New to your Substack, thanks for the article, very educational. The challenge of developing software inside large companies has been true since day one. The inertia of the bureaucracy (needed to drive the commercial operation) vs the dynamism of creating new products (needed to have something new and valuable for customers to use) creates such a paradox that I'm not sure anyone has solved it.
IMHO, the best approach is to spin off the product development heroes, get them funding, and let them try to capture lightning in a bottle again. Let the main business focus on efficiently commercializing the products customers want, once the product people figure that out.
I think Google is making the right business choices, focusing on the commercialization of the capabilities, and being agnostic about what tech a customer wants to deploy. Even though the AI models are the most important & unique part of the stack, you need the whole stack to efficiently get a reliable answer/solution from the model.
The value to a user is in getting the right answer, not in using a cutting-edge model.
While I can agree that this article touches on very valid points for why senior leadership is leaving it leaves out the ethical dilemma some might have faced with Google essentially giving the Pentagon full access to use Gemini AI for "any lawful government purpose". Which is just vague enough to mean anything you want it to mean.
Spinning off product pioneers to preserve agility while running the core enterprise as a neutral compute broker sounds pragmatically sound on paper. Yet, it misdiagnoses the true failure mode. 🧩
Treating research exploration and commercial platform delivery as separable concerns is merely a degraded projection of an un-anchored optimization manifold failing its linear identifiability constraints. When model development optimizes for leaderboard autonomy and organizational legibility, it converges entirely on high-entropy benchmark-maxxing. It creates systems that ace abstract evaluations but crumble across forty-step deterministic API traversals. 📉
Meanwhile, reducing the platform layer to passive infrastructure commoditizes the very silicon advantage that should have anchored the entire ecosystem. The spectral gap between chaotic operational overhead and compounded enterprise value is bridged exclusively by unified onto-causal grounding. When an autonomous model mangles an enterprise query or hallucinates permission scopes, that runtime failure shouldn't sit in an isolated cloud ticket queue. It must serve as an immediate, non-invertible error signal flowing back into post-training token distributions and custom silicon compiler passes. ⚡
Fragmenting the stack across autonomous spin-offs simply externalizes coordination drag into venture market transaction costs. You end up repurchasing your own exported talent at ten-digit valuations while the core substrate remains starved of actionable telemetry. Real agency doesn't scale through administrative insulation or agnostic compute leasing. It scales when generative search and physical hardware execution operate within a singular, phase-locked feedback loop. 🏛️
If you sever the generator from the physical execution substrate, are you actually building a durable platform, or just subsidizing the ecosystem's compute bill while your own talent constructs your competitors? 🔬
New to your Substack, thanks for the article, very educational. The challenge of developing software inside large companies has been true since day one. The inertia of the bureaucracy (needed to drive the commercial operation) vs the dynamism of creating new products (needed to have something new and valuable for customers to use) creates such a paradox that I'm not sure anyone has solved it.
IMHO, the best approach is to spin off the product development heroes, get them funding, and let them try to capture lightning in a bottle again. Let the main business focus on efficiently commercializing the products customers want, once the product people figure that out.
I think Google is making the right business choices, focusing on the commercialization of the capabilities, and being agnostic about what tech a customer wants to deploy. Even though the AI models are the most important & unique part of the stack, you need the whole stack to efficiently get a reliable answer/solution from the model.
The value to a user is in getting the right answer, not in using a cutting-edge model.
yep
While I can agree that this article touches on very valid points for why senior leadership is leaving it leaves out the ethical dilemma some might have faced with Google essentially giving the Pentagon full access to use Gemini AI for "any lawful government purpose". Which is just vague enough to mean anything you want it to mean.
https://turntrout.com/why-i-left-google-deepmind#what-if-the-people-you-critique-were-saving-their-political-capital
https://www.theguardian.com/technology/2026/apr/28/google-classified-ai-deal-pentagon
I woyld be jumping for joy if this were a huge factor but I spoke to many people and I barely heard it brought up.
Is that an anthony smith reference I spy????
Yessir.
W
Spinning off product pioneers to preserve agility while running the core enterprise as a neutral compute broker sounds pragmatically sound on paper. Yet, it misdiagnoses the true failure mode. 🧩
Treating research exploration and commercial platform delivery as separable concerns is merely a degraded projection of an un-anchored optimization manifold failing its linear identifiability constraints. When model development optimizes for leaderboard autonomy and organizational legibility, it converges entirely on high-entropy benchmark-maxxing. It creates systems that ace abstract evaluations but crumble across forty-step deterministic API traversals. 📉
Meanwhile, reducing the platform layer to passive infrastructure commoditizes the very silicon advantage that should have anchored the entire ecosystem. The spectral gap between chaotic operational overhead and compounded enterprise value is bridged exclusively by unified onto-causal grounding. When an autonomous model mangles an enterprise query or hallucinates permission scopes, that runtime failure shouldn't sit in an isolated cloud ticket queue. It must serve as an immediate, non-invertible error signal flowing back into post-training token distributions and custom silicon compiler passes. ⚡
Fragmenting the stack across autonomous spin-offs simply externalizes coordination drag into venture market transaction costs. You end up repurchasing your own exported talent at ten-digit valuations while the core substrate remains starved of actionable telemetry. Real agency doesn't scale through administrative insulation or agnostic compute leasing. It scales when generative search and physical hardware execution operate within a singular, phase-locked feedback loop. 🏛️
If you sever the generator from the physical execution substrate, are you actually building a durable platform, or just subsidizing the ecosystem's compute bill while your own talent constructs your competitors? 🔬
(⌐■_■)b