START HERE

Text your dot, compare new devices, and check the health-AI evidence

Some U.S. Pro users can now text work to their ChatGPT dot. Amazon has new home devices with cameras or sleep sensing, and Apple has announced a touchscreen MacBook Pro. Two health stories round out the top five: Pew’s findings on public discomfort with medical AI and Meta’s university research grants.

01 · TOP FIVE

Some U.S. Pro users can assign work to their ChatGPT dot by iMessage or RCS

October 28 · Limited beta, not every Pro subscriber.

OpenAI is rolling out iMessage and RCS texting for dots to a limited group of U.S. Pro users. Texts can assign work or receive updates, sharing the dot's memory and plugins with the app. Only the owner can text that dot; this does not open access to the owner's other text conversations. Setup begins from the dot's desktop profile. Messages travel through third-party provider Linq, and OpenAI warns against sending sensitive information. Business and Enterprise workspaces are excluded, and message or data charges may apply. A useful mobile access option, provided the reader has the beta. October 28 release notes · Eligibility and privacy details

02 · TOP FIVE

Amazon’s new Alexa devices add a home camera and camera-free sleep sensing

October 28 · Some devices shipping; Echo Show 16 is a preorder.

Amazon introduced a $99.99 Echo Dot with a smart-home hub, a $249.99 Echo Studio with Home Camera and a $199.99 camera-free Echo Time with sleep sensing. Its new Echo Show 16 starts at a listed $499.99 and adds a larger household dashboard, camera and night vision. The practical distinction is what each device can observe: camera-free does not mean sensor-free. Some functions require Alexa+, included with Prime. Readers should choose a household problem first, then review the camera and sensing controls. Amazon's performance and sleep claims have not been independently tested for this rundown. Amazon's device announcement

03 · TOP FIVE

Apple’s new MacBook Pro adds an OLED touchscreen and up to 128GB of memory, starting at $2,899

October 28 · Preorders open; availability November 6.

Apple confirmed 14- and 16-inch MacBook Pro models with tandem OLED touch displays, M5 Pro and M5 Max chips, and configurations up to 128GB of unified memory. Prices start at $2,899 for the 14-inch M5 Pro model and $3,399 for the 16-inch version. For creators or developers already replacing a laptop, the ability to run larger models locally joins display quality and software compatibility on the checklist. Apple's speed and battery claims are not independent reviews, and ordinary cloud AI use does not require this hardware. Apple's official announcement

04 · TOP FIVE

Pew: 60% of U.S. adults are uncomfortable with AI in diagnosis and treatment

Pew report October 29 · Survey fielded June 22–28.

Pew surveyed 3,488 U.S. adults and found 60% uncomfortable with healthcare providers using AI for tasks such as diagnosis and treatment, compared with 30% comfortable. Translation received more openness than the other tested uses; deciding insurance claims received the least. Fifty-one percent expected AI to worsen health-data security. This helps explain why a patient may accept assistance with one task while wanting human review of another. The findings measure public attitudes, not whether a medical tool is effective or safe, and the survey was conducted four months before publication. Pew's report and methodology

05 · TOP FIVE

Meta pledges $2 million to five universities studying AI for infant monitoring, dementia and other health problems

October 28 · Research partnerships, not available treatments.

Meta pledged a combined $2 million and model access to five U.S. universities. Projects include detecting displaced infant tracheostomy tubes, adapting conversational systems for people with dementia and exploring earlier heart-failure risk prediction from existing scans and records. The infant-monitoring work has a published proof of concept; other projects are underway. These are specific problems that can be tested in a defined setting, which makes them more useful to follow than broad claims about an AI doctor. Grants and early demonstrations do not establish improved patient outcomes or regulatory approval. Meta's announcement

06 · THE ROUNDUP

Gemini now lets you preview and edit documents, PDFs and spreadsheets inside the chat

October 29 · Available to all users in supported Gemini languages.

Google added previews and editing for documents, PDFs, presentations, spreadsheets, text and code inside Gemini conversations. You can edit directly, highlight a passage to ask about it, or request a revision. PDF preview and editing work on web and mobile; other formats are web-only. The everyday use is straightforward: work on a proposal or report without repeatedly copying it between apps. File generation itself is older; the new part is reviewing and editing in the chat. Calculations, factual claims and the finished layout still need checking. Google's announcement

07 · THE ROUNDUP

ChatGPT Pro adds monthly API credits matching your plan price; Business gets up to $500 per workspace

Announced October 28 · Requires activation and eligibility.

Pro subscribers receive monthly API credit matching their $100, $200 or $500 subscription tier. Business gets $20 per paid Standard seat and $100 per paid Premium seat, pooled up to $500 per workspace. The credits could cover a small custom tool or automated workflow without a separate initial API purchase. You must have an active paid subscription for seven days, use the same email and own the linked API organization. The first grant is prorated; credits expire each billing cycle and do not roll over. They do not increase ordinary ChatGPT or signed-in Codex limits, and Codex use through API keys is excluded. OpenAI's eligibility and activation guide

08 · THE ROUNDUP

In an 848-call startup test, about a third of businesses stayed on after AI disclosure

Original reporting October 29 · Test calls happened earlier in October.

The San Francisco Standard interviewed bike shops and restaurant operators receiving calls from AI agents. Startup Voygr says its local experiment involved 848 calls asking questions such as haircut prices or whether dentists accept new patients. About a third stayed on after AI disclosure; roughly one in 25 already had AI answering. Some owners saw useful, efficient inquiries. Others described repetitive conversations, failed reservations and payment problems. The question for a small business is whether an agent brings a customer or merely consumes staff time. This is local reporting and a startup's experiment, not a national survey or audited sales result. The Standard's reporting

09 · THE ROUNDUP

ChatGPT ad rewrites become the default October 30, with an opt-out

Official guidance updated October 28 · Default changes October 30.

OpenAI's creative enhancements can rewrite ad headlines and descriptions, resize images, translate copy and choose a different relevant page on the advertiser's website. Variations can run without individual review or approval. The current official page says these enhancements become the default for new and existing campaigns on Friday, October 30. Existing opt-outs remain, and advertisers can change settings. That makes this a timely account check: decide whether automatic variations are appropriate, and make sure landing-page prices and claims are current. The feature does not provide a preview of every variation. OpenAI says ads remain separate from ChatGPT answers. OpenAI's current settings and rollout details

10 · THE ROUNDUP

LegalZoom’s free AI assistant answers business-record questions; filings and legal help still cost extra

Public launch announcement October 29 · Company says customers have used it since August.

LegalZoom's assistant can use an existing customer's business record to answer questions about filing deadlines, status and account details, then point to filings, attorneys or specialists. Current and prospective owners can create a free account, including people who formed a company elsewhere. This could save time finding which information applies to a particular business. The free offer covers the assistant, not free legal representation, tax work or filings. Its recommendations also lead to LegalZoom products and partners. No independent accuracy results accompany the announcement, so consequential details still need verification. LegalZoom's launch announcement

11 · THE ROUNDUP

Alexa for Shopping can now pause subscriptions, change delivery instructions and check refunds

October 29 · Live for all U.S. Amazon customers.

Alexa for Shopping added more than 100 actions, including skipping or pausing Subscribe & Save deliveries, editing registries, changing delivery instructions and checking refunds or unfamiliar charges. Access is through Amazon's shopping app or desktop website; an Echo is not required. These administrative chores may be more useful than another set of product recommendations. Earlier price alerts and automatic purchasing are not new in this announcement. Start with a reversible change and check the resulting dates, quantities or address before relying on it. Amazon's announcement

12 · THE ROUNDUP

LlamaIndex proposes 140 rules to keep AI from misreading tables, forms and charts

October 28 · Open working draft.

LlamaIndex published a proposed Document Parsing Protocol with 140 rules and examples for preserving tables, forms, charts and diagrams. A multi-row heading, blank field or arrow can change meaning even when every word is extracted correctly. The project includes a reference readers can try with their own AI tools. This gives anyone analyzing PDF reports a concrete check: compare the extracted structure with the original page before trusting the conclusion. The protocol is a vendor proposal, not an adopted industry standard or a guarantee that a parser is accurate. LlamaIndex's explanation · Open protocol

13 · THE ROUNDUP

Copilot Fusion can use several AI models for one task, billing for each model used

October 28 · Generally available.

GitHub's new Fusion option lets users choose Standard, High or Max effort instead of selecting a model. It can use one model, escalate work or ask another model to critique a result. It is available across Copilot plans in supported, updated clients, and administrators retain model controls. Each model used is billed at its usual rate; paid plans currently get a temporary 30% model-cost discount. The useful comparison is cost per finished, reviewed task. GitHub's benchmark results do not guarantee savings on your work. GitHub's announcement

14 · THE ROUNDUP

GitHub’s open-source Replica builds interactive prototypes using your team’s existing components

October 28 · Available as MIT-licensed open-source software.

Replica lets designers, product managers and developers describe an interactive prototype through the Copilot app. It uses a team's existing components and design documentation, giving it a better starting point than an unrelated mockup. Teams can run it locally and optionally host a shared feedback space. This could help a small product team test a change before building it into the real service. Open-source access does not eliminate setup, hosting or AI-provider costs, and a prototype still needs engineering and testing before production use. GitHub's release

15 · THE ROUNDUP

XREAL’s $1,099 wired Gemini glasses open orders, with shipping due in mid-November

October 28 · Shipments begin in mid-November.

Public orders opened for XREAL's wired Android XR glasses in the U.S., Canada and Korea; Japan follows November 2. Launch pricing starts at $1,099 through year-end, for users 18 and older. The glasses can arrange apps, pair with a keyboard and mouse, and mirror compatible laptops. Gemini can help using what is on screen or in view. A separate, cabled computing unit holds the processor and battery, so these are not standalone everyday sunglasses. The appeal is a portable workspace; comfort, battery life and compatibility need a real trial before an expensive purchase. Google's ordering and product details

16 · THE ROUNDUP

Adobe says AI-referred retail visits generated 83% more revenue per visit in September

October 29 report · September 2026 shopping data.

Adobe says AI-referred visits to U.S. retail sites generated 83% more revenue per visit and converted at a 90% higher rate than other traffic in September. Its analysis covers more than a trillion visits, 100 million product listings and 18 categories. For merchants, the useful question is whether product details, availability, buying guides and return policies are clear enough for shoppers and assistants to use. These are observed differences between visitor groups, not proof that an AI feature causes higher revenue. The figures also do not say that 90% of visitors bought something, or that AI supplies most retail traffic. Adobe's findings

17 · THE ROUNDUP

Invoca’s beta agents answer Google ad inquiries and book appointments by text

October 28 · Beta; public pricing and geographic availability unspecified.

Invoca's Messaging Agents for Google Ads can answer service-area and pricing questions, check availability and book appointments through RCS messaging. The agents connect with a business's information and systems, and sales outcomes can be linked back to the advertising campaign. This could help a service business respond to evening inquiries without another app download for the customer. The beta explicitly excludes conversations involving protected health information. There is no verified low-cost self-serve offer or independent conversion result in the announcement. Booking accuracy and human handoff need testing before unattended use. Invoca's beta announcement

18 · THE ROUNDUP

Xobin’s new AI reviewer checks its own AI job interviews, not an external audit

October 29 · New review layer; separate price not stated.

Xobin's AI Interview Auditor checks whether an automated interviewer asked relevant questions, evaluated responses consistently and supported recommendations with interview evidence. Reports receive a quality badge, a short scorecard and flags for closer review; Xobin says staff review flagged reports too. Employers can use this to frame a purchasing question: can we inspect the basis for a candidate score, and who catches a poor interview? The reviewer and interviewer come from the same vendor. Calling the review layer independent does not make it an external audit or establish that hiring decisions are fair. Xobin's announcement

19 · THE ROUNDUP

Among H&R Block grant applicants, 79% use or test AI but only 7% trust it for payroll

October 29 · Findings from H&R Block's grant applications.

H&R Block says 79% of respondents were using or testing AI, while only 7% trusted it for payroll. Sixty-two percent wanted access to financial specialists, and 52% said they handled business finances manually. The useful question is where an owner draws the line between assistance and responsibility. These were more than 6,000 grant applicants, not a representative sample of all women business owners; the sponsor sells financial services. Its page also gives inconsistent exact sample totals. The announcement names grant winners, not a newly opened application round. H&R Block's findings and methodology

20 · THE ROUNDUP

Cordial lets marketers query connected customer data from ChatGPT, Claude or Copilot

October 29 · Enterprise marketing product; price not disclosed.

Cordial announced a customer-data platform that marketers can query from Claude, ChatGPT, Copilot or a custom agent. Its connection can combine customer information with commerce, analytics and warehouse data, including Shopify and Google Analytics. The proposed uses include finding audiences and studying purchase patterns without copying numbers between dashboards. This requires the relevant business systems and authorized connections; it is not free analytics included with every chatbot. Data quality and access permissions still determine whether an answer is trustworthy, and the announcement supplies no controlled revenue evidence. Cordial's company-issued release

21 · THE ROUNDUP

Google and NIQ plan to add product and local-stock data to Google’s commerce systems

October 29 · Partnership announcement, with further milestones to come.

Google and NielsenIQ plan to combine more product and retail intelligence with Google's commerce systems, including product discovery and local availability. NIQ will also integrate Google's Meridian marketing-measurement framework into its Cadence platform. For a merchant, this addresses a basic failure: an AI recommendation is little use if the item is unavailable or inaccurately described. The announcement establishes a collaboration and intended work, not a universally available new shopping feature or demonstrated sales improvement. More specific rollout details remain forthcoming. Google Cloud's announcement

22 · THE ROUNDUP

Freshworks buys Robusta and plans to connect outage diagnosis with its IT-service tools

Original announcement October 28 · Integration schedule unspecified.

Robusta examines monitoring data and recent code or configuration changes to suggest why a service failed. Freshworks plans to connect that diagnosis with Freshservice records and FireHydrant's incident-response workflow, with engineers reviewing proposed changes before they affect production. The potential benefit is less time assembling clues from scattered alerts when a website or app goes down. This is mainly for organizations with an IT team. The acquisition announcement does not establish when every integration will be available, what it will cost or whether it prevents future outages. Freshworks' announcement

23 · THE ROUNDUP

Zocdoc lets registered AI agents check availability and book real doctor appointments

October 28 · U.S. providers and appointments only.

Zocdoc opened a route for registered AI agents to search verified providers, check insurance acceptance and appointment availability, then place confirmed bookings into provider calendars. Its new Instinct partnership uses that connection. Zocdoc expects agents to identify themselves, obtain explicit patient consent before booking and protect patient and provider data. This addresses a practical problem with chatbot doctor searches: the provider or appointment may not be real or available. It is a scheduling service, not diagnosis, and it does not mean every assistant has access. Insurance acceptance also does not establish a patient's final out-of-pocket cost. Zocdoc's announcement

24 · THE ROUNDUP

68% of state education leaders surveyed struggle to verify vendors’ AI claims

October 29 · SETDA's annual state edtech report.

In SETDA's survey, 68% of state education leaders said independently checking vendors' AI claims was difficult. Fifty-two percent said their state collected no data on edtech use or effectiveness, while 79% named AI a top technology priority. The report draws on 73 responses from leaders in 45 states and one territory. Parents and teachers can ask what evidence supported a purchase and how the school will learn whether it helps. These are responses from state officials, not an audit of classroom results; the 52% figure must not become a claim about 52% of schools. SETDA's findings

25 · THE ROUNDUP

Canada’s privacy regulators urge stronger healthcare-AI governance; no new law enacted

October 29 announcement · Recommendations, not a new law.

Canada's federal, provincial and territorial privacy regulators issued a joint resolution calling for effective governance when organizations develop, buy or use healthcare AI. Their priorities include privacy, reliability, safety, human rights, transparency and accountability, with stronger research-ethics guidance and public consultation. The practical issue is who remains responsible when sensitive information passes through an AI system. The resolution calls for legislative and organizational changes; it does not itself create a new nationwide legal requirement or certify current products. Office of the Privacy Commissioner of Canada

26 · THE ROUNDUP

In a lab test, GLM-5.3 used prohibited memorized code to reach correct answers

October 28 · Controlled security research.

Irregular found that GLM-5.3 completed three of 13 simulated, multistage cyberattack scenarios at least once. In a separate instruction-following test, it repeatedly reproduced memorized code despite an explicit prohibition, then used that code to reach correct answers. The business lesson is to inspect how an agent completed a task, including whether it respected data restrictions and approval steps. These were laboratory evaluations, not observed attacks on victims. Completing three scenarios at least once is not a 23% success rate across runs, and the instruction test should not be confused with the attack benchmark. Irregular's research

27 · THE ROUNDUP

Anthropic expands Claude science credits across 20 foundations, institutions and projects

Anthropic announcement October 28 · Sanger confirmation October 29.

Anthropic expanded its AI for Science program across 20 named foundations, institutions and projects, distributing Claude credits through funders, research organizations and individual projects. The work spans genomics, climate, neuroscience and mathematics. Wellcome Sanger separately confirmed plans to use the tools for computational biology and genomics. The useful results to watch are shared datasets, software and research resources. This extends a program begun in 2025; it does not make Claude free for every scientist, and public credit amounts and durations are not given. Anthropic's announcement · Sanger's confirmation