Claude and the Riemann hypothesis: What independent checks have found
Anthropic mathematician Levent Alpöge says he found a proof of the Riemann hypothesis working with Claude. The Lean and Mathlib communities have since reported that independent proof checkers accepted the argument and that it proves the standard statement of the hypothesis.
The Clay Mathematics Institute still lists the problem as unsolved as of 22 November 2026, New Zealand time. The two updates answer different questions. The Lean checks tell us whether the formal argument holds. Clay’s listing records the institute’s recognition of the result.
Where things stand
Independent machine checks have passed. The Lean and Mathlib communities say they have full confidence in the proof. Clay still lists the problem as unsolved.
What Alpöge announced
In a note dated 20 November, Alpöge describes an October-to-November collaboration with Claude and credits the model with helping develop the argument near the finish. He released the argument as a library for Lean, a language and proof assistant that lets a computer check each deduction against formal rules. The Lean community puts the library at about 17 million lines of code. Alpöge also reported that the method extends to the generalised Riemann hypothesis for Dirichlet L-functions.
Read Alpöge’s original note and timeline (PDF).
What the hypothesis says
Prime numbers thin out as numbers get larger, but their spacing is irregular. Riemann found that the zeros of a function now called the Riemann zeta function encode information about how the actual count of primes deviates from its expected trend. The hypothesis says that all the nontrivial zeros, the ones relevant to this question, lie on a vertical line where the real part equals one-half. A proof would give mathematicians strong bounds on how far the count of primes can stray from its expected value.
Watch the zeta function take shape
3Blue1Brown’s visual explainer shows what the zeta function does and where the vertical line comes from. It’s background for readers coming to the problem for the first time.
Watch or read the illustrated lesson on 3Blue1Brown.
What independent verification found
The Lean and Mathlib communities first isolated the accepted mathematical statement. Two comparison tools matched Alpöge’s theorem to that independently defined target. Several proof checkers, including Lean’s official checker and separately implemented tools, then accepted the exported proof. This is substantially stronger evidence than a report that the files merely compiled: the checks cover both what the code proves and whether its proof is valid under the formal rules.
Read the Lean and Mathlib joint statement or see MIT mathematician Andrew Sutherland’s verification records.
What happens next
Mathematicians can now work through the argument, explain its central ideas in a form people can read and investigate what the method can do elsewhere. Clay follows a separate process for a Millennium Prize. Its rules call for qualifying publication, at least two years after that publication and general acceptance by the mathematics community before a proposed solution can be considered for the prize.
Check the Clay Mathematics Institute’s problem page.
What New Zealand teams can take from this
The transferable part is the way the result was checked. For a business testing AI-generated work, define the expected result independently, then test the output against it. If a model builds an invoice reconciliation, for instance, compare its totals with a known ledger and test cases, and have someone investigate discrepancies before approving it. Here, the checker could verify a mathematical deduction. In a business workflow, the equivalent might be a test suite, a reconciliation against source records or review by the person responsible for the decision.
Related coverage: What Anthropic’s Nasdaq listing means for Claude customers.
Anthropic’s Nasdaq IPO: What Claude customers can learn from the listing
Anthropic began trading on Nasdaq under ANT on 19 November 2026, US time. For New Zealand businesses that use Claude, the immediate change is that there’s a new public record to consult when assessing the provider behind the product.
An initial public offering is the first sale of a company’s shares to the investing public. The listing lets those shares trade on an exchange. Anthropic subsequently announced that the offering had closed. The IPO concerns the supplier’s ownership and funding. A Claude subscription or API integration still needs to be assessed against its own contract and service terms.
Back to the 22 November Weekly Highlights
Where to look
Start with the issuer’s disclosures for risk factors, commitments and financial updates. Use Nasdaq for the listing and Anthropic’s product documentation for service details.
What public filings can tell a customer
A public filing can show how a company describes its exposure to computing costs, infrastructure partners and competition. Read the risk factors and the business description together. For example, if a provider identifies capacity constraints as a risk, check which supported deployment options your own integration uses and what happens if those options change.
Where you buy Claude affects who you contact
A team might use Claude through a subscription, the Anthropic API or a cloud intermediary. Those are different purchasing and support arrangements. Find who bills you, who handles service incidents, and who can confirm how your account is configured. Keep the account owner and any contractual commitments in the same supplier record as the integration itself.
A supplier review for a New Zealand team
Consider an agency that uses Claude to turn approved client briefs into first drafts. Record which account processes the briefs and where the approved source files are stored. Keep prompts and templates in a location the agency controls. Run a small trial that recreates one approved output with a second model or a manual template, and measure how long the handover takes.
A question to revisit after the next filing
Did the provider disclose anything that changes your understanding of service cost, capacity or dependence on a third party? If it did, compare that disclosure with your actual agreements and workload. Track any action with a named owner and a date. An account-level question about retention, support or availability should go to the provider or reseller that supplies your service.
Read the filing
Pull out one dependency relevant to your workload.
Check the agreement
Locate the support contact, data terms and billing owner.
Test the fallback
Recreate one output and time the handover.
Questions readers may have
Does the listing change the Claude account I already use?
Consult your supplier’s product notices and agreement for account-specific changes. The listing announcement concerns the shares.
Where can a procurement team verify the listing?
Nasdaq’s announcement and Anthropic’s closing release give the listing and ticker details.
Sources and further reading
Gemini 4 Neon: How to test Google’s faster model on real work
Google introduced Gemini 4 Neon on 18 November 2026 for multi-step work such as coding, research and document-heavy analysis. If you’re considering a switch, measure how long it takes to get an output your team can use.
Google describes Neon as combining reasoning and speed. A faster response may improve a tool people use throughout the day. In work with multiple steps, the time spent checking sources and correcting omissions can matter more than the time before the first answer appears.
Back to the 22 November Weekly Highlights
Where to try it
Google’s launch page lists access routes including Google AI Studio and the Gemini API. Check the product surface, account and region your team actually plans to use.
Pick a task where a wrong answer is easy to spot
A digital agency could supply an anonymised monthly reporting pack and ask the model to explain three changes in project hours. Include two trends that can be verified from the data and one question the data cannot answer. Require a reference to the correct tab and period for every number. Grade the response against an answer key before opening the second model’s output.
Measure the time through approval
Run the same brief on Neon and the model already in use. Keep the input files, tools, access and expected output format consistent. Record completion time, incorrect figures, unsupported claims and minutes spent editing. Repeat on several representative cases, including one incomplete or inconsistent source file. Flag any error that would mislead the client as a failed run.
Watch an independent model comparison
Paul J Lipsky demonstrates pre-release tests with Neon on coding and design tasks. His examples show how task choice affects results. Use the same approach with your own brief and scoring criteria.
Treat model and product access as separate decisions
If the pilot works in AI Studio, rerun it in the intended deployment environment. A model can perform well in a browser while the API integration has different permissions, quotas or latency. Check the available settings and the data terms for the account you intend to use before giving it production documents.
A practical decision rule
Start with the workflow that has a named reviewer and good source material. If Neon matches your accuracy threshold and consistently reduces time to approval, keep testing with a limited rollout. If the results swing widely, look for the input condition that caused the misses before changing your default model.
Use a shared brief
Hold the inputs and output requirements constant.
Score the result
Check sources, missing-data handling and review time.
Repeat in deployment
Verify access, quotas and latency in the intended product.
Questions readers may have
Is Neon the same as Gemini’s Mac computer use?
Neon is a model. Mac computer use is a feature of the Gemini Mac app with its own availability and controls.
What should I do with vendor benchmarks?
Use them to select a few workloads worth trying. Decide based on your repeatable business tests.
Sources and further reading
Le Figaro’s OpenAI deal: Can AI discovery send readers to publishers?
Le Groupe Figaro and OpenAI announced an agreement to make Le Figaro journalism discoverable in ChatGPT. The companies describe an experience in which relevant reporting can be surfaced with source attribution and a link back to the publisher.
That creates another way for a reader to encounter a story. For a publisher, the commercial question is what happens next. A linked mention has one value as a citation in an answer and another if the reader follows it, reads the original or signs up.
Back to the 22 November Weekly Highlights
What the announcement covers
OpenAI describes linked summaries and excerpts from Le Figaro journalism. The announcement describes the partnership. It gives a New Zealand publisher no traffic baseline for its own site.
Follow the reader’s path
A reader asks a question, the answer cites a source, and a link can take them to an article. At that point, the publisher can observe the visit if the analytics setup receives a usable source signal. What happens inside the answer before the click is a separate measurement question. Keep referrals separate from any manually observed mentions or reported impressions.
Set up a small measurement experiment
Choose five evergreen articles with distinct questions they answer. Record existing channel and landing-page performance. In your analytics platform, inspect session source and landing page together. Build a view for identifiable AI referrals, track engaged sessions and one meaningful conversion such as a newsletter sign-up, and keep a separate category for unknown attribution.
Check the pages readers actually arrive on
On each article, make the author, publication date and subject clear. Review canonical URLs and make sure a returning reader lands on the intended article. Link to one natural next step, such as a related guide or an email subscription. Verify that the page works on mobile and that its story is visible under your chosen subscription or paywall rules.
Keep the commercial terms distinct from the analytics
A negotiated distribution agreement can involve permissions about content and presentation. Editorial, product and commercial teams may each care about different terms. For an independent publisher evaluating its own channel, start with the reader outcomes you can measure and keep source attribution, copyright permissions and commercial rights on a separate list of questions.
Discovery
A relevant answer links to a publisher’s reporting.
Arrival
A reader opens the original article and receives the publisher’s experience.
Outcome
Review engagement, registration or another chosen event.
Questions readers may have
Can I count every mention as a visit?
Use website analytics for visits and track observed citations as a different signal.
Where should I start with a small site?
Choose a few guides with reliable traffic and a clear next action. Compare the quality of identifiable referrals against your baseline.
Sources and further reading
US AI chip restrictions: What New Zealand cloud customers should check
Associated Press reported that Alibaba and Tencent expect tighter US restrictions on advanced AI chips to affect development or expansion plans. For a New Zealand organisation buying AI through a cloud platform, the question is whether a particular workflow depends on capacity it cannot readily replace.
An AI service relies on compute to train models and to run them when customers make requests. Cloud providers pool that capacity across products and regions. Restrictions that affect future deployments can change a provider’s expansion options while existing customer endpoints continue serving requests. Availability depends on the individual service and the provider’s published commitments.
Back to the 22 November Weekly Highlights
Read the reporting
AP attributed specific concerns to the companies. Alibaba described potential effects on some data centre expansion plans. Treat those comments as company assessments.
Trace the path from a button to the model
Start with one workflow that people would miss if it stopped. A support assistant might receive a message in a website form, pass it to an automation platform, retrieve an approved policy and call a cloud model. List the provider, model version, endpoint region and account for each step. Ask the vendor to confirm anything the dashboard leaves unclear.
Identify what can move and what must stay put
A second region or model can be useful only if it can handle the required task and fits the organisation’s data controls. If a backup would send the same input to a different jurisdiction or vendor, review that route before enabling it. Check the exact contract and account settings for data handling, capacity and support.
Rehearse a degraded service
On test data, simulate an unavailable model by disabling the test connection. Confirm that the application queues work or sends a clear message to an operator. Set a reasonable cap on automatic retries. For a time-sensitive customer enquiry, offer a plain route to a person while the AI function is unavailable. Time the recovery and record who has to intervene.
Keep monitoring specific
Track actual service signals such as repeated timeouts, quota failures, retirement notices or a confirmed region change. Compare them with the service-level terms that apply to your account. An export-control headline is a useful starting point for checking exposure. Service monitoring tells you whether your own workflow needs action.
Map the service
Record the provider, model, region and account.
Test continuity
Decide whether to queue, route to a person or use an approved substitute.
Watch service signals
Track real outages, quota changes and model notices.
Questions readers may have
Do the reported restrictions establish a New Zealand service interruption?
AP reports potential impacts on specific companies’ plans. Ask your own provider about a named service, endpoint and region.
What counts as a fallback?
A documented route that has been tested with the same class of inputs and has a named owner. It can be a manual workflow.
Sources and further reading
WordPress 7.2 release candidate: What to test before upgrading
WordPress published its first 7.2 release candidate on 17 November 2026 and announced a planned final release on 8 December. For a site owner, the candidate is a chance to see whether the next version works with the theme, page builder and plugins the business relies on.
A release candidate is a pre-release build made available for testing. WordPress describes RC1 as including further fixes and accessibility and compatibility improvements. Use a staging copy with the same setup as the production site so a successful test is relevant to the actual upgrade.
Back to the 22 November Weekly Highlights
Choose the test environment
Ask your host for a staging copy and confirm the restore point. Disable delivery to real customers and route forms and order emails to controlled test inboxes.
Record what works before changing versions
Capture the core, theme, PHP and plugin versions on the staging copy. Run a mobile enquiry and a desktop editing check first. Photograph or screenshot a representative page and note what happens at submission. Fix or document any pre-existing issue so it doesn’t get mistaken for a release regression.
Test the page builder and the visitor journey
For a WPBakery site, open a representative page in the builder and confirm that rows and modules load. Edit a small piece of copy, save it and view the result on mobile and desktop. Check a grid or query that pulls posts by tag. Then submit an enquiry and verify both the visitor confirmation and arrival in the test inbox or CRM.
Add store and editorial checks where they apply
In a store’s test mode, try a product variation, add it to the basket and step through checkout using a sandbox payment setup. Review shipping and order emails. For a site that publishes regularly, check draft previews, scheduled publishing and one post with a featured image. If your team uses the block editor, test reusable patterns and review steps with the roles people actually have.
Make the production decision from the results
List failed checks with a URL, steps to repeat, expected behaviour and a screenshot. Assign a person to check whether a theme or plugin update fixes each failure. Before moving the stable release to production, repeat the important checks against that release and confirm that a usable restore point is available.
Baseline
Record a passing test on the current version.
Stage
Install the candidate on the isolated copy and repeat the same checks.
Approve
Repeat critical tests on the stable release before the live upgrade.
Questions readers may have
Can I use RC1 on my public site?
Use a staging or test copy. WordPress publishes candidates so people can identify issues before the final release.
Will block editor improvements change my WPBakery pages?
Test the editing and public display of a representative WPBakery page. Core, theme and plugin compatibility determines how your particular site behaves.
Sources and further reading
- WordPress 7.2 RC1 and official testing instructions
- WordPress Core 7.2 project and schedule
- WordPress Test’s illustrated 7.2 testing guide
GPT-6.2 Astra or Sol? How to choose an AI model for work
OpenAI introduced GPT-6.2 Astra and Sol on 17 November 2026. Sol is positioned for regular professional work and Astra for harder tasks. For a business, the useful distinction is how much checking and rework each model leaves behind.
A quick draft can still be expensive if a reviewer has to rebuild it. Before choosing a model for your team, test both on the same work and measure the time it takes to get to an output you can actually use.
Availability at launch
OpenAI lists Plus, Pro, Enterprise and Edu for the launch in ChatGPT Work and Codex, with both models also offered through the API. Check your workspace for access and administrator controls. OpenAI release notes.
Header image: AI-generated editorial illustration for ADJ Systems.
What Astra and Sol are designed to do
OpenAI presents Sol as the everyday choice for work such as writing, analysis and coding. It presents Astra as the higher-capability option for harder projects. Both names refer to models. The product you use them in determines which tools and files they can access.
Start with Sol when the brief is clear
Try it on a structured task such as turning approved meeting notes into an internal action list. Define the format, give it the source material and ask it to mark decisions that still need confirmation.
Try Astra when the review load is high
Test it on a task that requires reconciling multiple sources, explaining conflicting figures or preserving a complex workbook. Give it the same inputs and measure how much additional work the reviewer has to do.
How to compare the models on your own work
Start with a monthly operations pack. Give each model the same anonymised sales export, inventory notes and definition of the metrics. Ask for a one-page summary and a short list of anomalies. Set a rule that each figure must trace back to a named sheet or row, and keep an answer key for the figures that matter most.
1. Use the same brief
Keep the prompt, sample files, tools and output format consistent. Include a messy example with a missing value or a mismatch between sources.
2. Review against the source
Check the totals, units and assumptions. Record unsupported statements and measure editing time. Mark a critical error as a failed task even if the draft reads well.
3. Compare accepted outputs
Compare the share of outputs a reviewer can approve, how long review took and your actual usage cost or plan limits. Repeat on a few examples before making a choice.
See OpenAI’s comparison examples, including a workbook and slide deck
Watch OpenAI’s paper-lamp design video under From CAD to a paper lamp. It follows a single Astra design from reference photos through revisions to a physical prototype.
What benchmarks can tell you
OpenAI publishes evaluation results and example outputs alongside the launch. They show how the models performed on those tasks under those conditions. A clean benchmark spreadsheet can be quite different from a workbook with local naming conventions, missing rows and undocumented formulas. Your own acceptance criteria should decide whether any improvement matters.
What New Zealand businesses can do now
Pick one repeatable job, set a review limit and run a small trial with approved material. If Sol meets the standard, use it for similar work. If Astra consistently reduces review time or errors on complex cases, reserve it for those jobs. Keep a person responsible for approvals and check that your workspace settings allow the files you intend to use.
Back to the 22 November Weekly Highlights.
Sources
Cloudways AI Website Builder: From first prompt to a launch-ready WordPress site
Cloudways has opened its AI Website Builder to general use. It can generate a WordPress or WooCommerce site from a written brief and give you a version to edit before publishing on Cloudways Flexible.
DigitalOcean says the generated pages use standard WordPress Gutenberg blocks. That gives a business an editable starting point. Its readiness depends on whether the navigation, content and integrations serve the customer’s actual needs.
Back to the 22 November Weekly Highlights
Where it fits
Try it for a new-site prototype or a small marketing site with a clear brief. If you already have a site, map the old URLs, data and integrations before committing to a rebuild.
Write a brief the builder can act on
A New Zealand renovation company might specify a four-page site with service areas, a project gallery and a quote enquiry. Supply approved service descriptions and say what happens after an enquiry. Specify a mobile-first navigation, the contact details, and which testimonials and photos are cleared for use. Mark claims the tool should leave as placeholders.
Review the first build with a customer task
Open the generated site on a phone and try to check whether a suburb is served, find a relevant project and submit a test enquiry. Record confusing labels, slow media and missing information alongside factual errors. Correct the page or brief where appropriate, then run the task again. Ask someone who wasn’t involved in the build to do the same.
Watch a build from prompt to site
This independent walkthrough builds a developer portfolio and shows the Cloudways staging workflow. Use the current help guide for the latest interface and options.
Connect the parts a business already uses
Match fields in the site’s enquiry form to the existing inbox, CRM or Make scenario and check the handover with a dummy submission. For WooCommerce, confirm currency, shipping zones, stock, payment settings and the order confirmation using test mode. Check tracking consent and conversion events on the same journey so the team can see whether it works after launch.
Budget for ownership and handover
Confirm who controls the WordPress administrator account, domain, backups and hosting billing. If the build replaces an existing website, test redirects and integrations before switching the domain. Cloudways’ support documentation says the builder’s AI editor works with sites created through the builder. Review migration and editing needs before choosing it for a replacement project.
Specify the job
List pages, approved content and one priority customer action.
Test the journey
Complete that action on mobile and confirm the handover.
Assign ownership
Identify who will run the site, the domain and its integrations.
Questions readers may have
Can the AI editor take over my existing WordPress site?
Cloudways’ FAQ says the AI editor requires a new site created with the AI Website Builder.
Can I use it to prototype a store?
Cloudways lists WooCommerce support. Use sandbox payments and check shipping and order emails before going live.
Sources and further reading
- Cloudways’ general availability announcement
- Cloudways step-by-step guide
- Cloudways builder FAQs
- DigitalOcean’s announcement on Gutenberg blocks and hosting
Canva’s AI controls: What each switch changes in your design workflow
Canva introduced two controls for Canva Free and Pro. One switches generative AI features on or off. The other changes what you see in the library by prioritising human-created content. Choose the setting that matches what you want to control.
If a client brief calls for supplied photography and an approved set of graphics, you may want the team’s workspace to make those assets easier to find. If the brief also restricts generating new creative material, the generative setting addresses a different part of the workflow.
Back to the 22 November Weekly Highlights
Where to find the settings
For generative tools, go to Profile, Settings, then Permissions. For library personalisation, open Profile, Your account, then AI personalisation. Canva’s announcement shows the controls.
The generative switch affects creation tools
Canva says you can disable generative tools such as Magic Write, Magic Design and Magic Expand. Assistive tools such as Background Remover and Magic Grab remain available. If a brief draws a finer distinction between retouching and generated content, give designers examples that show which edits are acceptable.
The library switch affects what appears first
The separate personalisation control lets you see less AI-generated material. It prioritises human-created content in library results. For a campaign that permits only assets the client has approved, start from the brand folder and check an element’s source or metadata before using it.
Try both controls in one sample design
Have a designer open a copy of a social template. Turn off generative tools and confirm the controls match expectations. Enable the library preference and run a search for something common such as a house illustration. Compare the results with the approved brand assets. Ask a second user to repeat the test so you know which preferences need to be set per account.
Build a quick review into your handover
At sign-off, record the origin of prominent photos and illustrations, any edits a reviewer needs to know about and where the approved originals are stored. For a multi-person team, confirm how its Canva plan handles settings and permissions. Give freelancers and clients the same brief about allowed assets and editing tools.
Creation
Use the generative AI setting to control access to generation features.
Discovery
Use the library preference to move human-created results higher.
Approval
Check the assets in the actual design before export.
Questions readers may have
Will the library preference remove all AI-generated results?
Canva describes it as showing less AI-generated material and prioritising human-created content.
Can I still remove a background?
Canva identifies Background Remover as an assistive feature that remains available when generative tools are switched off.
Sources and further reading
- Canva’s announcement and settings screenshots
- Canva’s screenshot of the generative AI switch
- The Verge’s walkthrough of the two controls
Microsoft Copilot Autopilot: What recurring work can you delegate?
Microsoft began rolling out Copilot Autopilot on 17 November 2026. It is designed to follow work over time, so it can pick up a thread or run a recurring task without someone entering a fresh prompt each time.
For teams already working in Microsoft 365, one practical use is pulling project updates together for a regular review. A team can keep decisions with the project owner while giving the agent a clearly defined job.
Read Microsoft’s Ignite announcement
Rollout and access
Microsoft is rolling Autopilot out progressively to eligible Frontier customers. Its rollout guide lists a Copilot licence, usage-based billing and required security controls. Check Microsoft’s Frontier requirements.
Header image: AI-generated editorial illustration for ADJ Systems.
What makes Autopilot different
A chat starts when someone makes a request. Microsoft says Autopilot can carry a goal and context across time within an organisation’s Microsoft environment. It can watch Teams channels, follow up on email threads and return to a project days later. Microsoft describes an agent identity, activity logs and controls that let administrators limit or block access.
Those capabilities make the operating rules part of the workflow design. The team should know which conversation or files the agent can use, what it should do when information is missing, and who reviews anything it prepares.
A pilot for weekly project updates
Imagine a small delivery team that writes a Friday progress note. People post updates in a Teams channel, the lead checks a project tracker, and someone chases missing information. Give the agent a narrow goal: compile the changes and identify gaps for the lead to review.
1. Define the inputs
Use one approved Teams channel and one project tracker. Specify a weekly cutoff and require a link back to each update used.
2. Draft and flag gaps
Group updates by project, surface unresolved risks, and draft suggested follow-ups. Put conflicting or missing information in a review list.
3. Keep the lead in charge
The project lead checks the links, confirms sensitive wording and sends the final update. Record what needed correction.
When a fixed workflow fits better
If the job is simply to remind everyone every Thursday at 3 pm, a scheduled rule is straightforward to test and maintain. Agent-based automation earns a closer look when the work involves interpreting varied messages, spotting gaps across several threads or deciding what deserves to be surfaced. A team can use a fixed schedule for the trigger and an agent for the draft if that division makes the results easier to review.
What New Zealand teams can do now
Check Frontier eligibility with the Microsoft 365 administrator and confirm how usage-based billing would apply. While access is being checked, run the process manually for a week and time the reading, writing and follow-up. Agree on one named owner, suitable source material and which actions are reserved for staff.
During a pilot, count missed updates, unsupported claims and unnecessary reminders alongside minutes saved. If follow-ups create more noise than they remove, tighten the scope before expanding access. Keep the results visible to the team so it can see whether the workflow is helping.
Sources
- Microsoft’s Ignite announcement and Autopilot overview
- Microsoft Frontier rollout requirements
- Ignite 2026 Book of News










