Will AI Companies Get Served? MS AI in School; 'Starlight Feeds My Core'

Today's AI Outlook: 🌩️

The Server Farm Meets the Town Hall

AI’s construction boom faces a local vote of confidence in Tuesday’s U.S. midterms. Roughly 20 communities are voting on data center bans or pauses, according to today’s reporting, as residents press for more control over projects in their backyards. OpenAI is reportedly preparing a post-election campaign around community benefits and a new name, “AI compute campuses,” while the industry faces the possibility of tougher scrutiny in Washington.

For communities, the debate is about the terms of hosting large computing facilities: who supplies the power, what benefits residents receive and how much say they have before construction begins. Report note that the White House has secured voluntary commitments for tech companies to supply their own power, while President Donald Trump has floated payments to local residents. Those proposals have different levels of commitment; a suggested dividend is still a proposal.

Why It Matters

Local opposition can delay or block the physical infrastructure behind AI services. Developers will need credible community agreements as they seek room to expand, and Tuesday’s results could change the political pressure they face.

The Deets

• Roughly 20 communities are considering bans or pauses, according to Politico’s reporting.
• Some Democratic candidates support a federal freeze on new data centers. That is a campaign position, not an enacted nationwide restriction.
• OpenAI reportedly plans to promote local economic benefits and a stronger community voice in new projects.
• The Neuron, citing The Information, reports that candidates favoring data center pauses could win up to 15 of 36 governor races.

Key Takeaway

The AI buildout depends on winning permission in places where the servers will actually live. A fresh name gives the sales pitch a polish; residents still have terms to negotiate.

🧩 Jargon Buster - Compute: The processing power computers use to train AI models and generate their responses.

🏛️ Power Plays

Microsoft Puts School Privacy in Writing

Microsoft’s AI Safety & Privacy Standard became available to U.S. school districts on Nov. 1, giving them a way to put student-data protections into new or existing Microsoft contracts without waiting for renewal. Announced with the American Federation of Teachers and United Federation of Teachers in September, the agreement makes those commitments enforceable once a district incorporates them. Schools still choose whether to use AI and must take the step of adding the standard.

The Neuron reports that districts adopting the standard can cancel and seek damages if Microsoft violates its commitments, with certification and audit requirements supporting enforcement. That gives school leaders a concrete way to hold the company to its promises. The agreement also leaves the classroom debate open: UFT President Michael Mulgrew has said AI use in grades K-8 is not appropriate right now. Setting conditions for a tool does not require a school to deploy it.

Why It Matters

Schools hand technology providers sensitive information about children and educators. Contractual protections give districts a way to make those boundaries enforceable, even as policymakers debate how existing privacy laws apply to generative AI.

The Deets

• Training is restricted: Student and educator data cannot train or improve models, except for a narrow safety and security exception. The training and data-sale bans are permanent.
• No data sales or advertising use: Schools retain control over retention, export and deletion.
• People oversee high-risk decisions: Human review is required, and AI companions are prohibited.
• Tracking is constrained: Microsoft’s fact sheet says tracking, profiling and related features require limits and review. That is more qualified than a blanket tracking ban.
• Accountability has a clock: Breaches must be reported within 72 hours; violations must be fixed or AI use paused or ended.
• Coverage still needs attention: The fact sheet does not list individual covered products. The Neuron reports interest from OpenAI and Anthropic, but no equivalent signed commitment from either.

Key Takeaway

Districts can now put stronger conditions on Microsoft’s school AI services. The next test is adoption by schools and whether other providers accept the same terms.

🧩 Jargon Buster - Model Training: The process of using data to change what an AI model learns, potentially shaping how it answers future users.

🔬 Research & Models

Google’s Orbital Computer a Poetic Success

Google says its Project Suncatcher satellite has answered an AI prompt from orbit, running the Gemma 4 model to compose a haiku and sending it back during a pass over Svalbard, Norway.

Blue marble below
Data streams in silent grace
Starlight feeds my core.

The experiment tests a building block of Google’s ambition to run AI computing in space. For now, the satellite’s power limits allow only one of its four AI chips to operate at a time, in bursts of about 20 minutes, according to today’s reporting.

Google’s announcement describes the response as an early milestone following the prototype’s Oct. 1 launch. Its next work includes testing the chips under radiation and thermal stress. A successful poem demonstrates that the system can execute a model and return an answer; reliable computing at meaningful scale still needs proving.

Why It Matters

An orbital computing system would need dependable power, durable hardware and fast connections between spacecraft. This test supplies an initial working example, while the reported operating limits show how much engineering remains.

The Deets

• The model was Gemma 4, running aboard the satellite.
• The response reached the ground during a pass over Svalbard, Norway.
• The newsletter reports that power constraints permit one of four Google AI chips to run at a time for roughly 20 minutes.
• Google’s next milestone is a pair of interconnected satellites in 2027; the newsletter says they will test laser links for sharing computing work.

Key Takeaway

Google has an AI response from orbit. The harder assignment is keeping the hardware useful, connected and reliable beyond a short demonstration.

🧩 Jargon Buster - TPU: A tensor processing unit, Google’s specialized chip for the calculations used by AI models.

⚡ Quick Hits

• Grok text gets an invisible marker. Developer documentation says watermarking starts Nov. 7 for supported text models, helping identify generated output through statistical detection. It is on by default, with model-level opt-outs subject to the documented conditions.
• Claude’s top plans get new names. Max 5x becomes Max and Max 20x becomes Ultra at unchanged prices, with the five-hour usage limit removed for verified Ultra accounts. The linked announcement on X could not be independently retrieved.
• ChatGPT and Codex will share a budget. OpenAI says Plus, Pro and Business plans will share a token-based usage allowance across both products beginning in December; Instant chat will not count against it. More demanding models and tasks will consume more of the allowance.
• Apple sets its China AI rollout. Apple says supported iPhones in China will get Apple Intelligence through iOS 27.3 in December, with a developer beta starting Nov. 4. Alibaba’s Qwen will support selected experiences, alongside Apple’s own models; the release excludes the next-generation Siri AI announced at WWDC 2026.
• Aramco buys into HUMAIN. The Neuron reports that Saudi Aramco took a 49% stake in the Saudi AI company, with about $2.6B in cash going in this year. Its linked report could not be independently retrieved.
• AWS grows fast, and Amazon spends faster. The linked earnings coverage reports AWS third-quarter sales of $46B, up 39%, while Amazon’s companywide quarterly property and equipment purchases reached $77.3B. Amazon’s trailing 12-month free cash flow was negative $42.4B, a different reporting period and broader business measure than AWS sales.

🛠️ Tools of the Day

• Arrow 2 and Inkscape: An editable-graphics workflow pairs Arrow 2’s generated artwork with Inkscape for changing headings and colors afterward. Check that your QuiverAI plan includes SVG export, and keep changing text in separate boxes.
• Phonic-1: The maker says its voice model supports more than 50 languages and can switch languages within a conversation while retaining the same voice. That makes it relevant to multilingual voice-agent projects; the performance claims remain the company’s.
• DeepSeek Super-Agent: This coding-agent prototype handles work for a GitHub repository, can listen to Feishu conversations and can return results through pull requests. It acts when asked or when an operator-defined rule permits it. Its README explicitly calls it research code, not an official DeepSeek product or production-ready software.
• Riverside Agents: Riverside’s production agents help prepare recording sessions, edit episodes, cut social clips and draft promotional posts. You review and approve material before it goes out.
• Ideogram Precise Edit: The Neuron describes Ideogram’s editing tool as changing a selected area while preserving the rest of the image. A product-label swap is one practical use; that preservation claim is reported here, not independently tested.

Today’s Sources: The Internet, The Rundown AI, The Neuron