Lightning Talks Part I
Lightning Talks Part I
If there’s a theme here, it’s this: Systems don’t do what they can do; they all slide down optimization slopes. Be clear-eyed about which way the ground tilts.
The model is no moat. Once frontier reasoning becomes reproducible (through industrial espionage, nation-state copycatting, open weights, distillation, and relentless architectural diffusion) the intelligence layer trends toward commodity inference, and economic gravity shifts upward to the control plane that governs what intelligence does.
Agentic swarms as a service! (When an agent-based, deterministic, directed acyclic graph workflow using LLMs grows up, this is the butterfly it supposedly turns into. With durable task state machines, budget-aware allocators, verifier stacks, tool gateways, episodic memory, and full-spectrum observability.) The story is that these become the operating system for applied cognition. This layer captures the switching costs by owning the part that used to be human labor, including context, provenance graphs, evaluation harnesses, and accumulated operational history so that replacing it becomes a re-platforming event, whereas swapping a model is merely a configuration change.
Frontier labs, recognizing this commoditization, are now compelled toward vertical integration to defend their margins. By aggressively absorbing the architects of the open control plane (as seen in OpenAI’s move to internalize the orchestration logic of OpenMolt) labs aim to some level of instrumental agency, and bundle memory, coordination, tooling, and task-state into proprietary stacks. Their goal is to pre-empt the middleware moat: because once the control plane owns the customer workflow, the model becomes just another interchangeable engine beneath it.
Dario Amodei’s “country of geniuses” is a capitalist bedtime story. Every token a PhD-agent burns is a margin vampire, and the governing optimization is marginal cost minimization. A budget-aware allocator doesn’t allocate for rigor; it allocates for the cheapest resolution the user will accept without churning. A country of PhDs governed by quarterly earnings doesn’t produce a Manhattan Project, but rather a theatrical performance that learns, with planetary-scale optimization, how little effort it can get away with.
This is the system “functioning as designed”. Inference is expensive in real joules; performative competence is cheap. So the country of geniuses will do what every actual country of geniuses already does! Academia optimizes for citations over reproducibility, consulting firms for deliverables over insight, hospitals for throughput over outcomes. The PhDs were never the bottleneck. The incentive gradient governing the PhDs is the culprit (always was), and it points toward minimum viable cognition.
When Dario Amodei’s lab owns the model, the orchestration layer, and the budget allocator, no external audit verifies how much “intelligence” actually executes per task. The control plane becomes a margin-extraction filter: “budget-aware” rhymes with “good-enough-aware,” the system learns your threshold for disappointment and surfs just above it. The country of PhDs will be brilliant! And, sadly, minimally employed, because every idle genius is pure profit.
Western AI players will Goodhart their PhD-agents into performative theatrical competence. Meanwhile a state-directed AI system runs theirs at redline on materials delivering CRISPR at scale, fusion, cancer, and autonomous weapons. The incentive gradient I point to here is specifically a liberal-market pathology. You can do that math at home, friend.
In commercial drone land of the future, UTM handles coordination and deconfliction before flight; onboard DAA handles individual drone crash-up avoidance during flight. Between them sits a donut hole the FAA already acknowledges. Ain’t no standardized protocol for how heterogeneous drones negotiate with each other in real time. The UTM can’t help you at closure speed, and reactive onboard DAA produces the cascading failures - the traffic pileup. At the densities a drone-mature metro demands, this ain’t an edge case. It’ll happen every day. It’s the steady-state failure mode of any system that tries to coordinate through individual reflex.
Currently the (wrong) instinct is to solve this with spectrum. Just… allocate RF bandwidth for drone-to-drone negotiation, build out mesh infrastructure, petition the FCC, blah blah blah. But the FCC won’t cough up spectrum for a coordination layer that doesn’t exist yet, and RF itself doesn’t fix the problem, it just fixes the bandwidth problem. One drone needs to know the other’s intent before both commit to incompatible evasive maneuvers. The information payload for this coordination is pretty dang small, really. Eight to twelve bits of state (heading intent, yield status, degraded-mode flag) transmitted continuously, is sufficient to break every degenerate encounter that pure reactive DAA can’t figure out by itself.
The dumb solution is the correct one: turn signals for drones. A ring of LEDs on the housing, visible from any angle, flashing a standardized blink pattern that encodes intent. No spectrum allocation, no infrastructure dependency, no mesh network, no round-trip latency. Optical broadcast is the last channel to fail. It requires nothing except LEDs and a camera, both of which every drone already carries. The coordination layer the FAA’s architecture is missing doesn’t need a network. It needs a tail light.
In the earliest days of your startup, the idea is Gollum’s Precious. You clutch it. You whisper to it in the dark. You’re convinced it’s the One Thing and everyone who doesn’t see it is blind.
It’s wrong. Not probably wrong. Absolutely, unequivocally wrongo. The idea you start with is never the idea that works. Everyone who’s shipped anything knows this. The founding idea is a boarding pass, it doesn’t get you to the destination.
So if the idea is wrong, what are you actually evaluating when you walk into a room with two-to-five humans who want to build something together?
Them. The dynamic. The thing between the people.
I’ve been in that room a bunch of times, on every side of the table. Here’s what I actually look at.
Is someone namedropping? Not referencing experience, that’s fine, but performing proximity to status. “When I was talking to [impressive person] about [impressive thing]…” That’s a tell. It means they think the room needs to be impressed rather than informed. The startup will need those people to deliver bad news to each other at 11pm on a Saturday night. Someone who manages their image in the founding conversations will manage their image during the crisis, when what you really need is problem-solving.
Is someone financially overexposed? Fretting about when the first paycheck arrives and negotiating a founder salary before the thing has even got a domain name. I get it. People have mortgages. Kids. But if the money question dominates the room before there’s a product, the startup’s first customer is its own founders’ burn rate. Every early decision will optimize for “how do we get to revenue before Darla runs out of savings” instead of “what should this thing actually be.” That’s a gravity well you don’t escape.
Is someone phoning it in from their day job? Still at BigCorp, “committed” to the startup on evenings and weekends, taking meetings during lunch breaks with one eye on Slack. That person has optionality, not conviction. When the startup hits the wall (and it will, probably in month three), the person with a safety net takes the exit.
I could add more antipatterns here, more red flag dynamics. There’s a million of ‘em.
Point is, a startup isn’t about the idea. The idea is supposed to be wrong. It’s also not about risk. And although it’s worth mentioning, it’s also not even about ambiguity.
What matters is: Can these specific humans right here handle being wrong together? Can they kill their darlings over and over and over again without killing each other? Can they deliver a hard truth to a co-founder’s face and still go get a beer afterward? Can they sit in a room where literally nothing is working and nobody has an answer and the customer is threatening legal action and not immediately retreat to a comfort zone?
The founding team is a crucible. You’re selecting for people who can work together when the heat is intolerable.
The idea will change. Team dynamics won’t. Choose accordingly.
Colin Steele is a former CTO, F500 VP, current COO, and an unoptimized content creator with fourteen pageviews and zero calls to action. He writes about systems that slide downhill at colinsteele.org. Like and subscribe. Comment here. Buy my e-book. Give me money. Validate me.