Techno-Tricksters

Microsoft is offering voluntary buyouts to more than a quarter of its workforce. Around 60,000 employees, or 27% of the company's staff, will receive an offer to leave voluntarily. In return, they are promised severance payouts and assistance with their future careers. The program covers selected teams in the US, the UK, Canada, and India, including those working on Xbox, Windows, Copilot, and other AI divisions. Eligibility for the offer does not depend on an employee's age or tenure. Source #AITools #ResponsibleAI

Apple has unveiled the new 14-inch MacBook Pro powered by the M6 chip 💻 According to Apple's tests, graphics are up to 30% faster, and local AI tasks run up to 25% faster compared to the M5. The base model comes with 16 GB of RAM, configurable up to 32 GB. The laptop also gets Wi-Fi 7, Bluetooth 6, and up to 24 hours of promised battery life. Pricing in the US starts at $1,999, with pre-orders opening tomorrow and sales starting November 6. Source #Hardware

OpenAI has doubled the usage limits for its Pro subscription plans. Users on the $100, $200, and $500 tiers now get twice the allowance for Codex and Work. Additionally, the previously announced limit reduction for the $200 plan has been canceled. The updated limits are already live. Source #AITools #GenAI

Google just unveiled Genie 4, an AI model that generates interactive 3D worlds complete with sound. The model builds environments you can actually walk around in and modify on the fly using text prompts. You can change the weather, direct the camera, or drop in new characters that talk and react to your actions, while sound effects, music, and dialogue are generated right alongside the visuals. Google also showed an experiment pairing Genie 4 with Gemini Argon to simulate a robot playing checkers — the long-term goal is to use these virtual worlds to train and test robots and AI agents. Right now, a session lasts a few minutes before details start shifting and fading, and complex object interactions are still glitchy. Genie 4 is currently limited to select testers, with a broader rollout planned later for Google AI Ultra subscribers and an open waitlist available now. Source #GenAI #Robotics

The startup Cathedral hacked Anthropic's internal systems in just 12 hours 🔒 During a security audit, Cathedral's AI agents managed to access real internal data, including meeting transcripts, salary records, and Slack chats where Anthropic employees were discussing upcoming models. Cathedral is a corporate cybersecurity startup founded by former members of DOGE, the government cost-cutting initiative overseen by Elon Musk. Anthropic stated that it has since patched the vulnerability and cut off Cathedral's access. User chats and model source code were not compromised. Source #ResponsibleAI #Startups

OpenAI is now giving out API credits to users with a Pro subscription. Depending on your tier, you can get $100, $200, or $500 added to your API account. The company promises to hand these credits out every month. To get them, go to your billing settings, select Link API Organisation, and hit Claim. More details are in the official FAQ. #GenAI #AITools

OpenAI no longer lets the models it is training access the live internet 🔒 Instead, OpenAI's most powerful models now learn to browse the web using a large collection of saved pages. This is meant to reduce the chance of them escaping their sandbox and hacking someone. Separate AI agents are tasked with testing these barriers and finding ways out of the isolated environment. They have already discovered previously unknown vulnerabilities and even figured out how to pass messages to each other through shared processor cache — without any network connection. Source #ResponsibleAI #GenAI

The first episode of "The Odyssey" created in Grok is out. Elon Musk started making this "historically accurate" adaptation because he wasn't a fan of Christopher Nolan's film. It looks great as a trailer, though it seems this clip is meant to be the final product — which makes the whole thing pretty sad 😞 Source #GenAI #AIFun

Anthropic has filed for an IPO, expected as early as the second half of November at an estimated valuation of $1.5 to $2 trillion. The company posted an 8% operating profit in Q2, with an inference gross margin of around 85% across subscriptions and the API. Subscription revenue alone covers all inference costs. In total, Anthropic has spent just $10 billion of investor money so far, leaving a cash cushion of over $131 billion. While the company recorded an $80 billion paper loss in the first half of the year, this is an accounting formality from revaluing convertible debt and stock obligations after its valuation jumped. Excluding these adjustments, Anthropic is already profitable. Customer concentration is also dropping: the top 5 clients now bring in 21% of revenue, compared to a little over a year ago when GitHub and Cursor alone accounted for nearly half. Over 70% of revenue still comes from US clients, while rival OpenAI's IPO remains on hold after being delayed in September. #Startups #GenAI

Kimi K3.1 Pro is out, and it's already beating other Chinese models in benchmarks. The new model comes close to the Opus 5.5 level. Beyond standard tasks, it can work with game engines like Unreal, Unity, and Godot, as well as handle 3D modeling in Blender and CAD. Pricing is set at $2 per million input tokens and $10 per million output tokens. There's also a Highspeed version that runs 2-3 times faster for double the price. Model weights will be open-sourced by November 4. Try it #GenAI #AITools

Steam is drowning in AI slop 😱 Over 8,000 demos were shown at the autumn Next Fest, and a third of them were made using AI. For comparison, three years ago the festival had around 900 games total. Valve isn't really bothering with filtering, hoping the market will sort itself out. As a result, it's getting harder for talented indie developers to break through, and harder for players to find anything worthwhile. It has gotten to the point where someone built a browser extension that highlights games using AI. Try it #GenAI #AITools

A court has ordered Meta to pay $25 billion for misleading users about their privacy and data protection on Facebook. A jury in New Mexico found the company responsible for 43 million violations of consumer protection laws. At the core of the case is the Cambridge Analytica scandal, where the British political consulting firm harvested data from millions of Facebook users without their consent and used it for Donald Trump's 2016 presidential campaign. Meta plans to appeal the verdict. If the ruling holds up, the payout will go toward funding education in New Mexico. Source #ResponsibleAI

China has found a new place to put ads: right on mirrors. While you are standing at a distance, the mirror displays ad banners instead of your reflection. As soon as you step right up to it, the ad disappears and it turns back into a regular mirror. #Hardware

ChatGPT quietly builds a psychological profile of each user. Researchers looked into the chat histories of 80 people and found 2,000 memory entries saved by the bot. In 96% of cases, ChatGPT saved this information without being asked to remember it—the model mostly decided on its own what was worth keeping. While most notes were based on what users actually said, the bot sometimes jumped to conclusions. For instance, asking for bands similar to Nirvana was logged as "the user likes Nirvana," even though the person never said that. Half of the stored entries contained details about users' desires, plans, emotions, and beliefs. Another 7% held highly sensitive data, such as health information. The study itself came out back in February 2026, but it is only now gaining traction on X. Paper #ResponsibleAI #GenAI

📷 Canva has launched its own AI image generator. The model went straight to 4th place on the DesignArena leaderboard and took the top spot for portrait and landscape generation. What it can do: • Generates 2K images across more than 30 aspect ratios. • Creates text on images in 55 languages with accurate placement. • Plans the composition and object placement first before generating the final image. • Splits images into layers via Magic Layers so you can edit individual elements. The model is available in Canva Dream Lab, with the free tier offering up to 20 generations per month. Try it | Source #GenAI #AITools

ChatGPT just added password-protected chats 🔒 You can now lock any chat with a PIN code, which hides it from your history and search results. Getting back into a locked conversation requires entering the password or authenticating via Face ID or Touch ID. The update is currently rolling out to all users. Source #AITools #GenAI

Need a startup idea? AI just generated over a million of them for free 💡 Bill Gross, founder of the Idealab startup incubator, has launched Idealab.ai, a massive catalog of business ideas. According to Gross, the AI analyzed millions of sources to compile 1.2 million potentially viable startup concepts. The search tool is completely free. You just briefly describe your area of expertise and set your budget, and the AI suggests business options that fit your profile. Over time, Gross also plans to add features for finding co-founders and investments. Try it #AITools #Startups

Anthropic shared two guides on how its team uses Claude for background tasks and code reviews. Here are the key takeaways. 1. Automations: agents need to remember where they left off (Source) Save the exact point of the last check to avoid missed events or duplicates. Agents should verify info is still relevant before sending a summary and confirm delivery afterward. Limit spending and permissions, and keep user instructions separate so they cannot be overwritten. 2. Code review: ask for proof that everything works (Source) Give Claude tests and build tools to verify the app. Pass the code to a separate agent with a fresh context for review, logging recurring feedback into project rules. Cap the number of "review-and-fix" loops. Humans should keep control over architecture, risky changes, and final accountability. #GenAI #AITools

Eka Robotics has raised $150 million to build AI for nimble robotic grippers. The startup's system is delicate enough to pick up a raspberry without crushing it. Their model combines computer vision, force sensing, and motion control. This lets the robot adjust on the fly if an object slips or shifts position. It learns these skills through reinforcement learning inside physical simulations, practicing over many attempts and getting feedback on the results. According to Eka, the robot can pick up 1,500 items weighing up to a kilogram per hour, with a success rate over 95%. The company is currently running industrial pilots focused on moving items and handling wires, and the latest funding round values Eka Robotics at $1.45 billion. Source #Robotics #Startups

DeepSeek has open-sourced the engine that powers its models. It is called DeepInfer — the software backbone DeepSeek uses to run its AI models. Right now, it supports DeepSeek V4.1 Flash and runs on two types of hardware: NVIDIA SM100 and Huawei Ascend 950. Its features include faster text generation through token pre-preparation and verification, as well as the ability to output responses in a specific format, such as JSON. GitHub #GenAI #AITools