7,000+ n8n automation templates just dropped, the largest guide so far for applying AI directly to your role, tasks, and daily operations.
🔸 Each workflow comes with step-by-step instructions, so you can implement automations without guessing or engineering everything from scratch.
🔸 Templates cover dozens of professions from marketing and sales to engineering, HR, operations, finance, support, and solo-founder use cases.
🔸 The library includes both AI-powered flows (ChatGPT, Claude, Gemini, local models) and classic n8n automations for email, CRM, analytics, and data pipelines.
🔸 For teams, it serves as a plug-and-play playbook: browse your job area, copy a workflow, adapt it in minutes, and ship automation that normally takes days.
This isn’t just a template pack, it’s a ready-to-deploy automation encyclopedia that can upgrade an entire workflow stack overnight.
⚠️ Chinese open-source models quietly become Silicon Valley’s hidden AI backbone
A new Bloomberg analysis highlights how much of the U.S. AI boom is now running on top of Chinese open-weight models, a shift driven by cost, licensing freedom, and sheer model quality.
🔸 Alibaba’s Qwen family has taken over Hugging Face: more than 40% of newly uploaded LLMs in 2024–25 are Qwen-based derivatives, dwarfing Meta’s LLaMA line.
🔸 Multiple U.S. startups are suspected of using Chinese base models under the hood: Cursor is linked to DeepSeek, while Cognition’s Devin is rumored to lean on Zhipu AI’s Z-series.
🔸 Developers cite price and permissive licensing, Qwen models are cheap to run, fully open-weight, and allow commercial fine-tuning without legal friction.
🔸 The trend is big enough that Nvidia’s Jensen Huang has warned China may “win the AI race” through developer scale even if the U.S. leads in advanced chips.
As Chinese models quietly become the substrate for U.S. AI products, the question isn’t just technical, it’s geopolitical: who truly owns the future stack that Silicon Valley is building on?