Today's thread
Half of today's news is about the terms of using AI changing. Anthropic revised the usage policy that spells out what you may and may not do with Claude, and some of the perks it gave startups were paused because demand outran supply. Naver was about to attach AI-written ad copy to search ads for real, then moved the date back to undecided.
The other half is about where AI runs. Google released an experimental app that processes meeting audio on a Mac instead of sending it to outside servers, and Microsoft revealed a dedicated chip to run its own AI services. When choosing a tool, look at where your data goes and who pays for the computing, not just at the features.
Anthropic revises usage policy — effective November 12
On October 8, Anthropic announced a revised Claude usage policy that applies from November 12. One new clause prohibits 'sustained and needless abusive or cruel behavior' toward its models. The company says it covers only extreme cases repeated with no discernible purpose, and does not apply to ordinary frustration, pushback, dark creative themes, or model testing and research.
Enforcement relies on Claude's ability to end a conversation itself. The announcement calls this 'the primary enforcement mechanism' and says it applies in the Claude web and apps and in Claude Code (Anthropic's coding tool for developers). Account suspension is not mentioned in the announcement.
The change that touches day-to-day work more directly is the high-risk use clause. Uses that affect health, legal rights, finances, livelihood, or essential services require a qualified person to review the output and notice to the people affected. Teams that use Claude for judgments about people, such as hiring, lending, or handling public complaints, should check before the effective date that a reviewer and a disclosure notice are in place.
No concrete threshold for 'sustained and needless' behavior has been published, and the announcement alone does not say what happens to the accounts of repeat offenders. The revision also explicitly bans tracking people without consent and using Claude to decide or recommend whom to investigate or arrest.
Claude startup perks — Team and $1,000 credit on hold
The Claude startup program page now says it is 'over capacity' on the Claude Team plan offer and the $1,000 API credit. API credit is a prepaid amount deducted from usage fees when you connect Claude to your own service or business software. The stated reason is very high application volume over the past few days.
All applications that have not yet received the offer will be re-reviewed, which may change the status of some of them. Anyone who has already claimed the offer keeps it. The partner discount and credit bundle (worth up to $45,000), biweekly 45-minute office hours, and events remain, and companies backed by a partner VC can still get up to $100,000 in additional credits through that VC.
Early-stage teams that built a prototype schedule around this credit should redo the math. Even with an approval notice, treat the credit as unconfirmed until it actually lands in the account, and set aside the first month's usage fees in the budget.
No date for resuming the offers and no re-review criteria have been published.
Google AI Edge Foresight — meeting notes that stay on your Mac
On October 6, Google introduced AI Edge Foresight, an experimental Mac app, on its developer blog. It fills out the short notes you jot down with details from the live conversation, picks up questions raised during the meeting and suggests answers, and lets you search images, documents, transcripts, and notes in plain language. It takes in both system audio and the microphone, so it works with any video-meeting service, and it runs without an internet connection.
Most AI meeting-note tools send recordings to a company server for processing; Foresight does all of it on your own computer. Inside are Google's open model Gemma 4 and EmbeddingGemma 2, a search model released the same day. EmbeddingGemma 2 compares text, images, video frames, and audio on one common scale so you can find them, and it has 740 million parameters (a number that roughly indicates model size).
It is worth a look for people in government, finance, or healthcare who find it hard to upload recordings to outside servers. But 'processed on my device' does not automatically mean it passes an organization's security rules, so getting sign-off from the information security team before using it in work meetings is still the right order.
Google calls it an experimental app, and the official post does not state supported languages (including whether Korean is covered), the minimum macOS version, or memory requirements. It also says nothing about operating systems other than macOS.
Naver Power Link AI summary text — launch date unset again
Naver moved the official launch of AI summary text for its Power Link search ads from the 13th to the 20th and then back to an unset date, according to an October 9 report. Additional testing that had been announced for October 8–15 is also unscheduled. AI summary text is a feature in which AI reads an advertiser's registered ad copy and landing page, then writes a summary matched to the search intent and shows it alongside the ad.
Naver has been pitching AI as the engine of its ad revenue. It said second-quarter ad revenue rose 7.5% year on year, with AI behind more than 60% of the increase. It also said AI Briefing ads, launched at the end of July and placed in the AI summary area of search results, have a click conversion rate more than 30% higher than conventional search ads and a purchase conversion rate more than three times higher. A conversion rate is the share of people who saw an ad and went on to click or buy.
Advertisers cannot edit the AI-written copy, and Naver says it does not guarantee the copy's accuracy or appropriateness. Instead, an automatic check filters out false or inappropriate expressions, and advertisers can switch the feature off per ad group (the unit used to manage a set of ads). If your industry is sensitive about price or efficacy claims, decide in advance which groups you will turn it on or off for as soon as the date is announced.
Hospitals and clinics, pharmaceuticals, medical devices, health supplements, and finance and insurance are excluded from the start. No performance figures for Power Link AI summary text itself have been released, and Naver says it will announce the new date once it is set.
Microsoft Maia 300 — in-house AI chip with 288GB of HBM4
On October 12, Microsoft detailed the architecture of its in-house AI accelerator, Maia 300, on the Azure blog. An accelerator is a chip dedicated to AI computation. Each chip carries 288GB of HBM4 high-bandwidth memory with 18TB per second of memory bandwidth (the speed at which data is fed to the chip), and the compute dies are built on TSMC's N2P process.
Compared with the previous Maia 200, memory bandwidth is 2.5 times higher and capacity is up 33%. Microsoft says a rack (one server cabinet) holding 96 of the chips delivers up to 7 times the token throughput of Maia 200, tokens being the units in which AI reads and writes text, and more than 2.5 times the performance per watt. It is designed to power Microsoft Copilot, GitHub Copilot, and Microsoft Foundry, and first production deployment has been reported for 2027.
You will never buy this chip yourself, but if a cloud company can do the same work on its own chips with less electricity than on Nvidia GPUs, that affects the price and usage limits of services like Copilot. If you plan internal AI tool costs over several years, which chips your vendor runs on is now worth watching.
All figures come from Microsoft itself and have not been independently verified. Whether Azure customers can rent it directly and at what price has not been disclosed; more details are promised for Microsoft Ignite 2026 on November 17–20. No schedule for Korea has been announced.
Things to try today
List the tasks where you use Claude for judgments about people, such as hiring, lending, complaint handling, or medical guidance. For each one, check before November 12 that someone reviews the output and that a notice tells people AI is involved.
If your team applied for the Claude startup perks, check in the Claude Console whether the credit has actually arrived. If it has not, budget next month's usage fees separately.
If you use a Mac and cannot send meeting recordings outside, get your security team's approval and then try AI Edge Foresight on a test meeting with no personal information. You will need to check for yourself how well it handles meetings in Korean.
If you advertise on Naver Power Link, decide in a table now which ad groups will have AI summary text turned on or off. The rule of thumb: turn it on only for groups where copy you cannot edit would be acceptable.