1Highlight: A smartwatch AI spotted signs of high blood pressure about as well as a first clinic check.
2Japan's freee says its own small AI model reads business documents at 1/50 the cost of a big one.
3A judge says Claude Max's '20x more usage' ads may mislead; GitHub rebuilt storage for AI agents.
Highlight of the day
Health research
65.8%
of high blood pressure cases the watch caught (first clinic check: 55.3%)
Merged from 3 sources
★ Highlight of the day
Smartwatch AI spots signs of high blood pressure without an arm cuff
Google Research says an AI system can spot signs of high blood pressure from an ordinary smartwatch's pulse and motion sensors, with no arm cuff needed to set it up. The peer-reviewed study was published in Nature on 5 Nov 2026. It flags who should get a proper check; it does not give you a blood pressure number.
What we learned
How it works: the watch's light sensor tracks blood flow at the wrist, and its motion sensor picks up tiny jolts from each heartbeat. The AI averages these over several days and adds your height to estimate the chance you have high blood pressure.
The test: 196 adults with no known heart disease and no blood pressure medicine wore a Google smartwatch for seven days. Results were checked against 24-hour blood pressure monitoring, the method doctors treat as the gold standard.
Results (from the Nature paper): the watch caught 65.8% of people who had high blood pressure, versus 55.3% for a first clinic reading. Both correctly cleared 90% of people who did not. Nature's press release says the gap was Statistically significant When a difference in a study is big enough that it is unlikely to be just chance; if it is not significant, the two results may really be about the same., so read it as 'about as good', not 'better'.
Limits: the watch still missed about a third of cases. Google calls this early research, not a current or planned product feature, though it says the work helped inform blood pressure trends in the Google Health app's Health Guardian feature.
Why it matters: More than half of adults with high blood pressure do not know they have it. If a watch ever nudges you, confirm with a proper cuff reading and a doctor, and remember a quiet watch is not an all-clear.
Detecting signs of hypertension using wearable sensors · Google ResearchRead ↗
Opportunistically Detecting Signs of Hypertension on a Consumer Smartwatch · NatureRead ↗
Technology: Smartwatches that can clock high blood pressure · Nature Portfolio press releaseRead ↗
Was this useful?Tap to vote. Hold, right-click or press Shift+Enter to add a reason or a note first.
More stories today
How companies use AI
1/50
of the cost freee reports versus a big cloud AI model
How companies use AI
Japan's freee says its own small AI model cut costs to 1/50
Japanese accounting software company freee reads about one million receipts, invoices and other business documents a month with AI. Instead of sending them all to a big general AI model, it built its own Small language model A smaller, cheaper AI text model, often trained for one narrow job, that a company can run on its own servers instead of paying a big AI service per request. for one job: picking out details like the date, amount and supplier. The figures are freee's own, from a talk at AWS Summit Japan 2026 that ITmedia reported on 7 Nov 2026.
Key learnings
Company: freee, Japan, Cloud computing Renting computers in someone else's data centre over the internet instead of buying your own. accounting and back-office software. Problem: big AI models read documents well, but at a million documents a month, paying per request cost too much.
What they built: an 8-billion-Parameters The internal numbers a model learns during training; more parameters usually means a bigger, more capable but costlier model.Open weights When a company publishes the trained model file so anyone can download and run it on their own hardware. given extra training on one task (Fine-tuning Giving an existing AI model extra training on your own examples so it gets better at one specific task.) with about 10,000 checked examples, plus tens of thousands more for testing. It runs on freee's own servers in AWS. A separate step still turns each image into text first.
Measured results (company's own figures): up to 10 points more accurate on text details such as description and supplier, 5 times faster, and 1/50 of the cost, compared with a Claude Sonnet 4.6-class model used through a cloud service. freee says the saving comes mainly from paying for server time instead of per request.
Lesson for smaller teams: this pays off only with steady, high volume and one narrow task. freee's rule of thumb is that if a small model like Claude Haiku 4.5 already does the job reliably, a self-run small model may compete. Measure real speed early, and keep testing after launch.
Why it matters: AI bills grow with every request. For one repetitive job at big volume, a small model you run yourself can be far cheaper, but someone has to test and maintain it.
freeeは「AIコスト問題」をどう乗り越えた？ 月100万件規模の処理で、汎用LLM比50分の1に (How freee beat its AI cost problem; translated) · ITmedia EnterpriseRead ↗
Was this useful?Tap to vote. Hold, right-click or press Shift+Enter to add a reason or a note first.
Your AI subscription
6x
Max 20x weekly use vs Pro, by the subscribers' maths for one model
Merged from 3 sources
Your AI subscription
Judge says Claude Max's '20x more usage' ads may mislead buyers
A US federal judge said on 6 Nov 2026 that Anthropic's promise of "5x or 20x more usage than Pro" for its Claude Max plans is "plausibly misleading". Key parts of a Class action A lawsuit where a few people sue on behalf of a larger group with the same complaint; a court must approve the group before it counts. by two subscribers can go ahead. Nothing has been proven yet: at this early stage, the court treats the complaint's claims as true.
Key learnings
The plans at the time: Pro cost about RM69 to RM81 (US$17 to US$20) a month, Max 5x about RM406 (US$100) and Max 20x about RM811 (US$200).
The dispute: Max gives 5 or 20 times Pro's use inside each five-hour session, but the Weekly limit A cap on how much you can use an AI tool over a whole week; when it runs out you wait for the reset, even if your shorter session allowance would allow more. is not 5 or 20 times higher. The subscribers calculate that for one Claude model, Max 5x gets about 3.5 times Pro's weekly use and Max 20x about 6 times.
The ruling: Magistrate Judge Alex Tse kept one subscriber's consumer-protection claim, plus claims of negligent misrepresentation and breach of contract. He dismissed the others but let the subscribers refile them. No class has been approved and there is no settlement.
Anthropic had asked for the whole case to be thrown out, arguing it promised more use per session and said weekly limits apply. An amended complaint is due by 2 Dec 2026, according to the case tracker LimitHit.
Why it matters: "More usage" can mean per session, per week or per month. Before you upgrade any AI plan, check the weekly limit, not just the headline multiplier.
Anthropic must face claims its Claude Max pitch misled users · Courthouse News ServiceRead ↗
Order on motion to dismiss, Case 3:26-cv-05763-AGT · US District Court, N.D. CaliforniaRead ↗
Claude usage limit lawsuits (2026): what's alleged and where they stand · LimitHitRead ↗
Was this useful?Tap to vote. Hold, right-click or press Shift+Enter to add a reason or a note first.
AI at work
7,200
updates a second to one project in GitHub's own test
AI at work
GitHub rebuilt its storage so 5,000 AI agents can share one project
GitHub says it rebuilt how it stores code so thousands of AI coding AI agent An AI program that can take actions for you, like writing code, filling forms or moving items, not just answering questions. can save work to the same project at once. In its own tests, 5,000 agents sent 7,200 updates a second to one project while reading stayed fast. These are GitHub's internal test figures, from its engineering blog on 5 Nov 2026.
Key learnings
Why: AI agents save changes far more often than people do. GitHub says a busy project used to mean dozens of updates a second, so it aimed for at least 100 times more.
The idea: most of a code update can be stored without waiting for anything else. Only the tiny step that records which version is now the latest needs everyone to agree, so GitHub gave that step its own fast service.
For people using GitHub, nothing changes: you save and fetch code the same way, with the same protections. Behind the scenes, nearly all projects now sit in Microsoft's Azure Cloud computing Renting computers in someone else's data centre over the internet instead of buying your own. storage.
Limits: the 5,000-agent figure comes from GitHub's own testing, not an independent check, and the post does not give a date for when every project moves to the new system.
Why it matters: Big tools are being rebuilt for many AI agents working on the same files. If your team tries coding agents, plan time for people to check what the agents changed.
Health research
