Omkar Deshmukh

Building things.

[Omkar Deshmukh @omkarishere]

I always am. I sometimes think.

[Omkar Deshmukh @omkarishere]

@jukan05 Local AI doesn't need HBM though - which is higher margin and eats up wafer capacity. If less HBM needed, easier to oversupply DRAM.

[rohit @rohitrajan]

Cranks & Engines

We are now entering the 4th year after the ChatGPT moment. The first wave of AI applications is here, promising the future of work. But if you look closer, they’re built on a broken economic model. They are Cranks, and they’re a dead end.

This architecture delivers linear returns that cannot sustain exponential growth in tokens. Its structural flaw is that it trades today's activity for tomorrow's churn. The market is mistaking this for long-term validation. How do I know this? We saw it firsthand. We built a Crank. Let me explain.

First, we built the crank.

A Crank is an AI application that requires the user’s constant presence to get any output, typically through the prompt box. It shifts the main operating mode away from direct manipulation.

We got in at the beginning of the first wave. We launched into the slides market by combining some publishing software we had built with GPT-3. Prompt to presentation in seconds. Very strong organic growth, poor retention.

The main problem is that the work is disposable. The output is a generation, not built to last. The user doesn't have the same connection to the work as something made by hand, making it easy to throw away. Turning the Crank involves frequent discarding, restarting, refining — and leads to ****massive token wastage. The output has a half-life of zero, and the surface resets itself on each turn.

There are ways to mitigate this. The best expression of a Crank is the Generative Workspace. In this case, the workspace accumulates context that makes it easier to turn the Crank forward. But the user is still the central operator and the primary bottleneck to compounding. High burn, no margin, linear returns.

Second, the ergonomics are poor.

Turning the crank is tedious labor. Each turn, the user is validating for accuracy, pruning and checking against edge cases. This is not creative work; it is expensive, repetitive quality control. Without this, the crank doesn’t work.

This labor is non-compounding. Each turn resets the frame, placing burden back on the user and demanding their attention and consideration. This cycle is constant, high-friction, and a disaster condition for torque, or in other words, growth.

Third, the value needs to “Compound”.

The solution is not a better prompt box, it is an Engine — an AI application that runs autonomously to create persistent, compounding value. It is a structural and economic departure based on the single principle that the system must compound on its own.

Engines are the only answer to the token waste and labor tax. They shift the paradigm to always-on, high token throughput. In this model, the system must use tokens to work while the user is absent. In other words, it must auto-turn the crank. This is how Engines build a structural moat. It is fundamentally different than memory. The system is building its own knowledge base.

With Cranks, memory is a key barrier to switching. The lock-in is behavioral. Engines are always running, building up knowledge over trillions of tokens. Switching away is non-trivial and expensive. The lock-in is economical.

The true breakthrough is in the ergonomics. The user is now consuming AI from the observation deck. Humans simply cannot be the primary consumer; we don’t read tokens that fast. Instead, the engine provides spot summaries on request. The user can consume on their own clock and step in as needed for high-level steering. This properly aligns the ergonomics with inference. It’s the necessary gearing that makes the Engine architecture mandatory.

Fourth, the architecture has to scale.

SaaS is the reference point and delivers 1x on your labor. Generative workspaces push that 2x. Engines deliver 100x leverage or more. This jump in magnitude isn’t bounded by headcount or user attention. It scales through compute and architecture. This is reflected in a different set of business benchmarks. Seats make sense in SaaS. Engines get measured on horsepower (token throughput) and efficiency (limiting down slop).

Engines are always on and constantly consuming tokens. They must process tokens thousands, if not millions of times faster than a manual Crank. Maybe more. In that sense the business develops vertically more akin to auto manufacturing. Value comes from improving the ergonomics of the chassis and making it easy to steer at high speed.

This model significantly changes the risk curve. The design makes leverage accessible to any team in the proverbial garage. One needs to focus solely on design and nailing the ergonomics. Once running, the architecture allows for trillions of tokens to flow through. This creates high market cap outcomes on a pound for pound basis.

Fifth, the most important reason.

We are flying to altitude through choppy air. Instead of looking to moderate lower, we need to push higher. Cranks by their design are turbulent. They create volatile, spiky demand, which fuels paranoia and bubble talk.

The AI build-out is fast and aggressive. It pressures energy demands upwards. The industry cannot justify trillions of dollars in infrastructure investment if the application layer is ultimately wasteful.

Engines are the necessary demand-side stabilizer. They process tokens smoothly and signal reliable, high-volume consumption. They are the critical foundation to absorb shocks in the supply chain. This encourages more research and education around token economics as more Engines come online. In turn, the capital keeps flowing steadily which opens up major arteries to spreading abundance wider and faster.

Conclusions.

The most successful Crank has introduced a new yardstick for growth. It’s tempting to replicate the playbook, but it is based on a flawed architecture. The incentives through the value chain from consumer to bare metal are misaligned. By their very design, Cranks cannot fulfill economic requirements to justify high-scale AI buildout.

Instead, one must start with the customer experience. The infrastructure demands extremely high token throughput. Humans don’t consume tokens that fast. Engines are the only path to stability in order to sustain a high growth rate. Most importantly, they create a favorable user experience and work towards limiting down the burden of labor. We are building the Orb to prove out this model. If you are building a Crank, it’s worth asking whether you should instead be in the Engine business. I think you should.

Rohit Rajan

November, 2025

[rohit @rohitrajan]

today we’re launching the Orb — a new kind of smart folder that learns from your files and helps you create new ones.

[Omkar Deshmukh @omkarishere]

Knowing < being

[Omkar Deshmukh @omkarishere]

A thing has value because we agree it has value (meme) or something deep within us compels us to value it (survival)

[Omkar Deshmukh @omkarishere]

What you do is who you are

[Omkar Deshmukh @omkarishere]

Everything is a memecoin.

[Omkar Deshmukh @omkarishere]

@alexalbert__ I can't really have long conversations with Claude. It keeps warning me to start a new thread. If you can fix that OR offer a way to one click summarize the current thread and start a new thread based on that, that would help. I often have valuable context in the current thread.

[Omkar Deshmukh @omkarishere]

The bigger the circle of the known, the larger the contact with the unknown.

[Siqi Chen @blader]

this christmas, i am humbly asking for your help to support dr. hankinson's research.

if not with $, i would be just as grateful for a repost or a like on this thread.

100% of your donation goes directly to funding the research efforts of Hankinson Lab: https://t.co/z8DcECGXGS

[Omkar Deshmukh @omkarishere]

Just because an event has low probability at the outset, doesn't mean it's not the likely outcome given the circumstances.

[Omkar Deshmukh @omkarishere]

@ShaanVP 65 35 is AA versus a flopped flush draw..not even runner runner

[Omkar Deshmukh @omkarishere]

@jmwind @vercel @typesense @supabase Supabase is great for db but for serving media I would highly suggest moving to Cloudflare R2. Cost could be ~10x less depending on the exact use case.

[Omkar Deshmukh @omkarishere]

@LongFormMath Thanks for sharing. To me, it has significance. Shows how much we don't yet know - and how hard it is to know something new. Humbling.

[Omkar Deshmukh @omkarishere]

universal basic insight

[Omkar Deshmukh @omkarishere]

@blakeandersonw how did you know what people want?

[Omkar Deshmukh @omkarishere]

flux is simultaneously both the most incredible and the most frustrating image model available right now

[Omkar Deshmukh @omkarishere]

@ShaanVP cash flow

[Omkar Deshmukh @omkarishere]

@antfu7 @VueUse useCursor to change cursor on an element