 |
|
ISSUE #357 | THURSDAY, NOVEMBER 19, 2026
Nvidia’s board has ruled out Jensen Huang’s return. Washington is weighing new disclosures about Chinese AI models. And a business-owner survey suggests the AI conversation is moving beyond experimentation, even if the promised gains have yet to arrive. Today’s edition separates what has happened from what is proposed, then gets into tools you can try and a small service business you could test.
|
Today’s Top Five |
1. Nvidia’s board rules out Jensen Huang’s return as CEO
Nvidia’s board said Nov. 17 it did not intend to reinstate Jensen Huang after he offered to return. Reuters also reported that Chairman Harvey Jones reiterated that position on the earnings call. This is a substantive development beyond the resignation or Huang’s offer: the board is pursuing a successor rather than a comeback. Interim CEO Colette Kress remains in charge. Separately, Nvidia’s own results report $115.8 billion in quarterly revenue, up 103% year over year. The leadership crisis and the continuing demand for AI infrastructure need to be understood together; strong sales do not resolve governance questions. Reuters board report, Reuters earnings-call report, Nvidia results.
|
2. Reuters: U.S. officials are weighing public disclosure of Chinese AI use
Reuters reported Nov. 18, citing three people familiar with the matter, that the Trump administration is considering requiring U.S. companies to publicly disclose their use of Chinese AI models. The reported aim is to discourage reliance without an outright prohibition. This is a policy under consideration, not an enacted requirement, model ban or compliance deadline.
The business question is concrete: do you know which models your software actually uses? A recognizable application name does not necessarily reveal the model underneath it. A useful preparation step is asking vendors which models support each feature, where business data goes, and how model changes are communicated. That is our practical recommendation, not a requirement reported by Reuters. Reuters report.
|
3. Bank of America survey: 84% of polled owners use AI, but expected gains are not proven gains
Bank of America’s Nov. 17 report says 84% of the business owners surveyed use AI in some capacity. Among AI users, 68% expect it to increase revenue over the coming year. Among respondents using or planning to use AI, 60% expect employees to move into new or redesigned roles, while 52% expect to hire additional staff to manage AI tools. Those figures describe different groups of respondents and what they expect, not what AI has delivered.
Ipsos conducted the online survey Aug. 7–20 among 849 small and 250 midsize U.S. business owners. The sample covers businesses with at least two employees and $100,000 in revenue, not a census of solo entrepreneurs. It is self-reported, bank-sponsored research, not evidence that AI caused higher profits or hiring. Report and methodology.
|
4. Citi says third-party AI transfer instructions count as yours, even when they contradict your prompts
Citi's November online-transfer terms say it will treat instructions from a third-party AI agent acting for a customer as authorized, including instructions that exceed or conflict with that customer's prompts. The document cautions customers against giving such agents access to their accounts. This is the bank's stated contract position, not a finding that customers have no legal protections. A useful boundary: an assistant can help organize a payment checklist without being given permission to move money. Read the specific account terms before connecting a financial account to an agent. Citi's current transfer agreement, section 4. November update, checked Nov. 18.
|
5. Google Vids makes personal AI avatars free for personal accounts
Google's Nov. 17 announcement opens personal avatars to personal Google-account users at no cost. The feature builds an avatar of the account holder, not an arbitrary person, and lets the user reuse that likeness in video scenes. Creation begins in the desktop Vids experience. A useful first experiment would be a short explanation of a service or a repeat training instruction, followed by a careful check of the words and presentation. Free access does not establish unlimited usage, and an avatar should not be passed off as a spontaneous live recording. Google's announcement.
|
More AI News |
6. ChatGPT adds live transcription and removes its ten-minute dictation limit
OpenAI's Nov. 18 release notes say dictated speech now appears as text while a user talks, rather than waiting until the recording ends. The old ten-minute recording cap is removed. The rollout covers all plans globally on iOS, Android, web and desktop. This is dictation into ChatGPT, not a claim that every voice conversation or recording product has changed. For writers and busy professionals, the experiment is simple: speak a rough explanation in your own words, check the transcript, then ask for structure without losing the original point. Official release notes.
|
7. Google's Genie 4 lets people change a generated world while exploring it
Google's Nov. 18 Project Genie update adds prompt-based changes during exploration and adaptive audio. Access is rolling out to eligible AI Ultra subscribers; Pro users have a waitlist. This offers a new way to test a setting or visual idea, but it remains an experimental world-generation system, not a finished game engine. Google acknowledges limitations including latency, physics, sound precision and persistence. Try it for creative prototyping: explore a scene before spending time building it, while keeping expectations about consistent mechanics modest. Google's announcement.
|
8. Google's free state AI-training partnerships now cover ten states, including Arkansas
North Carolina became the tenth participating state in Google's Nov. 17 announcement. Arkansas is already among the partners; it was not newly added in this announcement. The programs use institutions such as libraries, colleges and workforce organizations, with local access arrangements. This gives readers a better next step than buying an expensive course because they feel behind: check what their state partnership offers first. Availability depends on the local institution. Google's state-partnership announcement.
|
9. Google's Stitch Loop proposes ongoing app improvements, but it is still waitlisted
Announced Nov. 18, Loop uses connected code and operational evidence to propose improvements, build changes in an isolated environment and submit them for human review. Google describes security, accessibility and speed checks among its uses. Global waitlist, with eligible Ultra users prioritized. The interesting change is maintenance after an app's first version, not another instant-app promise. Google.
|
10. Suno adds music visuals and lyric videos, with separate visual credits
The Nov. 17 release adds image/video generation around songs, including lyric-video controls and extensions. Visual credits are separate from music credits, so a music subscription is not an unlimited video budget. Check source-material and commercial rights before using the results in a campaign. Release notes, credit guidance.
|
11. Gemini is reaching eligible lower-memory Android phones
Google's Nov. 18 update extends app support to eligible Android Go devices, including Live, camera assistance and image generation. It does not mean every old Android handset qualifies. For organizations helping people adopt AI, checking existing-device support may be more useful than recommending new hardware. Google.
|
12. ChatGPT for Sales connects customer research, call notes and sales paperwork
OpenAI's Nov. 17 sales announcement includes a Sales plugin and more than 30 new or enhanced partner plugins. Access depends on the plan, region, connected accounts and workplace settings. Start with a sourced account brief before authorizing record changes or messages. OpenAI.
|
13. AssemblyAI's new transcription models target recordings, live meetings and voice agents
The STT 3.7 family announced this week separates those workloads rather than treating every recording the same. Existing default prerecorded integrations are due to migrate Dec. 8 unless developers pin a model. This is infrastructure inside other tools, not a new free consumer notetaker. Teams should test names, figures and overlapping voices before a migration. AssemblyAI.
|
14. Arcade's new marketing-video pitch includes editable scenes, not just a finished clip
Shown on Product Hunt Nov. 18, Arcade can start from a prompt or screen recording and use a brand kit and voiceover. Its connected-source and publishing features are vendor-described capabilities, not a quality test. A useful evaluation is whether a short product tutorial preserves accurate details after edits. Launch, official product.
|
15. Pitch Avatar's HR Suite turns existing training materials into conversational lessons
Surfaced on Product Hunt Nov. 18, the suite combines avatars, questions, reminders and completion tracking. This is a showcase date, not a verified first-ever release date. Policy questions still need an escalation path to a person; a confident avatar is not the authority on benefits or employment rules. Launch, product.
|
16. Novo introduces conversational tools across small-business finances
Novo's Nov. 17 announcement describes transaction search, spending-pattern review and recommendations alongside banking, payroll, tax and credit services. Its reported association between AI use and customer cash balances does not prove causation. The practical question is whether an answer can be traced to the right transaction, not how polished it sounds. Company release.
|
17. BlackLine pairs an OpenAI collaboration with an AI-controls certification
BlackLine announced planned AI finance work with OpenAI Nov. 17 and an AIUC-1 certification audited by Schellman Nov. 18. Buyers should ask for the certification's actual scope and which workflows are live. Neither partnership nor certification guarantees error-free finance work. Collaboration, certification.
|
18. Microsoft and more than 40 organizations propose shared AI-transparency practices
The Nov. 17 initiative concerns independent evidence about models and agents, including access, actions and permissions. It is standards work, not a completed universal standard. For buyers, ask whether an agent's action history can be checked against independent records rather than trusting its explanation of itself. Microsoft.
|
19. Google's temporary AI Mode threads stay out of account history and personalization
The Nov. 17 feature rolls out globally in supported languages on desktop and mobile browsers. Google says these sessions are detached from the account, do not use past preferences and do not affect future personalization. That is not a blanket promise that confidential workplace information is safe to upload. Google.
|
20. Amazon Q Developer's GitHub app stops taking new installations
AWS documentation says new installations stop Nov. 18 and support ends January 30, 2027. This concerns the GitHub app, not all Amazon Q products. Existing users should identify affected repositories and assess a migration; AWS points to Kiro's GitHub app. AWS documentation.
|
21. PR Newswire survey finds 70% of communicators using generative AI for releases
A Nov. 17 report surveyed 1,016 communications professionals across 30 countries. Fifty-seven percent said they use clear, descriptive headlines to improve AI discovery. That tells us what respondents are trying, not whether the tactic earns citations. The research comes from a distributor with a commercial interest in press releases. Clear headlines are worth using for human readers, too. PR Newswire report.
|
22. Starburst introduces a shared store of business meaning for AI agents
Its Nov. 18 Context Studio announcement addresses the gap between finding data and understanding what a business means by it. A definition such as “active customer” needs to stay consistent across tools. The larger-enterprise product is less immediately useful to solo readers than the lesson: document business definitions before asking AI to compare figures. Official event and product introduction.
|
23. Nvidia says it has stopped shipping licensed Hopper chips to China
Reuters reported Nov. 18 that Kress said the company was no longer shipping these products to China and its outlook excluded them. Nvidia sold $2.4 billion of the licensed Hopper-generation products there in the preceding quarter. This is a distinct commercial and export-policy development, not a claim that every Nvidia product is banned in China. Reuters report.
|
24. Anthropic confirms $83 IPO price and 980-million-share offering
Anthropic has confirmed an IPO price of $83 per share, above its earlier $70–$78 range. The final offering comprises 980 million Class A shares: 956,351,354 sold by the company and 23,648,646 by existing shareholders. That replaces the preliminary 825-million-share offering. The company says shares are expected to begin trading on Nasdaq under ANT on November 19. That is the announced timetable, not confirmation of a first trade or the offering’s closing. Official pricing announcement.
|
25. Microsoft expands its own models for coding, security and images
Microsoft’s Nov. 17 model announcements include MAI Code 2 and Code 2 Flash, MAI Cyber 2 Flash and MAI Image 2.7. Microsoft is expanding the underlying model choices and competing on cost. Pricing and performance comparisons are vendor claims until independently tested. Availability varies by product and plan. Microsoft AI announcements, Ignite announcement.
|
Tool Spotlight |
Bluedot Phone Calls
The Nov. 18 Product Hunt showcase and maker explanation describe dialing from Bluedot web/mobile with a verified number, then receiving a transcript, summary and action items alongside meeting notes. Outbound calls only for now; country coverage is still expanding. Phone-specific quotas/prices were not confirmed. Obtain appropriate recording consent; the presence of an optional announcement does not settle legal requirements. Test with a consenting colleague before using customer conversations. Maker launch and limitations, official site.
|
|
AI Workflows | Thursday: New-opportunity identification
You could help small businesses get their billing paperwork ready.
On November 16, SpaceXSI described five Grok Bots handling parts of its accounting operation. The company says it gave them documented jobs, restricted permissions and human approval points. That is a company case study, with substantial setup behind it.
Here is a smaller opportunity worth testing: a billing-prep service for a few local service businesses. Your customer finishes the work. You help assemble the paperwork needed to bill for it: the agreed scope, relevant purchase order, proof of completion and questions that still need an answer.
The business hypothesis is that some owners lose enough time hunting through email and folders to pay someone to do this. An agent could reduce the gathering time enough to serve a small account profitably. Neither demand nor profit is established yet.
Test the service before you sell a monthly package
Speak with three owners you already know. Ask them to describe the last bill delayed by missing paperwork and how they resolved it. With one owner's permission, test ten completed jobs using copies of their records. Exclude bank credentials and unrelated customer data.
Use a company-approved agent with access to the selected folders. Grok Bot supports connected tools and scheduled routines; access requires an eligible plan and authorized sign-ins. Check that your chosen system can reach the needed records before committing to the service. Keep source systems read-only. You own the checklist and verify every packet; the client approves any invoice or message.
Give the agent this assignment: Review these ten completed jobs using only the approved client folder. Retrieve the agreed scope, purchase order and completion evidence. Create a separate checklist for each job with links to the source documents. Flag missing or conflicting information. Do not infer that work was accepted or that a charge is due. Do not change records, send messages, issue invoices or move money. Stop at a review packet for me. |
Measure your review minutes, errors and useful omissions caught. Then offer the owner a clearly priced, limited paid pilot that covers your review time and tool costs. A compliment is encouraging; a paid commitment tests demand. End the experiment if verification eats the savings, errors persist or owners decline the price.
If you manage a business, try the same service internally for one small account. Keep accounting judgments, tax questions and payment disputes with the people qualified to handle them.
Choose one owner and ask which piece of billing paperwork they had to chase most recently.
|
|
Thanks for reading mAIn Street. Aric Mitchell |