Issue #64: “Weird” Frameworks for AI Inside Your Business

Good morning.

This week is about putting structure behind AI and how you’ll be building around it through people, workflows, and decisions that actually hold up over time.

We’ll start with what works when companies scale AI across teams. Two new reports show that success has less to do with tools and more to do with how leadership defines scope, sets feedback loops, and assigns ownership.

This shouldn’t be a surprise, right?

But it’s incredible how many normally cool-headed people lose their minds when it comes to AI.

Anyway…

Futurehouse is building agent-based systems for science and health. The real value is in how they structure problem-solving. If your business depends on analysis, synthesis, or research, this is a model worth applying.

Sakana released a new model architecture that trades speed for reasoning. It’s designed to think in steps instead of jumping to conclusions. We’ll look at how this affects decision-making and automation inside your workflows.

Finally, Microsoft restricted internal use of DeepSeek because trust matters. We’ll break down what trust actually looks like when teams rely on AI daily and how to build policies that don’t slow you down or leave you exposed.

None of this is speculative. These are signals from companies doing the real work of making AI stable, useful, and built to last.

Finally, in the Cortex part of this issue, I’m sharing a framework for building better offers with AI (available to paying Cortex subscribers. If you see this on the website, you must log in for access).

Let’s break them all down for your business.

—Sam

IN TODAY’S ISSUE 🤖 

• How Businesses Are Structuring AI Rollouts

• The Futurehouse Blueprint

• Sakana’s Models Think In Steps

• Microsoft Just Banned Deepseek

• AI Could Make Humans The Second Smartest Species

• The Four Folds: A Lean Framework for Building Offers That Work

Let’s get into it.

How Businesses & Teams Are Making AI Stick

(This is a shorter version of the full issue, available to free subscribers and on the web. If you’re a Cortex subscriber, you get the full issue below—if you’re reading on the web, make sure you’re logged in. Cortex opens up once a month at the end of the month).

Most AI strategies are failing because the business never made space for it to work.

The models aren’t the problem. You most likely need fewer dead ends.

Two recent breakdowns, one on enterprise scaling, the other from a recruiter deep in the trenches, land on the same truth: 

The companies winning with AI are the ones building around it, not just bolting it on.

They’re changing how work gets done.

The biggest shift happens when companies centralize. Instead of putting AI into a separate innovation lab or hiring a few ML leads to “figure it out,” they embed it. Inside real teams. Tied to real outcomes.

Product managers are using models in early prototyping, not to build the thing, but to shape the question. 

Marketing teams are generating three campaign angles instead of one. 

Ops teams are pulling real answers from internal systems in seconds. 

Not because they were told to use AI. But because it works.

That shift comes from structure.

Every successful rollout starts small: 

One task, one owner, one reason to care. A prototype that saves ten minutes every day. A retrieval tool that keeps customer support from digging through Slack. Nothing flashy. Just useful.

And hiring? Same story.

Most teams still chase senior engineers with PhDs and zero product context. They hire for academic credibility, then wonder why nothing ships. 

The smarter move is hiring builders. The people who understand iteration, who write rough prompts, and who get things working by Friday, even if it’s messy.

You’re not looking for AI fluency. You want creative pressure. 

You want people who see an outcome and ask: ‘

“Could a model do this better than I can?”

That’s where scale comes from. Not in a PowerPoint deck. Manus and Genspark can create endless slide decks. Not from a board presentation. 

Scale comes from the middle of the team, solving problems faster than they used to.

If you’re running a team today, start with this: 

Find a high-volume task you hate. Pick one person to own it. Make AI part of the process without being a test or the pilot. Just make it something they’re accountable for delivering with.

That’s how you get AI to stick. Not with vision. But, funny enough, with friction.

Futurehouse Is Building a System That Works.

Futurehouse isn’t trying to make the next viral demo (which is why this is good). 

They’re trying to use AI to solve real scientific problems. That alone sets them apart from 90% of AI projects right now.

Their goal is big: 

Use agents to accelerate breakthroughs in health and science. 

But what actually makes their work valuable is the structure.

They’ve built a framework where AI agents are part of a tight loop between researchers, tools, and decision-making. 

Agents do what humans don’t have time for. They read the literature. Write code. Generate hypotheses. And the humans step in where nuance matters.

If you're running a business that relies on research, content, ops, or product development, the Futurehouse model isn’t out of reach. It’s just applied differently.

Imagine your team isn’t building agents to do everything. Just some things.

A content agency might use a research agent to pre-digest everything relevant to a new client vertical, pulling white papers, scraping sites, and summarizing positioning. The strategist then spends their time thinking, not Googling.

In less than a minute, I already got a rough draft modeled from ChatGPT:

If ChatGPT can do it with zero context, it can do it for your business.

Anyways, back to the use cases.

A B2B SaaS team could build an agent that monitors support tickets, looks for patterns in user feedback, and flags three feature requests that show up every week. That agent doesn’t make the roadmap, but it makes the roadmap better.

What Futurehouse gets right is that agents don’t replace expertise. They route it. They compress time. They shrink the space between question and answer.

And because they’ve committed to a single domain, science, they avoid the trap of trying to build one agent that does everything for everyone. That’s where most teams go wrong. They chase scale before specificity.

If you want agents to work inside your business, copy that. Start small, in one domain. Tight feedback loop. Clear scope. Let the agent own the grunt work, and give the humans better surface area to think.

That’s the Futurehouse blueprint. Not an agent lab. A thinking system that compounds.

Sakana’s Models Think Better

Speed has dominated AI development for years: 

Faster inference, faster response, faster output. 

But not everything in business benefits from speed. 

Sometimes what you need is depth.

That’s what Sakana is working on. 

Their new model architecture, called Continuous Thought Machines (CTMs), introduces a different mechanism: 

Instead of generating a single output in one pass, the model runs through multiple iterations of internal reasoning before producing a final result.

In plain terms, it thinks, checks itself, refines, and then responds.

This approach is more “brain-like.” And crucially, it's more aligned with how businesses operate when things matter. 

Nobody wants instant output on high-stakes questions. 

You want layered thinking. You want visibility into how an answer was reached. 

You want decisions that hold up, not just outputs that look right at a glance.

This shift has massive implications for AI products built for strategy, legal, financial modeling, and anything involving multi-step planning. 

If you’re using AI to generate outlines or summaries, you probably don’t care how it thinks. 

But if you’re asking a model to prioritize roadmap features or analyze a complex contract negotiation, you need it to reason across multiple steps.

And here’s the kicker: 

If you’re designing AI into your business, the way your team uses models now is almost certainly optimized for speed. That made sense in 2023. It makes less sense now.

It might be worth rethinking where you want speed, and where you need quality of thought instead.

This doesn’t mean dropping your current tools. It means looking at your use cases and asking: 

Which of these would get dramatically better if the model could pause, evaluate, and revise its answer before giving it to me?

If you're building a decision support tool, this is especially relevant. Customers will pay more for a slower system they can trust than a fast one they have to double-check. Sakana’s approach makes that tradeoff possible at the architecture level. The rest is on you.

When to Trust AI Tools and When to Cut Them Off

Microsoft banned employees from using DeepSeek last week. 

No scandal. No public breach. Just a quiet internal decision: this model doesn’t meet the bar.

It’s easy to look at that as a one-off, but it’s not.

This is a signal that trust is no longer a theoretical layer in AI usage. It’s operational. It determines which tools get embedded, which get locked down, and who’s allowed to use what.

And here’s the catch: most teams aren’t even tracking this. They’re still stuck on accuracy and latency. But in practice, trust is what defines whether a model belongs inside product, sales, HR, or nowhere at all.

That trust shows up in different ways:

• For leadership, it’s whether a tool’s data usage lines up with compliance.

• For product, it’s whether outputs can be explained or traced.

• For legal and finance, it’s whether the model does what it says it does—and nothing more.

• For the rest of the org, it’s whether people feel comfortable using it without asking twice.

Microsoft’s decision wasn’t about performance. It was about predictability and accountability. 

Two things most companies won’t think about until they’ve already rolled out a dozen AI tools across teams, all with different permissions, scopes, and terms of use.

If you haven’t written a policy yet, don’t overcomplicate it. Start with one question: 

What is each tool allowed to do? Be specific. If it’s just summarizing internal notes, great. Say that. If it’s used for outbound copywriting but not contract review, say that too.

You need to move beyond managing tools. Think of it like you’re managing capability boundaries and your team’s sense of what’s safe to trust.

When that trust gets fuzzy, tools get banned. Or worse, they get misused quietly until someone notices too late.

Microsoft made a clean call. Most companies won’t. Don’t be like most companies.

AI Could Make Humans The Second Smartest Species

What if the very intelligence that propelled us to the top of the food chain isn't unique to our biology, and something else is rapidly gaining ground, perhaps even surpassing us? 

What kind of future unfolds when the rules of life, as we've known them for millennia, begin to change because intelligence exists outside of flesh and blood? 

I’ve recently started posting videos again on my Youtube channel.

The first video is on exploring the potential for AI to make us the second smartest species (will it? Will it not?)

Now, funny thing about the Youtube algorithm…

If I link directly to the video and you only watch a minute or two, then pause or leave—my video will be demoted and Youtube won’t show it to more people. 

Crazy, right? 

It’s actually “better” for me if you go to Youtube, search and find it, then watch it. 

Because I’m in the process of “resurrecting” the channel, I’m going to ask for a favor:

Would you be willing to go to Youtube, input the search string below, and click and watch the video?

Here’s the search term to copy and paste into Youtube: 

Sam Woods AI Humans Second Smartest Species

It’s crazy that the most convenient thing to do (link directly to the video) could be the most harmful thing for the video. 

Anyway, it would mean the world to me if you searched, found, and watched the video.

I promise it’s worth ~10 minutes of your time. I think you’ll find it interesting.

A few discussions I’m having thinking through in the video:

• AI's Exponential Progression vs. Human Mastery.

• Intelligence Reimagined: Problem Solving vs. Human Mimicry.

• The Consciousness Myth or Fact?

• AI's Non-Conscious But Still Effective.

• The Singularity and Human-AI Integration.

• The Evolutionary Fork in the Road.

Let me know what you think. Hit the like button. Subscribe if you want to.

(If you’re a Cortex subscriber, the full, expanded issue is below).

There’s a big difference between adding AI and operating inside an AI environment. 

That’s what this week pointed to.

It’s about how your business and team handles coordination, speed, trust, depth, and responsibility, even when some of that work is now synthetic.

Hiring shifts. Workflows bend. Expectations get weird. One team runs ten times faster. Another gets cautious and waits for a process that doesn’t exist yet.

And then someone asks the right question: 

“What exactly do we expect this tool to do?”

That’s where it starts to settle. Where teams stop chasing novelty and start building around clarity. Who owns what, what’s automated, what needs a person, where the feedback goes, and which tools actually belong in the room.

If there’s one thing to take away this week, it’s this: 

AI gets more powerful when you shrink the scope. One tool, one job, one place to learn. Then build on it.

That’s how companies go from experimenting to actually evolving. Quietly, consistently, and without making a big deal about it.

Until next time,
Sam Woods
The Editor