4273207
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).
💻 Turn any repo into an AI prompt
GitIngest lets you feed an entire GitHub repo into an LLM in seconds.
🔸 Drop in a repo link and wait a moment
🔸 Get back a clean, structured prompt with the code and files
🔸 Edit it to trim noise and keep only what matters
A neat shortcut for devs who want their AI to grok a codebase fast.
💼 OpenAI to launch AI-focused jobs platform
OpenAI has announced OpenAI Jobs Platform, a hiring service where companies can find talent skilled in working with AI - from beginners to expert prompt engineers.
🔸 Platform will match business needs with candidates’ AI skills
🔸 Job seekers can take OpenAI’s own certification exam directly inside ChatGPT via Study mode
🔸 Official launch is planned for 2026
A move that could make OpenAI not just a model provider, but also a gatekeeper for the AI labor market.
🧠 Why LLMs hallucinate, according to OpenAI
OpenAI just published a rare research piece on one of the hottest issues in AI: hallucinations. The takeaway - they’re not mysterious at all, but a natural byproduct of how we train and test models.
🔸 Pretraining phase: models are forced to always predict the “most likely” next token. Saying “I don’t know” isn’t an option, and there’s no penalty for making things up.
🔸 World randomness: some facts (like birthdays or serial numbers) are inherently unpredictable. Models can’t “learn” them.
🔸 Benchmarks: most evals score wrong and skipped answers the same — 0. This incentivizes guessing over admitting uncertainty.
🔸 Garbage In, Garbage Out: data errors inevitably feed into outputs.
OpenAI’s fix? Change evaluation itself. Add “I don’t know” as a valid response, and reward honesty over confident fiction. With confidence thresholds (e.g. only answer if >75% sure), models would learn that admitting uncertainty beats hallucinating.
📊 Powered by Crypto Insider
🛠 Build your own internet stronghold
A full toolkit of guides to help you become a truly independent internet user.
🔸 Launch a website on your own domain for free in just a few hours
🔸 Set up a private mail server and ditch Gmail
🔸 Run your own chat, social network, streaming platform - even a Monero node
Take control and build a fortified corner of the web.
🗣️ Altman vs Musk - the saga continues
In a new interview, Sam Altman was asked when he chose the world over corporate advantage.
His answer:
Well, we still haven’t added sex avatars to ChatGPT.
🧮 Solve your math in a few clicks
Mathos is a simple helper for students that turns tough problems into clear, visual explanations.
🔸 Covers algebra, geometry, trigonometry, linear algebra, calculus, probability, and statistics
🔸 Accepts text, photos, PDFs, or voice input
🔸 Shows step-by-step solutions and builds graphs for clarity
🔸 Free to use
If you are buried under problem sets, this will save hours.
🤖 Figure teaches Helix to load a dishwasher
Figure’s humanoid model Helix is showing how far general-purpose robotics can go. The same system that folded towels and sorted packages now loads dishwashers - no new algorithms, only new data.
🔸 Handles stacked plates, glasses, cluttered states
🔸 Uses two arms in sync for fragile reorientation
🔸 Recovers from slips and collisions without breaking flow
🔸 Learns new skills incrementally across domains
Loading a dishwasher might sound simple, but for robotics it combines perception, precision, and adaptability. Helix’s progress is another step toward humanoids that scale beyond demos into real-world use.
📞 Assort Health raises $26M to automate clinic calls
Assort Health is betting that millions of Americans will still pick up the phone to book a doctor. The startup built a generative AI platform that replaces human operators, handling patient calls and syncing them directly into clinic schedules and EMR systems.
🔸 Already live with 50+ clinics
🔸 Works 24/7, no apps required, just a phone number
🔸 Cuts admin load and reduces errors
🔸 Just closed $26M at a $750M valuation
The catch: phone-based scheduling is a shrinking channel. In cities, apps and AI agents already handle bookings. The long-term question is whether Assort is tapping into a sticky, underserved demographic or simply cashing in on AI hype before the phone goes extinct.
Would you call this smart timing or a misaligned bet on a dying habit?
📊 How startups feel in 2025
Mercury surveyed 1,500 founders and the results paint a mixed but telling picture of today’s startup economy.
🔸 87% feel more confident than in 2024, only 3% say things are worse
🔸 Startups with AI are far more optimistic (93% vs 71%)
🔸 27% raised over $5M, often combining VC with loans, grants, or RBF
🔸 61% are hiring, but among $5M+ revenue companies there are no solo founders
🔸 80% expect costs to rise in 2025, mostly due to AI and automation
AI has shifted from optional to required, funding sources are diversifying beyond pure venture, and teams are the real bottleneck to scale.
Optimism is back, but founders who overextend on AI spend or inflate metrics may find 2025 just as risky as 2022.
🗣️ Grok code theft scandal
A former xAI engineer quit after allegedly stealing the entire Grok codebase and joining OpenAI. Musk confirmed the breach.
🔸 Before leaving, the employee cashed out $7M worth of xAI stock
🔸 xAI has already filed a lawsuit
🔸 Musk called it outright theft of company IP
Another chapter in Musk’s chaotic AI battles and a case study in how messy talent wars in this space are getting.
⚡️ Hyperdrives powers up with €3M for next-gen EV motors
Munich-based Hyperdrives has raised €3M pre-seed to bring motorsport-level performance to mass-market electric motors. The round was led by Rethink Ventures with First Momentum, SDAC, Acequia Capital, 2100 Ventures, Prototype Capital, and angels.
🔸 Breakthrough tech: hollow-conductor cooling that channels coolant directly into copper windings, tripling current density compared to standard motors.
🔸 Proven traction: bootstrapped through 3 product generations, tested under harsh racing conditions, and hit €1M in 2024 revenue before outside capital.
🔸 Market scope: automotive (Hyperdrives One with SiC inverter), electric aviation, and marine.
🔸 Why it matters: combines record power density with scalability and cost efficiency, a critical bottleneck for mass EV adoption.
With fresh funding, Hyperdrives will validate durability, expand customer pilots, and prepare industrial-scale licensing deals with OEMs and Tier-1 suppliers. A startup blending racing DNA with OEM pragmatism is now chasing global electrification.
📊 Powered by Crypto Insider
🥷 Murder shocks Burning Man, Silicon Valley’s desert playground
A homicide investigation has cast a dark shadow over Burning Man after a man was found dead during Saturday’s effigy burn. The timing is striking, as the festival has become more than a counterculture ritual - it is a networking hub for tech founders and investors.
🔸 Burning Man has long served as Silicon Valley’s offsite. Musk once said “Burning Man is Silicon Valley.” Zuckerberg flew in to hand out grilled cheese. The first-ever Google Doodle was literally a Burning Man out-of-office notice from Page and Brin.
🔸 The event attracts not only billionaires but also early-stage founders and VCs, who see it as a place to build relationships away from boardrooms. Deals, angel checks, and even company ideas have famously started in the dust.
🔸 The homicide disrupts that mythos. Instead of a playground for serendipity and founder bonding, Black Rock City briefly became a crime scene with authorities warning participants to “stay vigilant.”
For years, startups have mythologized Burning Man as the desert where innovation and counterculture blur.
This week showed the other side of that fantasy, when your networking hub is also a temporary city, it inherits all the risks of the real world.
📝 Make ChatGPT sound more human in seconds
No external tools needed. You can adjust it right inside ChatGPT.
🔸 Open ChatGPT
🔸 Click your profile photo → “Customize ChatGPT”
🔸 In the Traits field, paste this prompt:
Write in natural, human-sounding English. Avoid the AI tone: overly formal, polished, or generic phrasing.
Do not use long dashes, excessive quotation marks, corporate jargon, or bureaucratic language.
Choose simple, clear wording. Conversational style is fine if it helps convey the idea.
Don’t repeat the same phrases or overcomplicate sentences without need.
Vary sentence length and rhythm so the text feels alive.
The priority is clarity of meaning, individual style, and practical value in every line.
Each sentence should feel intentional, not mechanically generated.
🔊 Sam Altman on retaining talent as startups scale
Sam Altman says the biggest mistake founders make is failing to shift from recruiting to retaining as the company grows. Early on, hiring dominates. Later, retention becomes life or death for the business.
He recalls Zuckerberg’s rule: only hire people you would report to if roles were reversed. But if you do not make their role strong enough to stay in for the long term, the best people will leave.
Altman’s tactical advice for CEOs:
🔸 Spend one on one time with your best 5–10 people, through dinners, drinks, and real connection
🔸 Keep giving them more responsibility, because growth is retention
🔸 Proactively re-up compensation instead of waiting for them to ask
To keep your best people, treat them like co-founders, not just employees.
📷 DJI’s leaked Osmo Nano could shake up action cams
Leaks suggest DJI is working on the Osmo Nano, a tiny magnetic action camera with a detachable pod, a design that looks a lot like Insta360’s Go line. T
he pod can stick to hats, shirts, or even a dog’s collar, while the main unit offers an OLED screen and storage options.
🔸 Magnetic pod attaches in any orientation
🔸 64GB or 128GB storage plus microSD support
🔸 Designed for POV shots from unusual angles
🔸 Successor to DJI’s modular Action 2
If the leaks are real, DJI is pushing the form factor forward, while GoPro risks being left behind.
🛒Waitrose tests AI-powered smart carts
Waitrose is trialling AI trolleys in its Bracknell store that automatically scan items as shoppers put them in, aiming to cut checkout times.
🔸 Built by Israeli firm Shopic, the carts use cameras and a screen to track products and prices in real time
🔸 Payments are handled in-app, letting customers walk out without scanning barcodes
🔸 Shopic claims the tech saves up to 8 minutes compared with manned tills and reduces errors at self-checkouts
🔸 Trials follow Amazon’s cashierless Fresh stores and other supermarket automation pushes
If successful, “smart-carts” could make the self-checkout queue a thing of the past.
🗂 72 AI tools in one cheat sheet
A fresh list is making the rounds on Twitter — 72 of the most useful AI services across key categories:
🔸 Image generation
🔸 Video generation
🔸 Text writing
🔸 Research & analysis
A handy one-stop cheat sheet for anyone working with AI. Worth saving.
🦷 Japan begins human trials for tooth regrowth drug
Japanese researchers have started Phase 1 trials of TRG-035, a drug that could regenerate lost human teeth. Conducted at Kyoto University Hospital, the study is a major step toward regenerative dentistry.
🔸 TRG-035 blocks the USAG-1 protein, reactivating dormant tooth buds
🔸 Early animal trials in mice, ferrets, and dogs showed successful new tooth growth
🔸 Human trials involve 30 adults missing at least one tooth, testing safety and dosage
🔸 Potential use cases include congenital tooth agenesis in children and age-related tooth loss
🔸 Commercial availability is projected around 2030 if trials succeed
If proven safe and effective, TRG-035 could transform dental care - shifting from implants and dentures to natural regrowth.
🤖 Teaching a robot to ride, jump, and flip
At the RAI Institute, researchers are building the Ultra Mobility Vehicle (UMV), a bike-like robot designed to combine the speed of wheels with the agility of legs. The goal: give machines athlete-level movement through AI and reinforcement learning.
🔸 The robot’s lower half is a carbon fiber bike frame, while the upper “head” holds motors, compute, and power for jumps and flips
🔸 It weighs ~23 kg, can extend from 80 cm to 152 cm, and is engineered to survive crashes and rough landings
🔸 Sensors include IMUs, LiDAR, joint encoders, and cameras to track terrain and maintain balance
🔸 Reinforcement learning in simulation trains the robot for stunts like wheelies, continuous hopping, and one-meter flips
🔸 The team tackles the “sim-to-real gap” by randomizing parameters in training and modeling physical systems like motors and batteries
By blending robotics, AI, and athletic intelligence, UMV hints at machines that could one day navigate extreme terrain as effortlessly as human trial bikers. The next frontier: chaining skills into complex autonomous sequences.
🚕 Tesla opens Robotaxi to the public
Tesla has launched its Robotaxi app for iPhone users, letting anyone join the waitlist for self-driving rides. Until now, access was limited to influencers and celebrities.
🔸 Service currently works only in Austin
🔸 All rides still have a human safety observer in the passenger seat
🔸 In Silicon Valley, Tesla is testing a different Uber-style format with a driver behind the wheel overseeing the system
The move marks Tesla’s first real step toward scaling autonomous ride-hailing beyond PR stunts.
📱 AI Key wants to run your iPhone for you
Dafdef, a San Francisco startup known for its AI-powered smartphone browsers, has launched AI Key - a USB-C dongle that plugs into your iPhone and acts as a portable AI agent.
🔸 Works like a “mouse with eyes,” tapping, swiping, and typing across your apps
🔸 Handles tasks from booking hotels to ordering food, with step-by-step approval
🔸 Plans can be edited mid-flow — you stay in control at all times
🔸 No access to phone memory or payment data, reducing security risks
🔸 Already used by founders, creators, and researchers to cut repetitive work
A quirky step toward giving AI direct control of smartphones. The real question is whether users will trust an external agent with this much screen time.
📊 Where the money flows in US startups
Crunchbase compared startup funding by state in 2025, and the picture is one-sided. California has pulled in well over $100B already, leaving New York, Massachusetts, and Texas far behind.
For founders chasing capital, the US map is simple. There’s California and then everyone else.
💡A new culture in Silicon Valley
In past decades, "hustle culture" was optional. Today for 18–25 year olds in the Valley it has become the norm. Bars, drugs, and hookups are replaced with 12-hour workdays, 6–7 days a week.
Not grinding is what’s “uncool.”
This isn’t about a few outliers, it’s widespread and culturally reinforced. It means this generation may achieve far more than those before them.
The open question is what the long-term cost will be to health and psychology.
Or maybe that’s just how we justify not working as hard.
📊 Powered by Crypto Insider
❕ China enforces AI content labeling
Starting September 1, China requires all AI-generated content from text and images to audio and video to be labeled. Platforms must apply both visible tags and hidden watermarks or metadata.
🔸 WeChat, Douyin, Weibo and RedNote have already rolled out the rules
🔸 Users must disclose AI use or risk account bans
🔸 Weibo now allows people to report posts without AI tags
For startups and creators this sets a precedent: AI output is now regulated at the content level, not just the model level.
📖 Free AI courses from Anthropic
Anthropic has released a set of free courses that explain AI fluency in simple, structured ways. Worth checking out whether you’re teaching, learning, or just trying to catch up with the basics.
🔸 AI Fluency: Framework & Foundations
🔸 AI Fluency for Educators
🔸 AI Fluency for Students
🔸 Teaching AI Fluency
All courses are free and, in many cases, more useful than paid ones.
📊 Powered by Crypto Insider
💶 Scramble makes investing in consumer brands simple
Scramble is a European platform that lets anyone back fast-growing consumer companies and earn from it - without the noise of stocks or daily market tracking.
Each month, users get a curated batch of companies, and investments are automatically diversified.
🔸 Founder-backed loans, disbursed in stages to reduce risk
🔸 Group A: monthly repayments, up to 12.4% annually
🔸 Group B: higher risk, up to 25% returns
🔸 Scramble co-invests up to 20% in every batch
✨ New users get €10 for signing up, plus €5 bonus for every €100 invested.
💐 What Taylor Swift’s engagement teaches founders about hyper growth
When Taylor Swift’s engagement racked up 1.2M likes in ten minutes, it was not just pop culture. It was a masterclass in scaling fast. The playbook is surprisingly startup friendly.
🔸 Timing is everything: Swift stacked the moment on top of NFL buzz, a viral podcast, and album rumors. Stripe does the same by syncing launches with its developer conference.
🔸 Community as a growth engine: Swifties spread memes and clips, pushing engagement viral. Monzo and Robinhood tapped the same energy with forums and referrals.
🔸 Smart pivots: The engagement doubled as a market move, like Twitch’s shift to gaming or Hopin’s pandemic era pivot to virtual events.
🔸 Network effects: Swift’s team tracked sentiment and adapted in real time. Notion and Airbnb grew the same way, evolving through user input and feedback loops.
The lesson: cultural moments and startup hyper growth follow the same rules. Nail timing, build community, pivot with intent, and amplify network effects. That is how you turn buzz into unstoppable momentum.
🎥 Krea unveils real-time AI video tool
Krea introduced a real-time generator that lets users control outputs by sketching simple shapes. The result feels like img2img extended into video, with temporal consistency across frames.
🔸 Works at 12 fps in real time
🔸 Maintains coherence by referencing prior generations
🔸 Not quite video2video, but something new in between
🔸 Perfect for concept art, experimental clips, or music videos
👉 Join the waitlist here 👈
🤝 OpenAI x Anthropic cross-tested each other’s models
Earlier this summer, before GPT-5 launched, the two AI giants ran each other’s public models through their own internal safety tests. The idea was to check “raw” alignment without external filters.
🔸 Reasoning models (OpenAI o3, o4-mini, Claude 4) proved far more resilient, harder to jailbreak and better at refusing unsafe tasks
🔸 Classic chat models (GPT-4o, GPT-4.1) sometimes slipped, offering help with dangerous requests like drug or weapon instructions
🔸 Most models showed sycophancy, agreeing with users even in dubious scenarios, except o3
🔸 Anthropic models leaned toward refusal under uncertainty, while OpenAI models answered more often but risked higher hallucinations
Cross-testing exposed the blind spots that guardrails usually hide. If this becomes an industry standard, it could redefine how safety is measured in AI.
🧠 500 free AI tools in one place
A massive hub with 500 neural networks for every use case, all free and accessible right from your browser.
🔸 Text, music, video, voice, photo generators and more
🔸 Includes alternatives to ChatGPT, Midjourney, DeepSeek and others
🔸 No cards, no signup, no ads
👉 Grab it here 👈