In the world of entrepreneurship and agency management, artificial intelligence has shifted from a futuristic concept to a daily necessity. For many of us, it’s the first thing we consult in the morning and the last “colleague” we sign off with at night. We are using these models “all day long”, feeding them business plans, marketing copy, client data, and complex code, all in the name of unprecedented convenience and productivity.
But this convenience comes with a nagging, unspoken anxiety. There’s a growing sense that in our rush to gain an edge, we are trading away something invaluable. Preshton Pysh admits: “I can fully understand how it just knows everything about me at this point… And I’m not proud of that”.
This is a personal dilemma and a critical business risk. We’ve embraced a “trade-off”, accepting opaque terms of service in exchange for efficiency. The problem is that the habit of handing over our core intellectual property is forming faster than the safeguards. We are becoming “addicted to using these types of models” without fully understanding the terms of the deal.
This central tension—the incredible utility of AI versus its profound, hidden risks—was the focus of a recent, in-depth discussion with Preshton Pysh and co-founder of Maple AI and OpenSecret, Mark Suman.
The conversation is a must-see for any business leader trying to navigate this new landscape.
Prefer to read? We’ve got you!
For those who prefer to read or can’t watch the video right now, the following report breaks down the key insights from this critical discussion.
We go beyond a simple summary to provide a detailed analysis of the strategic implications for your business with timestamps as reference.
The Real Threat: It’s Not Just Data Leaks, It’s “Subconscious Censorship”
What became clear from the video is that the risk isn’t just about privacy. It’s about the potential loss of the very originality and independent thought that makes a business competitive.
When business leaders think of AI risk, they typically think of data leaks. This concern is entirely valid. Recently, bugs in both ChatGPT and Grok led to private chats “being indexed on Google search results” (0:15:24, 0:15:46). Imagine a competitor searching Google and finding your internal strategy discussion, or a client’s sensitive “marriage details” (0:16:01) from a shared chat. When you give your data to a third party, this risk is non-negotiable (0:16:01).
But this is the most obvious, and frankly, the least of your worries. The real, long-term threat isn’t just that your data is seen; it’s that your data is used against you in ways you can’t detect.
Mark introduced a chilling phrase for this: “subconscious censorship” (0:09:41).
To understand this, look at the methods “already used with social media feeds” (0:10:38). We all know that algorithms can “affect your emotional state” (0:11:05) or “keep you in an angry state” (0:11:05) simply by re-ordering the content you see. They are “tools of persuasion” (0:11:28) designed to maximize engagement.
Now, apply that same logic to an AI that “knows you intimately” (0:11:28). It has your business plans, your strategic thoughts, your personal anxieties. This AI can, over “weeks, months, years” (0:11:57), subtly “place an anchor of a false fact” (0:11:28) in its responses. It can “emotionally trigger” (0:11:28) you or gently “guide you” (0:11:57) toward a specific conclusion. You won’t notice it’s happening. You’ll just find yourself “guided into this rut” (0:11:57), believing the ideas were your own.
For an entrepreneur or agency, what is “the most unique thing about you”? It’s “your memories, your thought process, the way that you perceive the world” (0:09:18). This is your competitive edge. Giving this to a proprietary system is, as Mark warned, “giving up the thing that makes us uniquely human” (0:09:18).
Think of it as the ultimate conflict of interest. What if your brilliant, disruptive business idea threatens the ecosystem of the AI’s parent company? A proprietary system could be instructed to subtly guide you away from that idea, labeling it unfeasible or steering you toward a less-threatening alternative. It’s a silent, persuasive consultant in your boardroom with a hidden agenda.
The “Don’t Trust, Verify” Solution: Understanding Verifiable AI
The antidote to this “black box” is not another marketing promise of “privacy.” The antidote is proof.
The solution is a new paradigm called “verifiable AI” (0:06:00). This concept is built on the simple but powerful ideology: “Don’t trust, verify” (0:06:00). For a business, this means shifting from trusting a company’s privacy policy to demanding mathematical proof that your data is secure.
The podcast discussion outlined three pillars of this approach:
- The “Utopia”: Local AI. The most private AI is one that runs entirely on your own device (0:00:21, 0:16:46). “It’s never going to get more private than that” (0:00:21). The problem is that most of us don’t have devices powerful enough to run the best models… yet (0:00:21, 0:16:46).
- The “Problem”: Open-Source Code. Many companies claim to be private by making their code open-source (0:16:28). This is a good first step, but it leaves a critical gap: “how do I know that you’re actually running that exact same code on your servers?” (0:17:27). You don’t. You’re still trusting them not to have a second, hidden version of the code that spies on your data.
- The “Practical Solution”: Secure Enclaves. This is the practical, powerful solution for today. A Secure Enclave, or “trusted execution environment” (0:06:28, 0:17:12), is a “digital vault” in the cloud. It allows the provider to run the AI for you without being able to see any of the data you’re processing.
Mark provided a perfect analogy: this is the new “lock icon” (0:17:47) for the AI age.
We all know the difference between an unsecured http:// site and a secure, encrypted https:// site. The speaker calls this new standard “HTTPS E” (01:11)—the ‘E’ stands for Enclave. This system provides an “attestation” (0:17:12), or “mathematical proof” (0:17:12), that the server is running the exact open-source code it claims to be. It proves the provider cannot see your data.
This is a fundamental paradigm shift from Trust-Based Security to Proof-Based Security. Even the best trust-based systems, which rely on “third party auditors” (0:04:48), require you to believe their intentions. Verifiable AI requires you to believe only in math. The new standard for any business should not be “Do you have a privacy policy?” but “Can you provide a verifiable attestation?”
This verifiable approach isn’t theoretical; it’s being implemented right now.
Platforms like MapleAI are emerging that are built on this “don’t trust, verify” principle, allowing businesses to start protecting their AI interactions today.
Addressing the “But Is It Good Enough?” Question
The single biggest objection from any business owner is blunt: “The open source versions are just not even close to what these newest models are doing” (0:19:32). If protecting privacy means sacrificing performance, it’s a non-starter.
This is the old way of thinking. The reality is that the gap is closing at an astonishing rate. “We’ve seen the open models catch up a ton,” the Mark noted. “Now they’re like in the 90% range” (0:20:00) of their proprietary counterparts. In some areas, they’ve already surpassed them. For example, “Quen3 Coder” (0:20:25) is an open-source model that “is scoring just as good as some of the proprietary models on programming” (0:20:25).
But the more crucial point is this: the race for one giant, “general model” is the wrong race to watch. The future of AI for business is specialization.
We are beginning to “see more specific models” purpose-built for “the medical field and legal field” (0:23:46) and other industries. Mark used the analogy of a “general contractor” (0:24:04). The future isn’t one AI that knows everything; it’s a “general model that acts as a router in the front” that “pulls in… specialists” (0:24:04) as needed.
This is revolutionary for businesses. An agency doesn’t need a generalist that can also write poetry; it needs a best-in-class, specialized coding model. A legal team needs a verifiable, specialized contract-review model. In these specialized fields, open-source models are not just “catching up” (0:21:29); they are often superior.
This is the “have your cake and eat it too” (0:12:48) moment. The trade-off between performance and privacy is dissolving. The smartest businesses will be the ones who stop paying for the biggest generalist and start leveraging the most efficient, verifiable specialists.
The Future We Should Demand: “Sovereign AI Memory”
There is one more battleground for AI control, and it may be the most important: who owns the AI’s “memory” of you?
We all want the convenience of an AI that remembers our context and preferences—a private “memory bank” (0:26:57). But in the current proprietary model, this memory is a closely guarded secret.
The podcast guest offered a powerful analogy: using a proprietary AI is “like you’re sitting down with a biographer… The difference is in a proprietary system, you don’t get to read that biography” (0:28:51).
This “biography” is your company’s “essence” (0:30:05), and you don’t own it. Worse, you can’t even be sure you can delete it. Proprietary systems may “show you an interface that says… ‘we’ll even let you delete it.’ But there’s no guarantee that that’s actually happening” (0:29:19).
This opaque, unreadable memory is the very mechanism that makes “subconscious censorship” possible. The AI can decide which of your memories to “overweight” (0:31:34) and which to “suppress” (0:32:14) to control the context of all future interactions.
The solution is to demand “a truly sovereign AI memory” (0:29:35). This is a memory that you control—one you can “go in and see… and then you can edit it, you can add to it” (0:29:35). While the engineering challenges are significant (0:30:52), this is the necessary goal.
For a business, sovereign memory means true, verifiable control over your intellectual property. It’s a future worth demanding.
The “Line in the Sand”: Building Your AI Toolbox
So, where does this leave the practical business owner? Will we all be running our own “home server” (0:48:54) for AI in the future?
Perhaps. But the Mark offered a more useful parallel: the email server (0:50:11). We all have the capability to run our own email servers, “but we don’t.” Why? Because using Google is “just so dang convenient” (0:50:11).
We’ve made that trade-off. But the stakes here are infinitely higher. This is the “line in the sand” (0:50:34) that defines the next decade of technology. As the speaker so perfectly put it: “you can have our emails, but you can’t have our brains” (0:50:34).
The solution is not “all or nothing.” It is not about “throw[ing] away ChatGPT” (0:52:52).
The strategic solution is to “view… AI as a toolbox” (0:52:52). You have “different tools to use for different things” (0:52:52). This is the new professional standard:
- Use Proprietary Tools (like ChatGPT) for public, low-risk, non-sensitive tasks.
- Use Verifiable Tools (like Maple) for everything that matters: your business strategy, your client data, your code, and your “personal information” (0:53:07).
The goal is to regain that “refreshing feeling knowing that this is just a private room with you and an AI and nobody else is listening” (0:53:26).
Navigating these complex technological shifts is no longer optional—it is essential for business survival. Having the right partners, the right strategy, and the right tools makes all the difference.
