Fab’s Friday Field Notes
We had one of our climbing friends over for dinner this week. Beautiful person, exceptional climber, but suffering right now for reasons that have nothing to do with climbing.
We didn’t offer advice. We offered dinner, conversation, and laughter. A place at the table. Sometimes the most profound thing you can do is just be there — not as an expert dispensing wisdom from a distance, but as a fellow human sharing proximity.
This is what real support looks like.
My body’s teaching me the same lesson. I’m climbing relatively well considering I’m not training properly — investigating an injury with doctors after M fell on my head two months ago. I heard that crack. Can’t unhear it. Today we agreed: diving into the neck before treating the peripheral issues.
The doctors want protocols. I want to listen to what my body’s actually telling me.
This weekend, if my neck and arm cooperate, I’m riding West Wood again. Three or four laps of the perimeter — one of my favourite Winchester trails. The terrain will tell me if I’m ready to undergo surgery. Not the diagnosis. Not the protocol. The actual roots, the real gradient, the feedback in my shoulder on each compression. I just want to know if I can get to Easter before surgery, otherwise I should plan it between January and February.
The Distance Problem
This week a few people asked me to evaluate some AI startups. We’re at the end of 2025 and when I expressed interest to invest, they sent a message about “seeding in Q1 2025.” Then I looked at their testimonials — the same names appearing on similar companies across the globe. Testimonials that don’t exist from people who don’t exist.
My old colleagues at Omnitel called these guys “sbarlafuss” — operators living in a completely separate dimension from reality. They’ve perfected the performance of expertise without ever touching the terrain.
Turns out this isn’t just a startup problem. It’s an expertise problem. And I found the perfect illustration buried in two contradictory AI reports.
Cartographers vs Climbers
In April 2025, a team of researchers published “AI-2027” — a detailed scenario planning exercise. Impressive credentials: one former OpenAI researcher, another ranking top on forecasting leaderboards. Their prediction: Western AI dominance through 2027. China falls six months behind due to chip restrictions and compute constraints.
Classic sbarlafuss logic: more chips, more data centres, more dollars — therefore, more progress. Resources determine winners. The implication is clear.
Three months later, Neptune.ai published something different. Not a scenario exercise — a report based on interviews with 13 teams across North America, Europe, and Asia who are actually training foundation models. Not theorising. Doing.
Their finding: Chinese teams achieved state-of-the-art performance requiring a fraction of the compute resources their Western competitors needed.
DeepSeek — operating under severe chip restrictions — trained their R1 model for roughly $6 million while matching models that cost billions. Moonshot’s Kimi K2 reportedly trained for around $4.6 million (though the CEO says this figure isn’t official), competing with frontier models.
The Neptune report’s conclusion: success comes from “building deep expertise across all aspects of training” — not from having the most GPUs. One team they interviewed trained foundation models on as few as two GPUs. The median hovers between 24 and 32.
Not exactly billion-dollar infrastructure.
What The Terrain Actually Shows
The AI-2027 authors were cartographers mapping from a comfortable distance. Brilliant credentials, extensive tabletop exercises, expert interviews — but fundamentally disconnected from the actual terrain.
The Chinese developers were climbers. Embedded in impossible constraints: export restrictions, limited chips, a fraction of the compute budget. So they focused on what they could control: engineering elegance, software efficiency, creative problem-solving. Meta forked Nvidia’s networking library for their Llama models. Moonshot optimised training paradigms. They couldn’t throw resources at problems, so they had to understand the problems deeply.
Small teams. Deep expertise. Embedded in constraints. Finding creative paths through impossible terrain.
Same pattern as those sbarlafuss startups versus the companies actually building things. Same pattern as protocols versus listening to your body. Same pattern as advice from a distance versus dinner with a suffering friend.
The scenario planners predicted what should happen given resource advantages. The market showed what actually happened despite resource constraints.
The Partnership Principle
Proximity beats prediction.
You don’t help a suffering climbing friend by analysing her situation from across the room. You invite her to dinner.
You don’t diagnose injuries by following generic protocols. You listen to your body’s specific signals — this shoulder, this neck, these trails.
You don’t win AI races by accumulating resources and extrapolating trendlines. You embed yourself in constraints and adapt to terrain as it reveals itself.
You don’t build generative partnerships by deploying capital and expertise from a distance. You embed yourself in their reality, understand their specific constraints, co-create solutions from proximity.
And you don’t evaluate startups by reading their testimonials. You look for people who’ve actually touched the terrain — and you can always tell the difference.
This is what “embedding over advising” means in practice.
The Neptune report’s most telling finding: when they asked successful teams for recommendations, the top answer was simple: “Nothing beats getting started and trying.”
Not “acquire more resources.” Not “follow best practices.” Not “hire experts to advise you.”
Get close. Get embedded. Read the terrain. Adapt.
That’s what I must do now with all the colleagues and excellent partners working with us and getting into the trenches and getting this stuff done.
Adventures Start at Home
This weekend, those laps of West Wood will tell me more about my injury than any diagnostic protocol. Friday or Sunday I’ll try to send those 7a too. The trail doesn’t care about predictions or resource advantages. It cares about reading terrain, adapting to reality, and finding the line that works now.
Just like our climbing friend doesn’t need advice — she needs dinner and presence. Just like those Chinese developers didn’t need more chips — they needed deeper engineering and closer attention to reality. Just like your business partners don’t need consultants — they need people embedded in their constraints, reading their specific terrain.
That’s also why I love the climbing analogy: you either pass and climb that route or you don’t. Binary. No shadows of grey. No sbarlafuss. The rock doesn’t care about your credentials.
When experts contradict, ride the trail.
When reports diverge, trust the market.
When in doubt, get closer to the ground truth.
The trails, the rock and the snow will still be there next week. But this weekend is time to find out if my body’s ready. Not because a protocol says so. Because the terrain will tell me.
See you on the other side of those laps.
— Fab