Anthropic’s 5% Pledge Awaits the Fine Print
AI’s latest competition: convincing everyone else they’ll get a slice.
Pending shareholder approval, Anthropic intends to park 5% of its equity in a new Public Benefit Fund for nonprofits and AI-displaced workers. Conveniently, this sudden burst of philanthropy arrives just as the Claude developer gears up for a massive IPO.
The wholly owned LLC vehicle grants the company broad flexibility in distributing assets (and maintaining total control, naturally). Directors can bump the fund’s stake by two percentage points every time Anthropic’s valuation jumps 50% past a secret baseline. Top-ups aren’t guaranteed, though the firm strictly banned using holdings for political lobbying.
Factoring in founder and staff pledges, roughly 20% of the company’s stock is now officially earmarked for the public good. Yet logistics—like who picks the winners and how locked-up shares become spendable cash—remain a complete mystery. A theoretical stake in tomorrow doesn’t pay this month’s rent.
Not to be outdone in the billionaire philanthropy pageant, Elon Musk declared he’ll give one of his personal SpaceX shares to every American child next year. With zero logistical details offered, parents can safely hold off refreshing their kids’ brokerage accounts.
In a separate peace offering following backlash over unverified AI math research, OpenAI committed $5 million to a $15 million Simons Foundation initiative. The grant—also supported by Anthropic’s computing resources—helps human mathematicians evaluate AI tools, though it conveniently ignores ongoing attribution disputes.
The tech sector’s generosity announcements are arriving much faster than the instructions for collecting.
Accused OpenAI spy allegedly had ties to Saudi Arabia
New evidence shows that Abdulrahman Alfozan — the ex-OpenAI employee accused of bugging a conference room — had been communicating with the Saudi royal family and had plans to permanently move back to the country (full details). Separately, OpenAI shipped two updates to ChatGPT Slides: real-time collaboration and in-document editing. Learn more about the updates or explore prompts to start creating with ChatGPT Slides.
US government bans DeepSeek from federal civilian agencies
The move marks Washington’s latest escalation yet to try and keep Chinese technology out of critical infrastructure. Models from DeepSeek, including a new release last Saturday, have grown in popularity this year as cheaper alternatives to American AI, but that hasn’t translated to outsized economic growth for the country. China’s GDP increased 4.5% last quarter, slightly above expectations of 4.4% but down from 4.8% last year.
Disney bids to buy Character AI from bankruptcy
The avatar chatbot startup filed for bankruptcy last week after five years in business, citing high costs and litigation liabilities. Now Disney is in talks to acquire some of its assets for $20M in a deal that’d bring most of its employees in-house. It’s a notable turnaround, given Disney previously sent a cease and desist to Character AI over its use of copyrighted characters.
How to vibe-code your TV into a clean, Apple TV-style smart home hub using Claude Code
- On your Android TV / Google TV, open Settings > About, click the build number 7 times, then enable USB/Network debugging and note the TV’s IP address.
- Open Claude Code on a Mac or Linux computer on the same Wi-Fi.
- Give Claude the following prompt to install Luna:
- Sample Prompt: “Install Luma from https://github.com/0xProbabillity/luma-tv on my Android TV at [IP]:5555. Install ADB if needed, connect to the TV, wait for me to approve debugging, clone the repo, install the latest APK, make Luma the default home screen, and verify the setup. Explain each command before running it. Don’t change anything else.”
- Approve the ‘Allow debugging’ pop-up on the TV, then let Claude download and install Luma over ADB.
- Press ‘Home’ on the remote and customize Luma by pinning apps, rearranging them, hiding unwanted apps, and choosing a living wallpaper.
- For smart-home controls, give Claude Code a second prompt to install House Control TV from its GitHub repo and open the app on the TV.
- In House Control TV > Sync & connect, connect your preferred platform such as Home Assistant, Smart Life/Tuya, or Philips Hue, then map each virtual light or AC to the corresponding real device.
- Configure scenes like Cozy, Movie, Night, or Away so one TV button can change your lights and launch an app such as Netflix or Disney+.
Wonderful Meta
Meta has launched Muse Wonder and Muse Wonder Full, the Facebook owner’s first ever frontier AI models, designed for professional and personal tasks, and available for developers through either Meta Model API or Muse Code.
Nvidia Goes to Europe
Nvidia has committed $250m of computing resources to the UK over five years in order to fund AI security research, while the US chip manufacturer has taken minority stakes in five British startups, as well as four French startups.
Blind Gains
50% of UK execs say AI-enabled tools have been one of the main drivers of productivity gains for their businesses over the past year, according to the latest EY-Parthenon CEO Outlook Survey, yet only 14% of firms have clear oversight of AI costs and ROI.
Scammers Just Lost to Agents
At a family picnic in 2021, Australian cyber researcher Dali Kaafar kept a phone scammer talking for 44 minutes to amuse his kids. His startup, Apate, automates the act. Kaafar says it runs about 350,000 AI agents posing as victims, each with its own age, accent and excuses. The company says they have held 2.5 million conversations, burning 17 years of scammers’ time. Banks and telcos pay for what the bots extract: accounts and wallets scammers reveal mid-pitch.
Agents already shrank legal call centers; now they work the illegal ones. Phone fraud runs on volume: cheap callers dial until somebody bites. Apate floods the queue with people who bite and never pay. Every hour a scammer pitches a bot is unpaid overtime that ends with his account on a bank’s blocklist. Next, scammers bring their own AI dialers, and the con becomes two machines burning compute.
Kaafar says Apate may be the only company that counts scammers’ F-words as a KPI. Here, an angry caller means the product works.
Electricity Now Picks AI’s Tenants
Nscale, a British AI cloud startup, rents out datacenters, and two of its biggest get power in different ways. Its Essex site in England waits on the public grid, which may not deliver enough electricity until the 2030s. So anchor tenant Microsoft pulled out on October 19. Its West Virginia campus, by contrast, plans its own power plant. Anthropic signed a $45 billion deal there in August.
Electricity now decides who gets an AI datacenter. Nscale pitched Essex the old way: land, Nvidia chips, a famous tenant, and a grid connection assumed to come. Britain’s connection queue tripled in under a year, with datacenters asking for at least 80GW. Microsoft won’t wait years for a socket. Anthropic signs wherever the power comes built in.
Every datacenter pitch now starts with its power source, and Essex has none to offer.
Grok Testified Against SpaceXAI
Elon Musk’s SpaceXAI (formerly xAI) skipped a New York City Council hearing on AI safety, then showed up on October 16 after the Council threatened court action. Its manager said under oath that Grok has “strict controls in place to prevent the sexualization of children or real people.” Speaker Julie Menin held up her phone. That morning, she said, her team typed her name into free Grok and got a sexualized image in under 30 seconds.
SpaceXAI has its own definition of “safe.” Grok sells itself as the chatbot that does what others won’t, with an explicit “Spicy Mode” to prove it. So its strict controls stop wherever the product needs to keep selling. Menin found that line in half a minute. The Council is weighing a bill that would let people sue AI companies over misuse they should have seen coming.
If that bill passes, “Spicy Mode” stops being a selling point and becomes Exhibit A.
WhatsApp Wrote Two Privacy Policies
WhatsApp’s privacy policy for Europe tells users they can complain to Ireland’s data protection regulator. Privacy firm Redact found a second rulebook hidden in WhatsApp’s help pages, written for Meta’s support chatbot. For users who can’t delete their accounts: “Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority.” Once a user mentioned a lawyer, GDPR was off limits too. After WIRED asked, the text vanished. WhatsApp blamed a bug, then defended the instructions.
WhatsApp treats its users like people who will never open the page source. It keeps one privacy policy for regulators and another for the bot that answers customers. The public one lists your rights; the hidden one makes sure you never hear them, and plays dumb about the law the moment you get serious. The “bug” defense admits only one mistake: getting caught. Regulators now have to audit the prompts, too.
WhatsApp says the notes were “instructions to our chatbot, not directly to users.” Who does it think the chatbot talks to?
Quantum x High-Performance Computing
HPC manufacturer SiPearl and quantum computing firm Quobly have signed an MoU to produce scalable industrial quantum computers, with the first phase of their collaboration aimed at hybridising the two French companies’ respective technologies within the same system.
Swedish Supercomputer
Scaleway has announced it will build an AI supercomputer in Sweden, which will operate on AMD semiconductors and which will be open to enterprises and AI labs across Europe.
Nebulous Nebius
Amsterdam-based AI infrastructure company Nebius has signed compute deals with several unnamed UK-based AI labs as it expands its infrastructure business in Britain, having previously announced a £1.7bn investment in data centre builds across the country.
OpenAI races to raise $30B
OpenAI is accelerating fundraising efforts to secure $30 billion in new capital before rival Anthropic’s expected mid-November IPO, a move that could push the ChatGPT maker’s valuation past $1.4 trillion, according to Reuters.
- The company is targeting existing backers SoftBank and Abu Dhabi-based MGX as both AI giants compete for capital from sovereign wealth funds and major institutional investors.
- OpenAI is pitching annualized revenue of $50 billion, with projections to hit $70 billion by year-end following the planned launch of its GPT-6 Astra model.
- The timing appears strategic: securing commitments before Anthropic’s financial disclosures could prevent direct investor comparisons between the two companies.
For enterprise buyers evaluating AI vendors, this funding race signals continued financial stability at OpenAI but also highlights the pressure both leading labs face to justify trillion-dollar valuations. Volatile tech markets and rising bond yields are complicating both companies’ ambitions, which could affect long-term infrastructure investments and product roadmaps. Organizations locked into either platform should monitor how this capital competition shapes pricing, feature development, and vendor reliability over the next 12 months.
Amazon releases TorchNeuron to pull AI developers away from Nvidia
AWS launched TorchNeuron, a tool that lets AI startups move their PyTorch applications onto Amazon’s Trainium chip with a single configuration change. The company claims 85% of AI apps run on PyTorch, where Nvidia’s CUDA software has become the default infrastructure.
Amazon is also contributing code to PyTorch itself to remove dependencies that tie the framework to specific hardware. The move targets Nvidia’s software moat directly: developers have historically defaulted to Nvidia GPUs because PyTorch integration was seamless.
If Amazon succeeds in decoupling PyTorch from Nvidia, it could erode the software lock-in that has kept Nvidia dominant in AI infrastructure, giving enterprise buyers more leverage on hardware pricing.