In partnership with
The voice-to-text app that built its own speech model
Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.
It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.
Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.
Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.
November 2026 · Week 2
November 9, 2026
Good morning!
Cheaper video, faster attackers and stronger hiring share one lesson this week: a headline number is only the beginning. Here is what changed, what the evidence actually says and where our attention belongs next.
AI AND CREATIVE WORK
A cheaper clip is not the same as a finished video
Original AI generated concept illustration, not a Gemini output or product demonstration.
The price of experimenting with AI video is falling again. On November 5, Google announced general availability of Gemini Omni 1.2 Flash on Google Cloud. Its advertised 720p output rate is seven cents per second, 30 percent below the preceding version. Clips can run up to 15 seconds, compared with ten previously.
That puts a single 15 second output at $1.05 at the stated rate. It is arithmetic, not a quote for a finished advertisement: discarded attempts, other service charges and the surrounding production work can change the bill. A consumer subscription is also not the same purchase as an API request.
Google says the model better follows visual references and supports conversational changes to an existing clip. Those are company claims, not independent evidence that every product detail or scene will survive an edit. Audio references require allowlisting, another reason to check access before planning around a demonstration.
For a small creative business, our interpretation is that the opportunity is more inexpensive exploration, not an automatic replacement for production. More candidate shots could help a team settle on a direction before committing to a larger shoot. They can also create a larger pile of almost right material that someone still has to evaluate.
The useful comparison is the cost of an approved result. Include selection, corrections, factual accuracy, rights clearance and the final edit. A cheap scene that misrepresents the thing being sold is not a bargain. Falling generation prices make judgment more valuable, because generating another option is becoming easier than deciding which one deserves to exist.
SECURITY
Exposed credentials now have faster company
Original explanatory graphic. Access depends on the compromised credential's permissions.
Microsoft reported on November 5 that attackers were using agentic tooling to search public GitHub repositories for exposed credentials and then act on what they found. In one observed case, the company says the path from scanning to extracting data from Azure Key Vault took five minutes.
The finding is narrower than a claim that AI can break into any business in five minutes. These attacks used usable credentials. Their consequences depended on what those credentials could access. Microsoft's report describes observed activity, not a universal benchmark for autonomous hacking.
Security provider Todyl has also documented application credential abuse, with a November 6 update connecting its investigation to Microsoft's findings. The reports reference one another, so repeating them does not create two independent confirmations of every technical detail.
The immediate lesson is familiar but more urgent: removing a leaked key from a file does not invalidate it. Microsoft's guidance includes rotating exposed credentials and reducing application permissions. If a contractor manages your website or cloud accounts, ask who owns those credentials, how exposure is detected and how access is revoked when the work ends.
THE ECONOMY
Hiring strengthened, but one month is not the whole market
Original chart from BLS seasonally adjusted payroll figures. The comparison is monthly growth, not total employment.
The November 6 US jobs report showed nonfarm payrolls rising by 165,000 in October, compared with an average monthly gain of 48,000 over the preceding twelve months. The unemployment rate was 4.1 percent, which the Bureau of Labor Statistics described as little changed.
Professional and business services added 70,000 jobs, with gains also in health care and transportation. That is relevant to people selling professional services, but it is not a direct measure of freelance bookings or a guarantee that every industry has more work available.
There was a caution inside the same release: the number working part time for economic reasons increased by 339,000 to 4.8 million. Those people wanted full time work but faced reduced hours or could not find it. The headline and that underlying pressure can both be true.
Our reading is encouraging, not conclusive. These estimates can be revised. For a small business, the better question is whether stronger national hiring is appearing in actual inquiries, signed work and reliable payments. Keep the macro picture in view without letting one release substitute for your own demand evidence.
YOUR RESET
Our reset for the week: choose one number worth checking beneath the headline. For AI video, it is the cost of a usable result. For security, it is the access a credential permits. For the economy, it is whether interest becomes paid work. Better decisions start with the denominator.
Thanks for reading,
The Great Reset
Help us spread the word and tell a friend:
Want to advertise with us?
DISCLAIMER:
This newsletter is strictly educational and is not investment advice or a solicitation to buy or sell any assets or to make any financial decisions or investments. Please be careful and do your own research.
