Beyond LLMs: The Geometry of Strategy and the Coming Wave of AI-Driven Competitive Advantage
A completely new type of AI is coming.
Its unique ability to predict what it hasn’t seen will radically change your business and your investments.
The era of non-LLM AI is coming. There will be big winners and big losers.
The LLM Blind Spot
Unless someone’s technical, what they typically label as "AI" is almost certainly a Large Language Model or a diffusion model - but are those the best options?
LLMs are extraordinary: they write, reason, code, and converse.
Because of their dominance, you may not realize two things:
They have significant limitations,
And entirely different types of AI are emerging alongside them.
LLMs and diffusion models revolutionized how we create text, images, code, music. But a new class of model is emerging that could revolutionize how we decide and discover.
The impact will shape everything from where to position products to what molecular structures to pursue to cure diseases to how to allocate resources and even when to enter a market.
FlowBoost is one of the closest examples of implementation-ready AI we've seen, and we're going to discuss it here with the researchers who invented it in an upcoming podcast, so you can get ahead of the curve.
Introducing FlowBoost
FlowBoost is a new AI framework from researchers at Aarhus University and the Max Planck Institute for Mathematics in the Sciences in Leipzig.
On the surface, it solves difficult mathematical optimization problems such as packing spheres, arranging circles, minimizing discrepancies in point distributions.
It does this with remarkable efficiency: a 2-million-parameter model running on a single GPU in 1–2 hours, achieving results that match or beat Google's AlphaEvolve system which requires billions of parameters, cluster computing, and over 1,000 hours of compute time.
But the real story for us isn't geometry puzzles - it’s marketing & product strategy.
The real story is: how many of our most important and strategic product, marketing, and sales challenges will soon be solvable at scale because they can be most effectively solved when transformed into geometry problems.
Stated more simply: strategy is geometry
Think about these high-level strategic marketing problems:
Positioning a product in a new market
Determining the ideal feature set for an individual customer
Creating an individualized marketing mix
‘Under the hood’ all of these challenges are really complex geometry problems.
And if we now have an AI system that solves geometry problems at scale for low cost, we are looking at a new era in business strategy.
The Geometry of Strategy
I'd like to introduce a concept that I believe is novel: Geometric Strategy.
Geometric Strategy is the practice of transforming strategic decisions into optimization problems within constrained, multidimensional spaces.
This isn't a metaphor. It's a precise structural claim.
Every time you make a strategic decision, you're doing some version of the following:
Placing things: products, features, messages, investments, ad placements, team members
Inside a space with boundaries: budget limits, regulatory constraints, market size, competitive dynamics, time horizons, manufacturing capacity
Trying to optimize a score: revenue, market coverage, differentiation from competitors, return on investment, customer lifetime value
I believe that structuring objects, constraints, objective functions is exactly the mathematical structure - not approximately, not as a loose analogy.
And now AI models like FlowBoost are designed to solve those types of problems quickly and cheaply.
Here’s Why That’s Important
Strategic Positioning as Geometric Strategy
Consider how positioning works today.
A strategy team identifies market whitespace through qualitative research and frameworks. They produce a language-based narrative: "We are affordable, luxury urban furniture."
This is a semantic approach that hides a deeper geometric structure.
You can create a positioning solution (a positioning story) by leveraging semantic tools, but more powerful options open up if you expose the deeper geometry.
Your product exists in a space defined by every dimension customers care about: price, quality, aesthetics, sustainability, availability, brand perception, support experience. Your competitors occupy positions in this same space. Optimal positioning means finding the point that maximizes distance from competitors in the dimensions that matter most to your target customers, while staying within your cost and capability constraints.
We see glimpses of this strategy when a marketer creates a positioning diagram on a whiteboard. A Forrester or Gartner diagram is literally a 2D competitive space. But real competitive positioning involves potentially thousands of dimensions simultaneously. Examples include the digital twins that underlie Amazon's and Netflix's recommendation engines.
However, those models require lots of data and are only good at predicting in areas they've already "seen." In contrast, geometric systems have the potential to build a causal model of the world. As a result, they can make better predictions in new areas (aka out-of-distribution data), such as entering a new market.
Let’s look at that as a concrete example:
Imagine you're entering a new market.
You can and do complete research, but you and your team don’t have any actual data for your product in that market.
In addition, you have budget constraints, regulatory requirements per region, five established competitors with known positioning, and twelve customer segments with different willingness-to-pay profiles.
The question is: where do you place your product in this competitive landscape to maximize total addressable coverage while minimizing head-to-head competition with incumbents who have resource advantages?
A strategist solves this with frameworks, experience, and spatial intuition that maxes out at 3 to 6 dimensions.
A FlowBoost-class model could solve it across all dimensions simultaneously, respecting every constraint, and proposing configurations a human strategist literally cannot conceptualize.
Product Development as Geometric Strategy
Product development can be seen a sequencing-and-selection problem through a dynamic landscape. You have limited resources (constraints), a market that shifts over time (changing space), competitive responses to anticipate (adversarial dynamics), and you need to pick the path through feature-space that maximizes value at each stage while maintaining optionality for future moves.
This is a constrained optimization problem through a changing landscape exactly the problem class FlowBoost addresses.
And unlike static analysis, FlowBoost-class models have the potential to evaluate counterfactuals: if we build Feature A and our competitor responds with Feature X, what's our optimal next move? What if they respond with Feature Y instead?
What Geometric Strategy Is NOT
I don’t believe that Geometric Strategy needs to be a replacement for judgment, market intuition, customer empathy, or relationship-building.
Nor do I believe that the models that underlie Geometric Strategy should or will replace LLMs. There’s a reason that the brain is modular: different challenges require different or hybrid approaches.
However, models like FlowBoost can integrate dimensions of strategy: positioning, resource allocation, coverage, competitive distance, portfolio construction into rigorous and computable causal models at scale. The in-the-world dimensions still need humans. And LLMs are well-suited to the qualitative work interpreting customer language, generating messaging, synthesizing market narratives. The power comes from integration.
Here’s a point I think is important to call out: the "at scale" part matters enormously.
Rigorous optimization of a single strategic question is valuable.
Rigorous optimization of thousands of strategic questions per segment, per customer, per moment creates categorically new opportunities that weren't previously possible.
This level of abstract, causal intelligence has the potential to unlock everything from self-adapting, AI-First products to autonomous life science research to individually positioned messaging.
Before We Get Too Excited…
FlowBoost has been demonstrated on pure mathematics problems: packing, optimization, discrepancy minimization.
The business applications we're describing are inferences based on structural similarity, not tested products. But the structural similarity is precise, not metaphorical. The same formal framework configuration spaces, hard constraints, objective functions, high-dimensional non-convex landscapes describe both sphere-packing and competitive product positioning. The gap between proof and application is real but narrow.
The Opportunities
Opportunity 1: Democratizing Rigorous Strategic Planning
Large and mid-size companies can afford corporate strategy teams, McKinsey engagements, and dedicated analysts. Smaller businesses and startups operate on intuition and best guesses.
A system built on FlowBoost-class models has the potential to provide rigorous, individualized strategic optimization at a fraction of the cost and time. Not because FlowBoost does strategy but because it solves the underlying mathematical structure of strategic problems. Combined with an LLM for market intelligence, customer insight, and natural language interaction, a hybrid system could offer expert-grade geometric strategy to any business with data.
This isn't abstract. FlowBoost runs on a single GPU in hours. That's laptop-scale, not data-center-scale. When the computation barrier to rigorous optimization drops dramatically, the competitive implications are real: small businesses gain access to strategic tools that were previously the exclusive province of the well-resourced.
Opportunity 2: Strategy at the Speed and Resolution of the Individual Customer
Today, strategic decisions are slow, expensive, and centralized.
Positioning gets decided once and applied broadly. Pricing strategy updates quarterly. Product configurations target segments, not individuals.
FlowBoost-class models are fast enough and lightweight enough to change this fundamentally. The decision cycle could collapse from quarters to seconds, and the unit of strategy could shift from "market segment" to "individual customer."
Consider: a system powered by geometric strategy could ingest real-time market intelligence from a narrow customer segment, compute the optimal product positioning for that group, and generate individualized messaging through an LLM all continuously, adapting to customer receptivity and competitive moves.
This is the same shift that happened with pricing.
Teams used to set prices quarterly. Dynamic pricing didn't arrive because anyone got smarter about pricing theory, it arrived because computation got fast enough to run pricing at transaction speed. Geometric Strategy could do the same thing to positioning, product configuration, and market targeting.
Opportunity 3: Deeper Strategic Intelligence Counterfactuals and Competitive Modeling
Beyond speed and scale, FlowBoost-class models offer causal depth. Because they operate in continuous, high-dimensional spaces, they can simultaneously account for more factors than traditional analysis, e.g.: current market conditions, projected shifts, competitor response scenarios, resource trajectories, and regulatory changes.
This means the system doesn't just answer "what's our best positioning?" It can answer "what's our best positioning if Competitor A does X, and how does that change if they do Y instead, and what's our optimal sequence of moves over the next three quarters given both scenarios?" all while respecting every real-world constraint simultaneously.
Competitive simulation at this resolution has been possible in theory for decades. The computation to make it practical in real-time is what I believe is new.
Opportunity 4: Accelerating Foundational Research and Discovery
While this article focuses on strategic implications, it would be incomplete without noting FlowBoost's direct impact on research.
For life sciences startups, materials science labs, and R&D teams, FlowBoost-class models don't just improve decisions about existing products, they could unlock the discovery of entirely new structures, configurations, and molecular arrangements that humans wouldn't conceive of.
FlowBoost already beat Google's AlphaEvolve on circle packing problems finding configurations with higher total radii while using a thousandth of the computational resources. For a biotech startup or pharmaceutical researcher, that kind of capability applied to molecular or protein structure optimization could compress discovery timelines from years to months.
When combined with physical automation and recent advancements in automated laboratories, the potential to increase the research-to-result cycle time has the potential to shrink by many orders of magnitude. This opportunity means many more discoveries in less time in a way that completely alters the economics of modern research investment.
The Meta-Opportunity: Why this concept matters beyond FlowBoost
FlowBoost and similar models need not solve all of these problems themselves; what they represent, along with other emerging specialized models, is more important than any single application.
Marketers & Product Leaders must now think beyond the LLM
First, it demonstrates that what most business leaders think is possible with "AI" is just the very beginning.
New foundational models and hybrid systems offer the potential to not just replace or augment existing workflows, but create categorically new opportunities that weren't previously possible at any price.
Disruption without AGI
Second, it demonstrates that truly disruptive opportunities with massive profit implications do not require anything close to human-level AI.
FlowBoost has 2 million parameters. GPT-4 has over a trillion. FlowBoost beat Google's billion-parameter AlphaEvolve on specific problems while running on a single GPU. The AI that disrupts your industry probably won't look like ChatGPT. More than likely it'll be a small, specialized model that solves a specific class of problem orders of magnitude faster and cheaper than current methods.
This fact is true even without FlowBoost as we are still discovering that highly-connected LLMs have enormous untapped potential to create categorically novel applications. But FlowBoost-class models add precision, rigor, and breadth that LLMs alone cannot provide.
The Platform Is Already Emerging
The next paradigm of AI-enabled business where AI is the primary user interface, products automatically reconfigure around individual users, and systems connect at scale to accomplish user goals does not require FlowBoost. It requires LLMs, tool integration protocols like MCP, and a few other technologies that exist today or are arriving within months.
This revolution is already happening, FlowBoost-class models just supercharge it.
They add causality and optimality to what would otherwise be LLM-driven heuristic decisions. They add the capacity to evaluate counterfactuals. They add depth: more factors considered, more scenarios modeled, and more constraints respected. The combination of LLM flexibility with FlowBoost-class precision is more powerful than either alone.
So How Does FlowBoost Actually work?
Choose Your Own Adventure…
For the Strategist (CMO, Investor, Founder)
FlowBoost is a "design-first" AI model. Instead of generating language (like LLMs) or images (like diffusion models), it generates candidate solutions directly inside a continuous design space, the kind of space where solutions must satisfy hard constraints (no overlaps, stay within boundaries, respect budgets) while optimizing a score.
Think of it as a system that learns the "shape" of good solutions. It studies thousands of decent, valid solutions, learns the geometric patterns that make them work, and then proposes new configurations that respect all the rules while pushing toward higher scores. Crucially, it gets feedback on every proposal and improves; it doesn't just guess.
For the Technical Leader (CTO, Founder, Lead Architect)
Three things make FlowBoost fundamentally different from prior approaches:
It works in continuous space. Unlike systems that discretize problems into tokens or code, FlowBoost operates directly in the continuous configuration space. This preserves nuance; it doesn't have to round off or simplify the problem to work on it.
It builds constraints into the generation process. Rather than generating candidates and filtering out invalid ones, FlowBoost enforces geometric feasibility during generation through a technique called Geometry-Aware Sampling. This is dramatically more efficient. Imagine the difference between generating 1,000 random product configurations and throwing out the 990 that violate your constraints, versus generating 1,000 configurations that all respect constraints from the start.
It has a closed feedback loop. Each candidate the model generates gets scored against the objective, and that score directly improves the next generation of candidates. Prior systems either had no feedback (just pattern-matching on previous winners) or relied on massive LLMs as the creative engine. FlowBoost's closed loop converges in 1–10 iterations, compared to 100+ for open-loop alternatives.
For the Researcher (Scientist, R&D Lead)
FlowBoost combines three components into a unified pipeline:
Generator (Conditional Flow Matching): A learned velocity field transports random noise into structured candidate designs through an ODE integration process like a guided drafting process that starts rough and becomes precise. The model (~2M parameters, permutation-equivariant Transformer) learns to map from a simple prior distribution to the distribution of high-quality configurations.
Constraint Handling (Geometry-Aware Sampling): During sampling, flow integration is interleaved with Gauss-Newton projection onto the constraint manifold and proximal relaxation. This maintains approximate feasibility throughout the trajectory without requiring gradient information for constraints during training.
Closed-Loop Optimization (Reward-Guided CFM): Online fine-tuning via importance-weighted reward with a teacher-student consistency regularizer. The frozen teacher prevents generative collapse (mode collapse to a single high-scoring configuration) while reward weighting systematically shifts the distribution toward higher-objective regions. Convergence typically achieved in 1–3 boosting rounds.
The comparison that matters:
| Feature | PatternBoost | AlphaEvolve (Google) | FlowBoost |
|---|---|---|---|
| Parameters | ~10M | ~10B+ (frozen LLM) | ~2M |
| LLM Required | No | Yes (essential) | No |
| Iterations to converge | 10–100 | 10–1,000,000 | 1–10 |
| Compute | Single GPU, 10–100h | Cluster + API, 1000h+ | Single GPU, 1–2h |
| Loop structure | Open (no direct feedback) | Open (evolutionary selection) | Closed (reward-weighted updates) |
Thought Experiments for Your Role
For the CEO or Founder
What happens when your bootstrapped competitor can run positioning optimization that used to require McKinsey on a laptop, in an hour?
How do you prepare for a world where wide and deep strategic analysis is no longer a constrained resource advantage?
For the Investor
Which of your portfolio companies' strategic bottlenecks are actually configuration problems in disguise?
How does your thesis change when specialized AI models can solve constrained optimization problems 1,000x faster and cheaper than current methods and your competitors' portfolio companies are already adopting them?
How do your selection, allocation, and support processes change when categorically superior analysis is widely distributed?
For the Marketing and/or Product Leader
If your positioning could update per-customer in real time…
What would you stop doing today?
What data would you start collecting?
How would your messaging change?
What organizational structures would become obsolete?
What is possible if your products and elements of your services can become ‘strategic?’
For the Researcher
FlowBoost beat Google's AlphaEvolve on circle packing with a thousandth of the compute and no LLM dependency. It runs on commodity hardware.
What problem in your domain has the same mathematical structure objects in constrained spaces, an objective to optimize, a landscape too complex for brute-force search?
The Science of Innovation - The Creation of FlowBoost
While FlowBoost and similar models represent a truly novel and seismic shift in commerce, the insights that led to FlowBoost are useful in their own right.
As you know by now, the researchers did not set out to build a business strategy tool; they set out to solve a fundamental problem in mathematical and scientific discovery.
Their process reveals exactly why this architecture is so potent for next generation, AI-First applications.
The Pivot: Escaping the "Posterior Tail"
When the team set out, the goal was to find a non-LLM tool to search for extremal solutions to the absolute optimal, edge-case configurations in mathematics.
In many areas of science, finding these optimal solutions is akin to finding a needle in a high-dimensional haystack. The standard approach was to use a Deep Generative Model (DGM) to learn the probabilities of good solutions and sample from that data. However, the team realized a critical limitation: if you only sample from known, existing data (the "posterior tail"), you will never discover anything truly novel.
To create genuine breakthroughs whether in molecular structures or competitive market strategies the AI needs the capacity to push beyond known examples and explore "uncharted territory." This mandate became the philosophical foundation of FlowBoost.
> Insight: Causality as an escape from the in-distribution data trap.
The Physics Connection: From Diffusion to Flow Matching
Initially, Diffusion models (the technology mentioned earlier that’s behind AI image generators) were the state-of-the-art choice. However, the breakthrough came by looking at the intersection of AI and particle physics.
Baran, bringing extensive experience from deep generative modeling in particle physics (specifically the "LEGO" project at the Technical University of Munich), recognized that Flow Matching offered a superior path. By mapping the generative process as a continuous "flow" (a velocity field), the model became much faster to train, more precise, and easier to control than Diffusion. This cross-domain innovation applying the rigor of physics simulations to AI generation is what allows FlowBoost to operate so efficiently on a single GPU.
> Insight: Physics as a core first-principle strategic decision framework.
The Breakthrough: Eliminating 50% Waste with "GAS"
The most severe bottleneck in AI discovery is wasted compute. The team discovered that standard deep generative models were highly inefficient when dealing with hard constraints. If the AI was tasked with generating points inside a sphere, but accidentally generated them inside a surrounding cube, 40% to 50% of the samples were instantly invalid and had to be thrown away. In higher dimensions, this waste scales exponentially.
Their solution is arguably FlowBoost's most significant commercial feature: Geometry-Aware Sampling (GAS).
Instead of waiting until the end of the generation process to check if a solution broke the rules (a post-hoc correction), GAS controls the entire sampling process. It actively guides the generation in real-time, ensuring that every candidate stays within the complex boundaries of the problem. For business leaders, this is the equivalent of an AI that doesn't just brainstorm a thousand marketing strategies and throw away the ones that are over budget; it only imagines strategies that are financially viable from the first synapse.
> Insight: Filtering is an expensive decision framework
The AI-First Paradigm: Closing the Loop
The final piece of the puzzle directly mirrors the transition to an AI-First economy: moving from Open-Loop to Closed-Loop systems.
Traditional models (including many LLM workflows) are open-loop. They rely on fixed inputs, generate an answer, and stop. They cannot self-correct or monitor their own outputs against a goal.
FlowBoost introduced a closed-loop optimization system. It relies on continuous feedback to dynamically adjust its parameters in real-time. It generates a configuration, scores its success, and instantly feeds that reward signal back into the model to correct deviations and guide the next iteration deeper into uncharted, highly-optimized territory.
This is the exact mechanism required for the next generation of business technology: agentic systems that do not just execute a prompt, but autonomously iterate, adapt to market disturbances, and optimize for revenue goals in real-time.
Join Us
Hopefully, you’ve gotten new useful insights and ideas. But the fun doesn’t have to stop.
In mid-June, I will be interviewing one of the FlowBoost researchers.
The video will be dropped soon after that and this article will be updated with insights we uncover.
If you’d like to be notified when it drops or suggest questions you’d like answer, feel free to provide your email below:
We will not spam you. We’ll only use your email to contact you about this and related content.