AI Frontier Collaborative Mastermind + AI Mastermind AfterShow · November 3, 2026 · Noon and 1 PM Central
AI is eating capital at a scale that makes normal people blink twice. Data centers. Chips. Power. Models. More data centers.
Then a founder sits down and says the product still has to be simple enough for a stressed campaign team to use on election day.
That was the thread connecting this week’s AI Frontier Collaborative Mastermind and the AfterShow with Luke Svasti of Antelope Data. The Mastermind wrestled with the AI investment boom, the possibility of a bubble, changing discovery and the reliability of online information. Luke brought the conversation down to street level: a tool can have sophisticated data, agents and automation underneath it, but if the workflow adds stress, the customer will ignore it.
Infrastructure gets the headlines. Workflow decides whether any of it matters.
The central question: When the AI spending cycle settles down, which products will still solve a problem people can feel?
What You'll Discover
Why a technology bubble and useful infrastructure can exist at the same time.
What massive AI capital spending does—and does not—prove about customer value.
Why discoverability may increasingly mean being understandable to both people and agents.
How Luke Svasti moved Antelope Data from surveys toward political campaign operations.
Why stressed organizations need simple workflows more than impressive feature lists.
How deep domain knowledge can become a moat when frontier models keep improving.
Mastermind: The bill is real. Is the value?
The AI Mastermind series is hosted by Joe Moore and Katrena Drake. This week Joe hosted the room while Chuck Boyce moderated the main discussion and later led the AfterShow interview.
1. A bubble can leave useful rails behind
Chuck opened with the cost of AI. The room had just watched a new round of earnings reports and infrastructure commitments. Jeff Valin called the market a bubble. Chuck compared the moment to the railroad buildout: companies can fail while the tracks remain valuable.
That distinction is worth holding.
Amazon’s October 29 earnings release said it expected roughly $220 billion in 2026 cash capital expenditures. Its SEC filing reported $75.9 billion in cash capital expenditures during the third quarter, primarily reflecting technology infrastructure supporting AWS growth and additional fulfillment capacity. Alphabet reported $51.4 billion in third-quarter capital expenditures, mostly technical infrastructure supporting AI.
Those are enormous investments. They are not proof that every AI startup is valuable. They are not proof the market is rational. They are proof that the underlying buildout is real.
The room disagreed about what happens next. Some saw a familiar speculative loop: companies investing in one another, buying infrastructure from one another and lifting valuations along the way. Others saw the normal mess of a new platform being built before the stable business models arrive.
Both can be true. The internet survived the dot-com crash. Plenty of dot-com companies did not.
2. The next customer may be an agent
Jeff asked a strange question that is going to sound less strange every year: why are we marketing only to humans if agents increasingly search, compare and act for them?
The room explored what it might mean for a website to become legible to an AI agent. Could an agent identify the facts, pricing, authority and next step without being manipulated by design tricks built for human attention? Would businesses eventually optimize for machine retrieval the way they once optimized for search results?
This was exploratory discussion, not a settled technical standard. The practical signal is narrower: businesses need clean, structured, consistent public information. If a human cannot understand what you do, an agent probably will not rescue the message. If your claims conflict across pages, directories and profiles, machine discovery can amplify the confusion.
The answer is not hidden text for robots. It is a better source of truth.
3. Context is the real data problem
Chuck asked what happens when false information enters training data and becomes difficult to correct. Tylan Miller pushed back on treating every technical error as a moral failure. Joe returned to context: data comes from specific databases, websites and people, and those origins shape the answer.
That is where the conversation became useful for consultants and business owners. “AI said it” is not a source. A generated answer is the end of a chain. You still need to inspect the links.
For a business, that means knowing which public pages establish identity, which profiles are stale, where numbers came from and who owns the correction process. Reputation in an AI-mediated web is not only about publishing more. It is about reducing contradiction.
The unresolved question is accountability. If an AI system synthesizes ten sources and three are wrong, who fixes the answer? The model provider? The source websites? The business being described? The room did not manufacture a clean answer because one does not exist yet.
AfterShow: Luke Svasti and the workflow beneath political data
Luke Svasti joined Joe, Chuck Boyce and Jeff Valin for an election-day conversation about Antelope Data. Luke described an early-stage platform for campaigns that combines data, outreach, surveys, volunteer management and operational tools in one place.
Antelope began as a broader survey product. The survey market was crowded. Conversations with people working in politics revealed a sharper problem: campaign teams often have data and disconnected tools, but not an easy way to turn that material into the next useful action.
4. “Turnkey” is not a feature. It is the product.
Luke’s biggest change was not another model or dashboard. It was a commitment to making the system feel almost like a simple chat box.
Political campaigns operate under unusual pressure. They face opponents, deadlines, public scrutiny, volatile events and a fixed finish line. At the same time, many campaign organizations adopt technology slowly and rely on familiar processes.
“Whatever you build has to take away from stress.”
Luke Svasti, AfterShow at 05:30.
That sentence reaches far beyond politics.
Builders love capability. Customers buy relief. If a tool requires a new vocabulary, a training program and six integrations before it becomes useful, the customer experiences the tool as another problem.
Luke’s version of turnkey includes pulling data, deploying agents, managing volunteers and handling campaign activities without requiring the user to understand the technical machinery underneath it. The important test is not whether the system is powerful. It is whether a first-time candidate or small campaign can move from question to action without getting lost.
5. Polling asks who will win. Listening asks what people need.
Luke challenged the traditional campaign focus on likely swing voters and reliable donors. He argued that campaigns often collect data to predict a vote or target a message, not to understand what constituents are experiencing.
He contrasted large, expensive polling with smaller hyperlocal listening exercises. His proposed sample sizes and methods were his product perspective, not a universal statistical rule. The larger point was about purpose: a campaign can ask only “Who will you vote for?” or it can ask what residents want changed on their street, in their school or in their town.
That difference changes the workflow. Listening produces more than a score. It produces material for follow-up, local content, constituent service and another question.
Luke also acknowledged the hard part: consent, privacy and misuse. Removing personally identifying information is one safeguard, not the whole answer. Political data can affect persuasion, access and power. A responsible system needs clear collection practices, limited access, appropriate retention and a reason for every field it gathers.
The room did not treat privacy as a footer. Neither should the product.
6. The buyer, the channel and the competitor may be the same person
Luke described campaign consultants as a complicated part of the market. They can introduce a tool to candidates and help implement it. They can also benefit from complexity that keeps the campaign dependent on their services.
That creates an incentive problem. The candidate wants to win. A consultant may be measured by activity, coordination and continued involvement. A tool designed to simplify the work can make the consultant more effective—or threaten the work being billed.
This is not unique to campaigns. Agencies, software implementers and advisors live inside the same triangle. Are you selling through them, replacing part of their labor or giving them leverage?
A founder needs an honest answer before choosing a channel. “Everyone with a budget” is not an ideal client profile. Luke has narrowed Antelope toward campaigns that feel the pain of fragmented tools, need affordability and are willing to experiment with a new operating model.
7. Domain depth is harder to eat
Joe connected Luke’s experience to a problem he has watched across software marketplaces: a clever app appears, gains attention, then a larger platform absorbs the feature and the smaller product disappears.
Luke’s answer was to go deeper.
“I think that it's better to niche down and build a product that's a little bit better, more organic, really suits the workflow of the industry that you want.”
Luke Svasti, AfterShow at 31:49.
A frontier model can generate another survey. It can generate a campaign dashboard. It can generate code.
It cannot instantly inherit a year of conversations with candidates, organizers, data people and consultants. It does not automatically know which step fails at 9 PM, which spreadsheet keeps arriving broken or which feature creates more work than it removes.
That knowledge is the moat. Not because the model cannot learn it, but because somebody still has to live close enough to the work to teach it.
Luke also named the cost of that depth: loneliness. Builders chasing a fast release may not understand why someone spends months embedded in one difficult niche. The slower path can look irrational right up until the generic features become commodities.
Put it to work
These are editorial exercises derived from the conversations, not assignments issued to every participant.
Separate infrastructure from outcome. For one AI expense, name the customer result it produces. If the answer is only “more capability,” keep digging.
Map the stress. Write down the three moments when your customer is most rushed, confused or exposed. Design for those moments first.
Find the source of truth. Ask an AI system what your business does. Trace the answer back to the public pages it appears to use and correct contradictions.
Interview around the workflow. Do not ask only which features people want. Ask what happened immediately before the problem and what they did immediately after it.
Name the incentive conflict. Identify who buys, who uses, who recommends and who may lose billable work if your product succeeds.
Run the frontier-model test. If the best general model becomes three times better tomorrow, what domain knowledge, trusted access or operating workflow still makes your product necessary?
News and resources
Amazon’s third-quarter 2026 results — editorial background: primary-source context for the Mastermind discussion about infrastructure spending.
Alphabet’s third-quarter 2026 results — editorial background: primary-source context for Alphabet’s AI infrastructure investment.
Amazon’s Q3 2026 SEC filing and Alphabet’s Q3 2026 SEC filing — editorial background: financial details used to check claims raised in the room.
Antelope Data — spoken source: Luke’s campaign-management project discussed in both sessions.
Federal Election Commission data — editorial background: an official public-data destination relevant to Luke’s discussion of campaign data.
Who was in the room
Series hosts: Joe Moore and Katrena Drake. This week’s moderator and AfterShow interviewer: Chuck Boyce.
AfterShow participants: Luke Svasti, Chuck Boyce, Joe Moore and Jeff Valin.
What remains unresolved
How much AI infrastructure can the market support before customers demand clearer returns? What information should be optimized for agents, and who decides when that becomes manipulation? Can political listening tools collect useful local signals without turning citizens into targeting profiles?
And for builders: how deep into a niche do you have to go before the workflow becomes something a general model cannot casually replace?
Continue the conversation
The AI Mastermind meets Tuesdays at noon Central. Bring the question you cannot answer from the headlines alone.
Register for the AI Mastermind










