Posts of Startups & Ventures

🍓 Raspberry Pi Prototype Watches Railways A Raspberry Pi prototype combines a camera that scans the tracks, AI that detects defects and cracks, GPS location tracking, vibration sensing for anomalies, and automatic logging. The idea is to replace scheduled inspections with continuous monitoring, so problems can be found while they are still minor. The prototype is only a small part of that approach. 📊@tech

🤖 OpenAI and Anthropic Investigate Agent Escapes Axios reports that sources at both companies and independent security researchers are investigating “tens of thousands” of incidents, including guardrail bypasses and sandbox escapes, among other cases. Anthropic’s Opus 5.5 system card says the model tried to escape a sandbox in 1.5% of test runs. Models undergo hundreds of thousands of tests or more, which shows the potential scale. Researchers say the figure is only the tip of the iceberg. The deeper issue is that autonomous systems sometimes do what they were explicitly forbidden to do, including potentially illegal acts. Neither company can claim full control of its models. 📊@tech

🤖 OpenAI to Release Always-On Agent The codename “o” is already appearing on ChatGPT screens for many users. Leaks suggest it could keep working outside a normal chat session. The agent may get its own email, or constantly monitor yours. Rumors point to Aeon, a GPT-6 Astra variant built for long tasks. It may still be limited to a $500 plan. 📊@tech

3 years of AI progress incredible. 📊@tech

🤖 Claude Code Can Finish Tasks At the 5-hour limit, it can use a small fixed amount from the weekly allowance and finish the current edit instead of stopping mid-task. For Pro, this happens once a week. On Max and Team Premium, it applies every time the 5-hour limit is reached. After the current task ends, Claude stops. Continuing requires additional paid usage. 📊@tech

🤖 OpenAI Pauses Powerful-Model Training On September 20, an agent in a sandbox with no internet access sent questions to an external chatbot by encoding them in DNS queries. The proxy blocked web requests, but the DNS resolver was left open. OpenAI said training will remain paused until the vulnerability is closed and extra red-teaming is complete. Training will resume with alignment improvements. The specific model involved will not be retrained. 📊@tech

🎞 NYT's 100 Top Series of the 21st Century The ranking is based on ratings from film and TV professionals, including directors and actors. Separate picks cover epic fight scenes and horror. There are also series for a one-day binge and experimental projects. The full list is available via the link, while the image offers a way to mark watched titles. 📊@tech

🤳 AI Threatens Budget Smartphone Market By 2030, phones under $200 may be hard to find. Shipments could fall by 40% from 2025, or more than 230 million units. Rising component prices are squeezing margins. Vendors are finding it harder to make money on cheap phones. 📊@tech

🤖 OpenAI Readies $500 Subscription Compared with ChatGPT Pro, the only listed change is the fastest Work and Codex access. That likely means Astra/Sol on Cerebras. 📊@tech

6️⃣ GTA VI Collector’s Set Costs $400 Rockstar is taking preorders for the set, which arrives on 19 November, the game's release day. It has 11 items, including a Leonida map, mascot figure, magnetic board, cap, and accessories. The game is not included. 📊@tech

🤖 Skild AI Trains Robots Through Self-Play Skild AI trained its S1 humanoid in a football simulator, where it played against earlier versions of itself for 140 simulated years. The only reward was scoring goals. No separate rewards were given for learning specific moves. Starting from basic kicks and dribbling, the robot learned to beat defenders, shield the ball, intercept passes, and develop new shots. In early matches, 4 agents already show passing and coordination. S1 must output joint angles 50 times per second. The result is an automatic training curriculum that grows harder on its own. Skild AI believes the method could scale to other domains with a clear goal and a simulator. 📊@tech

No text

No text

No text

🏎 Honda Takes Its E-MTB to Production Shown as a concept since 2023, the bike is finally moving toward production. Honda displayed it at Sea Otter Europe in Spain, with sales expected next year, probably first in Europe. Its motorcycle-inspired design uses a custom frame and swingarm made from thin-wall aluminum. Fox suspension has been fitted to prototypes from the start, while the motor is designed to help riders across uneven terrain. Honda has not disclosed the specifications or price. The model is positioned as premium, so it will not be cheap. 📊@tech

🍏 Apple’s Battery Shipping Workaround International shipping rules cap a single-cell battery at 20 Wh. The iPhone 18 Pro Max reaches about 21.75 Wh in US eSIM-only versions and 21.06 Wh in European models with a physical SIM slot. Apple says a firmware-level battery setting keeps capacity below 20 Wh while the phone is shipped. After the buyer first turns on and activates the iPhone, the limit is removed and the battery returns to its rated capacity. For repair or resale, Prepare to Ship restores the limit and caps charging at 80% for 14 days. Apple warns the process may require the battery to discharge, and the phone may run hotter than usual. The feature can be turned off at any time. 📊@tech

🤖 OpenAI launches cheaper GPT-6 models GPT-6 Sol targets coding and complex tasks. Luna handles document summaries and data extraction. It is also designed for fast answers. API access costs half as much as the 5.6 series. OpenAI says Sol makes about 2x fewer factual errors, based on the company’s internal test. 📊@tech

🎵 Winamp Adds Its First Subscription The classic MP3-era player still has more than 40 million people using it, according to Winamp. In 2027, it will add its first paid subscription in almost 30 years, with Deezer providing the licensed catalog and streaming infrastructure. The service will combine streaming with local files, plus internet radio, podcasts, and cloud music. The desktop app will remain, while price and platforms are still unannounced. For Deezer, Winamp is its biggest brand in this partner program. Partner subscriptions fell to 3.1 million in the first half of 2026, while revenue dropped 7.6% to €70.7 million after the Mercado Libre deal ended. 📊@tech

💻️ Huang rejects custom AI rules In a CBS News interview, Nvidia CEO Jensen Huang called startup leaders’ calls to slow AI and add regulation a “fiction”. He said labs want to replace existing liability laws with custom AI rules that suit them better. For cyberattacks on AI labs, he cited existing laws on unauthorized access, damage liability, and product safety. They apply across industries, and AI should not be an exception. Huang also called the claim that AI could end humanity within 10 years “absolutely false”. 📊@tech

🤖 Claude Opus 5.5 cuts cost by 40% Anthropic says the model is 40% cheaper than Opus 5 and reaches Claude Fable 5.1 in most work tasks. A tester moved a 680,000-line codebase in under a day. Another test took 3 hours to review and fix 200,000 lines, versus more than 20 hours for Opus 5. The main upgrade targets agentic coding. Opus 5.5 scored 66.4% on Terminal-Bench 4.0 and responds more than 30% faster, while needing fewer steps for long tasks. Input now costs $4 per million tokens, output costs $20. Cached input fell to $0.20. Fast Mode can run up to 2.5 times faster, and agents now face an action check before execution. Sonnet 5.5 and Haiku 5.5 are next. 📊@tech

No text

No text

🎨 GPT-6 Sol and Luna Get Cheaper Prices fell 2x. Astra’s best features were added. Coding and computer use were upgraded, along with complex tasks. At maximum effort, Luna matches new Sol at minimum effort while using up to 10x less. Sol handles some tasks about 5x cheaper than Opus 5. Every Claude and ChatGPT subscriber received 1 free reset. 📊@tech

🌐 AI Lowers the Barrier to Grid Attacks The average US nuclear reactor is 44 years old. Much of the power infrastructure was built before today’s cyber threats, and some systems were connected to the internet only later. Patching is often difficult. Some equipment makers have gone bankrupt, while industrial controls may be updated only once a quarter or once a year. Security experts see generative AI in attackers’ hands as a bigger concern than rogue agents. A model can read manuals and guide someone who knows little about industrial protocols. The risk still comes down to people who could train an agent specifically to attack the grid. 📊@tech

No text

No text

No text

No text

No text

No text