Jörg Doku

Building AI infrastructure. @Fluidstack, Former Head of ML @ @RunPod. @Google Brain, @Meta FAIR, Cruise. systems + inference. @Cornell

[Jörg Doku @JorgDoku]

Cadence is the clearest business-model analog for a new class of AI-era infrastructure companies: platforms that turn complex, high-stakes physical design work into repeatable, software-driven, compounding workflows.

Cadence for AI factories.
Cadence for self-driving science factories.
Cadence for AI-era power and grid infrastructure.

[Zeming Lin @ebetica]

We're in a place where codebases are no longer a thing that produces an action, but the product of a thing that produces an action.

In some sense LLMs compile intent into code. Just as compilers compile higher level abstractions into bytecode. It makes perfect sense that good engineering semantics for code now no longer work when some of it is compiled and some of it is hand written.

I think it's interesting to think about how to find the next paradigm of engineering practices. What happens when your compiler is super powerful, but slow, probabilistic, and expensive.

[Jörg Doku @JorgDoku]

Turns out fashion sense compounds too.

[Jörg Doku @JorgDoku]

Learn from the best. Forget the rest.

[Jörg Doku @JorgDoku]

@rachelrapp Nice!

[Jörg Doku @JorgDoku]

@jamesmishra @laran_hq Nice!

[Jörg Doku @JorgDoku]

@neildevani Very true!

[Jörg Doku @JorgDoku]

Of course all waves are great waves.

[Jörg Doku @JorgDoku]

It’s interesting how great engineering breakthroughs align the architecture with the binding constraint, often with remarkable economy of design. The same likely holds in markets: you can ride the right wave with the wrong business model.

[Alfred Lin @Alfred_Lin]

Don't confuse motion with progress. Many startups are very busy, but they aren't actually moving the needle. Velocity is speed with a specific direction.

[Jörg Doku @JorgDoku]

It's remarkable how critical presentation is for enterprise customers.

[Jörg Doku @JorgDoku]

@ariel_bendavid 😂

[Elon Musk @elonmusk]

@nikitabier @wholemars Thanks for everything

[Jörg Doku @JorgDoku]

Classic.

https://t.co/lkGEhMANLO

[Jörg Doku @JorgDoku]

@Javaughn12 well said!

[Jörg Doku @JorgDoku]

Very obsessed with everything Aswath Damodaran.

[Jörg Doku @JorgDoku]

Throwback Thursday!

[Jörg Doku @JorgDoku]

It’s fascinating to think that integrating new processors into autonomous vehicle fleets, including porting, optimizing, and validating the software stack, could once require teams of 100–1,000 engineers and years of work. Designing custom silicon could add another year or more. AI might eventually compress much of the chip design and software development work into months then weeks then eventually seconds.

[Jörg Doku @JorgDoku]

It feels like the key question is whether customization unlocks enough economic value, after accounting for migration, verification, maintenance, and operating costs, to justify a recurring purchase. So that way it's not purely a technical novelty.

[Jörg Doku @JorgDoku]

Programmatically generating software systems, including entire databases, from composable subsystems enables software to be tailored to a company’s specific workloads, performance requirements, and operational constraints, rather than requiring the company to adapt its workloads to the tradeoffs of a general-purpose product.

The prevalence of general-purpose software reflects, in part, the scarcity of specialized engineering talent and the high costs of development, validation, maintenance, and support. Vendors spread these costs across many customers, favoring architectures that serve a broad market rather than optimize for any one company’s requirements.

At sufficient scale, however, eliminating workload-specific bottlenecks can generate enough value to justify developing systems in-house. Meta’s development of Presto, a distributed SQL query engine for interactive analytics, and Google’s development of Spanner, a globally distributed database, illustrate this approach.

One of the biggest benefits of greater programming efficiency could therefore be to make architectural specialization economical far below hyperscaler scale. More companies could afford tailored databases, storage systems, search engines, caching layers, stream-processing systems, schedulers, and inference infrastructure. Customization could increasingly extend beyond configuring an existing product to choosing and composing the underlying algorithms and architectures, making capabilities once economical only for the largest companies accessible to much smaller ones.