Why Causally — Hongkai He

Preface: people keep asking why I’m doing this. What follows is excerpted from an internal investor memo I wrote on October 8, 2025 — sensitive parts stripped, the rest reshaped for a public read.

What Causally is

Causally is an AI Studio, founded with two missions:

To explore AI-native organizational form, unlocking development and iteration speed 100x what current technical-management approaches deliver.
To enter vertical after vertical with AI-native product capability, reshaping how humans interact with information and how tasks get delivered.

I. What LLMs fundamentally change

First, cognition — the core means of production — has shifted from scarce to copyable and scalable. Like electricity and compute, and the capability is still climbing.

Second, AI has moved beyond being a tool. It can do the work and deliver value on its own. Put differently: the productive force that drives a process — today it’s cognition, tomorrow embodied AI will deliver in the physical world too — no longer originates only from humans. AI is the labor.

That cascades into a string of consequences:

Large swaths of executional jobs get replaced fast, the way textile workers and scribes/typists did in their time. Humans no longer need the underlying operating skill (we don’t need elegant penmanship or mental-arithmetic chops anymore).
Traditional org structures, division of labor, and workflow paradigms get rewritten. These shapes were forged by the industrial era’s need to organize human labor. When the cost structure of the key input changes, the form that grew around it has to evolve.
The bar for producing software keeps falling, approaching the bar for producing content today. But like content: low barrier to start, sustained high-quality output still takes an organized team. Anyone can post articles and videos, brilliant individuals always emerge — but the top accounts that consistently put out quality are still professional teams.
Traditional ARR and PMF logic may not last. A chunk of consumer software will become throwaway-product-like (10M in sales this year, vanished next year). Building durable moats and customer stickiness beyond the code itself becomes the long game — and a first-principles consideration when picking what to build.
Org scale shifts from headcount to revenue per employee (RPE). Using AI is a massive lever on labor productivity.
Org forms, business models, and product shapes previously bounded by human attention and brainpower become possible. Things like a “dedicated salesperson” for tens of millions of users, a “personal physician” for ordinary people, “phone negotiation” for millions of accounts.
AIGC content explodes, dominating the absolute majority of internet content. The volume and variety of information any individual can reach jumps another order of magnitude or more.

II. What changes slowly

But some things, over long timescales, change very slowly:

Human I/O changes slowly. People remain the rate-limiting step in sending, receiving, and controlling workflow.
Human nature is basically stable, and the contexts that generate demand are relatively constant (food, shelter, work, relationships, parenting, running a business).
Mature organizations struggle to reorganize around AI, and can’t extract AI’s full leverage in the short run. Structure is fossilized, inertia is massive, stakeholders constrain each other — major change gets blocked.
The VC–startup ecosystem has path dependency and inertia around big-money, big-market narratives.
Technology pushes individuals toward becoming “super-individuals,” but the value of team coordination is still huge — only the organizational shape changes. A team of supermen still beats a single superman.
Plenty of non-technology-frontier domains have real demand and entrenched traditional solutions, but the practitioners don’t lead with technology, and large enterprises plus VC-backed startups routinely ignore these markets.
Time remains the most fundamental scarce resource for any individual. A day is 24 hours, growth and lifespan have biological constraints. Human work-hours are the single largest cost line in economic activity, and they don’t scale or stretch.

III. The opportunity window that opens up

Put those two sets of facts together and the opening becomes clear.

Organizational forms that fully exercise AI productivity are deeply worth exploring, and they’re starting to emerge. AI-native organizations can run at 100x efficiency and high flexibility, outclassing traditional teams — and outclassing solo developers by deploying “human team + many AIs.”
In such a team, the division of labor shifts from skillset (mastering Python and iOS, or knowing Photoshop cold) to mindset and task ownership: every individual is a manager and collaborator of AI, capable of recruiting and coordinating multiple AIs end-to-end to deliver what they own.
AI tools will keep evolving at speed, and AI-native organizations must co-evolve with them.
AI’s accelerated iteration makes “build as you think” real, and trial-and-error becomes an advantage: small apps in two weeks, large apps in a month; RL-style “implement → feedback → improve → redeploy” cycles can compress to one or two days. The iteration efficiency will dramatically lift the hit rate of good products — versus the traditional software-engineering pattern of long research + long build.
Against a backdrop of information explosion plus technical enablement, the paradigm of how humans interact with information keeps evolving. Looking across the history of tech, we’ve already passed through three paradigms:

Directories (libraries, Yahoo, Sina) → Search (Google, LinkedIn) → Recommendation/matching (ByteDance, Uber)

AI is now pushing us inexorably toward a fourth — the assistant / curator paradigm:
- Intent understanding
- Context awareness
- Autonomy
- Flexible delivery of results and information
Vertical applications previously dismissed as “too small a market” now carry meaningful margin because costs collapsed. Business models previously infeasible become feasible thanks to the AI substrate (e.g., apps requiring heavy human interaction at low per-customer revenue but non-trivial aggregate volume).
Asymmetric competition: for 2B/2C niche apps whose core is search or recommendation, evaluate the value of redoing them on the fourth paradigm — especially in contexts large companies and VCs dismiss as “not big enough, not sexy enough.”
Cutting out the human layer: tasks that high-net-worth users already pay humans to do, and that can be delivered in the virtual world (booking travel, scheduling meetings, organizing personal calendars and to-dos, following up with people on work, household management, personal health advisory, family logistics — the whole personal-assistant space) all warrant evaluation for fourth-paradigm AI redos.
Competitive displacement: against expensive niche enterprise apps (> $1000/month) or low-TAM legacy categories, ship a better version at 1/10 to 1/100 the price and actively shrink the TAM to drive competitors out (capital doesn’t want to fund a single-application play whose TAM is shrinking). Or have AI do the work directly rather than selling tools per seat — deliver at the same price point, but strip out the human operator and their cost.
Billing paradigm shift: software moves from “usage-based tool subscription” to “outcome-based productivity pricing.” What you’re competing for is no longer just the software budget — it’s the labor budget.

IV. How we’re executing

Form a 3–5 person initial team in China. Take on some B2B custom projects early to refine the management model and team chemistry, while trialing 2–3 apps with potential for stable cash flow. In parallel, accumulate general-purpose components (sign-up, payments, marketing, customer service) and explore mechanisms for sourcing and validating demand.
Develop or recruit partner-level capability into gaps as we go — current focus is finding a co-founder strong in AI-native sales and growth.
Strictly speaking, this AI-native team itself is the first product Causally is building.
Once the model is stable, push 10–20 products per year (mostly software, possibly some smart hardware) into the US market. Data feedback decides whether each product gets doubled down on or shut down fast.
Each product’s fate gets decided by what reality says: the steady-runners get maintained; the high-potential ones get heavy investment and a dedicated team; the ones big enough to stand alone can spin out and raise independently.

V. It’s an organizational experiment

Causally’s operating philosophy is essentially real-world reinforcement learning, an organizational experiment: reality as the reward function, iterate fast.

At the start of a shift like this, we’re walking a road everyone is only beginning to explore. We’ll meet a lot of the unknown and a lot of the unexpected. It’s entirely possible that a year from now, most of the hypotheses and views in this piece will be overturned or revised — and we should hold that openness.

Not a prediction. An experiment.