Codex

Autonomous Business Incubator

April 4, 2018

Contents
1. Value Proposition
2. Introduction
2.1 Where existing platforms are lacking
2.2 The potential for autonomous businesses
3. A Platform for Autonomous Businesses
3.1 The simplest autonomous business structure
3.2 A first use case for AI: "Hypebot"
3.3 Adding complexity with commission-based work
3.4 Minimum viable product(s)
3.5 From direct democracy to fully algorithmic governance
4. The AI-driven VC Fund
4.1 Codex and the CDX Token
4.2 Token Launch Sequence
5. Appendix

1
1 Value Proposition

Codex is an incubator for autonomous businesses. Like traditional business incubators, it will
connect new ventures with resources such as market research, professional services, accounting
and financial management, and funding opportunities. Unlike in a traditional business incubator,
all ventures on Codex will launch with their own secure paywall and landing page, a Stellar
token representing ownership of business revenue, and a publicly auditable DAO (Decentralized
Autonomous Organization) that interacts with bots, freelancers, and investors to manage
business operations on behalf of token holders. We will provide a low-cost testbed for product
market research as well as a working low-touch sales and marketing channel for a wide range of
existing products and business models, and last-but-not-least, a novel avenue for AI research.
Codex itself will operate as a DAO seed accelerator as part of its role as an incubator.

● Producers​: an end-to-end business model for a product can be launched in minutes for a
dollar, with the majority of costs paid out of revenue, and services added only as the
revenue stream grows to support their costs. Alternatively, funding can be added early to
accelerate business growth.
● Freelancers​: with auditable DAOs, business structures will be transparent -- freelancers
can know ​what​ they would be working on and whether they would be paid fairly for their
contributions, and choose among ventures accordingly. With opportunities to work for
commission and equity, there could be outsize rewards for contributing to successful
products early.
● Customers​: with business structures transparent, buyers, too, will be able to see where
their money goes when they make a purchase (for example, are freelancers being paid
fairly?), and choose among products accordingly. They will also have the opportunity to
profit from products via affiliate links and equity tokens.
● Investors​: transparent business structures, early and novel opportunities, dividend-paying
business tokens (tradable on the Stellar decentralized exchange), and a dividend-paying
platform token (CDX) conferring governance rights
● Researchers​: a novel, complex task environment in which to test AIs

2 Introduction

2.1 Where existing platforms are lacking

2
People launching small online businesses typically need to go to (1) a platform like Kickstarter
or Indiegogo to handle crowdfunding (2) a platform like Amazon Mechanical Turk (AMT) or
Fiverr to handle tasks like web design (3) a platform like Stripe to handle payment processing.

Kickstarter and Indiegogo are great platforms for putting fledgling products before an audience
of would-be early adopters. However, they do not support equity crowdfunding natively, and
demand (sometimes formally) that products be already near-enough completion and well-enough
marketed prior to launch that investors will be willing to take on risk in exchange for only
non-monetary prospects (such as a copy of the finished product). Meanwhile, various ICO
launch platforms may provide a good solution in the case of equity-crowdfunding complex
startup businesses, but can be expensive overkill if the goal is to fund a simple venture such as an
ebook. Codex will enable $1 equity token creation.

Fiverr and AMT are a natural solution for businesses wishing to outsource freelance piecework
or "microwork", but lack support for payment in business equity or on a commission basis. Such
an arrangement may be desirable for businesses that lack initial funding, and for freelancers who
wish to leverage an eye for potential market successes. Similar arrangements are frequently made
informally among crypto startups and traders in the form of ​bounty programs ​in which token
equity is given in exchange for marketing tasks. Bounty0x has found some success at capturing
this particular market.

Another issue with Fiverr and AMT is transparency. Freelancers are frequently unaware of what,
exactly, they are working on. This is ethically dicey for a number of reasons, chiefly that
unethical businesses can access resources they wouldn't be able to access otherwise. and
freelancers can't always see if they're being paid fairly -- a group of workers may, unknowingly,
jointly handle the entirety of a business's product development and marketing while receiving
only a tiny fraction of its revenues, and that business may be funding God-knows-what. Codex
will offer the option of commission and equity-based payment for freelance work and make
business's identities and equity structures transparent to freelancers.

Lastly, Stripe and similar platforms can be quite expensive per transaction. Platforms like Stellar,
meanwhile, offer virtually feeless transactions, as well as the ability to cheaply codify revenue
splits at the point-of-sale; buyers may sign smart contracts that pay out a number of revenue
stakeholders, limited only by a function of the item purchase price and the number of actions that
must be taken on the Stellar network to carry out the contract (each action costs a very small fee,
under a tenth of a cent). Codex will use Stellar and ​SatoshiPay​ for these cheap, trustless
transactions, and to give autonomous algorithmic entities direct control of funds.

3
In sum, while existing solutions can be great at their specialized tasks, they each have significant
limitations. Further, they do not constitute an end-to-end business solution. A creator must sit in
the middle of these disparate business services and make many business decisions of their own
as they synthesize disparate data streams. These decisions are made all the more complex by the
varying profit models employed by each service -- some are subscription-based while some take
a cut of revenues, and others demand an up-front investment. For a wide range of online business
models, there is room for a simplified end-to-end solution, where each component service has
been built from the ground up with fluid interoperability, and ​autonomous​ operability, in mind.

2.2 The potential for autonomous businesses

Investor Peter Lynch said it's wise to ​"invest in businesses any idiot could run because someday
one will." [​source​]​ If so, one should codify one's business processes into a formal, step-by-step
procedure that anyone with half a brain can follow -- i.e., make it ​algorithmic​. It would seem that
many people are taking Lynch's advice these days, but perhaps only incidentally, for there is an
additional, and perhaps greater benefit to making business processes algorithmic: they become
amenable to AI solutions, some quite smart within their limited domains of expertise. For
example, JPMorgan Chase introduced a system to review commercial loan contracts, saving the
company hundreds of thousands of man hours. [​source​] Companies like Talla, Salesforce and
Optimizely offer any company AI solutions to problems like marketing analytics and enterprise
information retrieval. Great strides are being made in specific domains, but one place where
humans still outsmart AIs is in holistic business management, which entails executive
decision-making across a variety of interconnected business functions.

Currently no one would trust an AI to run their billion dollar company, but last year no one
would trust an AI to drive them around in a two-ton hunk of metal down the highway, and that
sentiment is quickly changing. Further, even ten years ago someone might trust an AI to drive
them in a go-kart around an empty bumper track at 10mph -- what's the worst that could happen?
Similarly, AIs can certainly be devised ​today​ to run end-to-end businesses given a sufficiently
restricted and low-stakes task environment. For a concrete example, take BitBarista, a "fully
autonomous corporation" in the form of a coffee machine that accepts Bitcoin for coffee and
pays Bitcoin to customers for completing maintenance tasks. [​source​] In the case of coffee, an AI
business is just another sales channel for a product that already has many -- what's the worst that
could happen?

The best that could happen is a massive increase in profitability and efficiency garnered, like any
other gains made by automation, by eliminating the need for a whole class of human labor -- in

4
this case, business management and corporate finance. Brotherlab Collective has been able to
reduce costs to clients while maintaining healthy rewards to employees and contractors by
eliminating the operating costs needed to support levels of management. Collectives like
Brotherlab achieve this by placing management and finance functions under direct-democratic
control, or else by operating in a distributed manner that does not require centralized control.
Collectives, however, are not always workable models. Where centralized, quick
decision-making is necessary for business operations, a collective may not be up to the task, but
AI offers a solution: cheap bots, agreed upon by the collective members, could support, or be
delegated the power to entirely handle, specialized business decision-making at a fast pace. Bots
can certainly be made auditable -- ​Botchain​ by Talla and the MIT Digital Media Lab's
TuringBox​ propose two different means of opening AI code to scrutiny by business stakeholders,
customers, or anyone whom its actions affect.

In an ideal scenario, a business collective managed entirely by bots, in which the majority of
labor is performed also by bots, is profitable and able to support both the members of the
collective and the bot owners who provide bots as services, while offering products and services
to customers at a fair price without negative externalities. We can make this ideal more possible
by (1) making business code auditable (2) making business code visible to everyone it will
affect, and the business itself answerable to these people and (3) making the business code
effective at generating profit. The first we aim to achieve right away with available technology.
The second is a complex social and political state of affairs we hope to promote by inviting the
right people to our platform and by making the right investment choices as a fund. The third is a
complex technological undertaking we hope to promote, again, by welcoming the right people to
our platform and making the right investment choices as a fund -- AIs themselves, formalized as
data services accepting micropayments, will be an ideal product to launch as autonomous
businesses on our platform

3 A Platform for Autonomous Businesses

3.1 The simplest autonomous business structure

A single product such as an event ticket, ebook, article, or video, or a serial periodical of these
items (such as a weekly blog) can be launched using no more than a paywall, a landing page, and
a straightforward affiliate marketing mechanism to help it find customers -- namely, offering a
percentage of revenue to the person who shared the link.

5
This model constitutes an autonomous business, if not an ​intelligent​ autonomous business, in that
incentives are so aligned as to require little-to-no oversight from a business operator, human or
otherwise; affiliate marketers are incentivized to share links in those places and with those
descriptions and advertisements they believe will maximize revenue through their links. Links to
the product spread virally, especially if buyers themselves are offered the opportunity to earn
back the cost of purchase by spreading affiliate links through their own channels, and especially,
of course, if the product is a good one -- it is clear how this model might work especially well for
special-interest event tickets and special-interest ebooks, items that already spread more easily
through word-of-mouth channels than through large marketplaces designed for general-interest
audiences.

3.2 A first use case for AI: "Hypebot"

To the above business model, a simple AI called "Hypebot" may be added to increase revenues.
Hypebot would look at (1) a description of the product provided by the creator -- a vector of tags
such as "event", "music", "electronic", "Boston", perhaps logically extended into a richer
description -- and (2) publicly available data on where people interested in similar things
congregate, and generate ​leads​ for affiliate marketers; for example, posting a link on
reddit.com/r/electronicmusic, or posting a QR link on a physical flier outside the House of Blues
in Boston. Hypebots’ advice would be checked by the affiliate marketers; they will follow it to
the extent that they believe it is good, and ignore it otherwise. All the while, Hypebot trains with
data collected on where links were actually posted and where they were or were not successful.
There is an inbuilt "data network effect" in this arrangement: the more products that launch with
Hypebot, the more useful it becomes. Hypebot may be formalized as its own autonomous
business, owned by its developers, which contracts with other businesses for micropayments.

Similar bots can be devised to handle a variety of other business tasks -- further automatable
tasks relevant to the above model include creating optimized landing pages (see ​The Grid​) and
ensuring an optimized pricing structure. For example, given an event date and a funding goal for
an event, a simple ​pricing bot​ may increase or decrease ticket prices and affiliate marketer
commissions as the event date approaches, depending on how far the event is from its funding
goal.

3.3 Adding complexity with commission-based work

Following this example, the next logical step is to allow contractors to contribute changes to the
business's landing page and flyers. In a commission-based model, contributors would earn a
reward, similar to link sharers, when a product is purchased through a flyer or landing page they

6
designed. Hypebot can be upgraded to manage the A/B testing of a number of contributed flyer
and landing page versions and select a winner to be used permanently. Part of its function would
be to look at click-through rates and decide if and when new flyer and landing page designs are
needed, as in when a product is underperforming relative to past, otherwise equivalent products
the bot has marketed.

Allowing work to be done in the form of bounties, for set cash or equity payments upon
completion, creates the need for a separate ​finance bot ​to decide (1) when to channel revenue
back to token holders, and when to add to the DAO's operational funds -- as in when Hypebot
determines that work on flyers or the landing page may be necessary to increase revenues.
Financebot may be employed by the human members of the DAO to give advice, or to make
decisions directly. With DAO's placed in control of wallets, the Financebot concept may be
extended to cover not only these intra-business "investments" in projects, but investments in
general, opening the door to autonomous mutual funds and hedge funds.

On top of marketing materials like flyers and landing pages, work may also be contributed to the
products or services offered by a DAO. Determinations as to current product quality and
product-market fit, and, in turn, whether it will be profitable to fund work on the product, would
be the jurisdiction of a ​Productbot​, or a product team that looks to a Productbot for advice. Such
a Productbot may also CDXde the work itself -- it may be feasible in some applications for a bot
to identify a multi-part task as being necessary, break the task down into atomic bounties, then
stitch the atomic results together and commit the final change to the product. An NLP-heavy
example is identifying the need for a new article on a specific topic by scanning newsfeeds,
identifying the topical sections that will be necessary to fully explain the topic, and issuing
bounties to write each section before finally composing them into a finished article for sale.

3.4 Minimum viable product(s)

Our MVP will provide the basic infrastructure necessary to create a ​digital product-oriented
business​ (in which human customers use SatoshiPay on a web page to purchase digital content),
an autonomous fund, or a ​data service​ -- a basic toolkit will include a readymade web app, data
service, and bot templates with standard APIs, and a suite of working bots to support blogs and
funds. On Day 1 we plan to enable users to create a business with a button click for $1.

Each ​digital product business​ will be given a SatoshiPay-enabled template web app deployed
from a special ​staging repository​ under Codex's control. Changes to both the app code and the
product inventory will be made by editing the code held in this repository. Codex will also
provide a standardized ​bounty board​ application to all businesses. Lastly, it will provide tokens
which (1) entitle the holder to an agreed-upon share of revenues generated through the business's
7
SatoshiPay channels and (2) grant liquid-democratic voting rights within the DAOs governance
structure. Every DAO on Codex will include immutable components that create a
liquid-democratic governance structure for tokenholders and make dividend payments according
to the agreed-upon schedule.

Fig 1. Basic Codex components and their relationships

DAOs will be deployed alongside apps and placed in direct and exclusive control of these apps,
and of the business's funds. DAOs will make only those payments and commit only those
changes to the business's apps (including its bounty board) which have been agreed upon
democratically. In particular, changes to the app code will be routed through a staging repository,
placed under the DAO's control by means of Codex as a proxy -- users will require on-platform
identities to commit changes to the code held in this repository, and Codex will only commit
work that, according to the DAO, the user had the voting power to commit alone, or else it will
open a vote for a period of time before rejecting the commit.

8
Voting power will be granted fluidly, on a category-by-category basis, to ensure that specialized
actions can be taken quickly by the people and bots best attuned to make them, without extensive
voting.

Fig 2. Plural democratic voting model. Everyone starts with voting power in each and every
category proportional to their token holdings. In this particular case, no bots (blue squares) are
token holders, and all actions taken by the DAO fall under only four departmental categories.

9
 Fig 3. The power distribution after token holders have delegated their votes. Actors now have
different voting power in different categories.

In the plural liquid-democratic model we will enforce, each user starts with the voting power
proportional to their token holdings, the same in each decision-making category within the DAO,
which can then be delegated to human or bot specialists who may not themselves be token
holders (Figures 2 and 3). Voting power can be delegated transitively (someone may re-delegate
voting power that was delegated to them) and can be revoked at any time.

All ​actions ​upon the business's wallet, codebase, and inventory will be formalized using an
action language​ to fall under a predefined set of decision-making categories -- for example, the
action "edit to frontpage.html" may be categorized under "marketing", such that only people or
bots with >20% voting power in the category of marketing are able to commit edits to the file.

10
These categories may simply be functional departments, as in Figures 2 and 3, or they may be
finer-grained: there may be distinct ​projects ​within departments (such as single products in the
product department), distinct ​components​ within projects, and finally distinct atomic actions
upon these components, to which voting power may be specifically delegated. Behind the scenes,
every single action is formalized as a proposal and called to a vote, with the person committing
the action ​de facto ​voting in favor of it. But if the actor on their own holds sufficient voting
power to pass the resolution, it simply passes, before the proposal is broadcast for formal voting,
and the action occurs seamlessly.

This model will work just as well for single humans or bots as it will for large organizations
comprised of a mix of humans and bots, and will allow organizational power to be transferred
from people to bots gradually as AI solutions gain trust.

Data service businesses​ will be structured in the same way as digital product businesses, but will
accept H2M and M2M Stellar micropayments rather than SatoshiPay payments and will have
standardized APIs -- all DAOs on Codex will be able to autonomously send each other payments
and, where the functionality has been enabled, employ each other's services. Bots will follow this
model.

Funds​ will simply control a set of wallets and make investment decisions on the basis of input
from humans as well as trading bots -- all DAOs on Codex will be able to make and accept offers
on the Stellar dex, thereby enabling them to hold each other's tokens (and any other Stellar
assets). We will add support for non-Stellar wallets over time to broaden funds' horizons beyond
Codex and Stellar.

Codex will release its own paid weekly blog (a ​digital product business​) and a Financebot,
Productbot, and Hypebot as ​data services​ specialized to manage blogs, all as autonomous
businesses. Codex itself will be structured as an autonomous business that provides services in
exchange for payments and makes investments (acts as a ​fund​). So, prior to public launch, we
will build working instances of each business type we aim to support. In the course of doing so,
we will hammer out the details of what works and what doesn't and create easily copyable
templates for the creators that come after us.

Lastly, we will provide a simple frontend for registering user accounts, launching new
businesses, and browsing businesses and bounty boards (see figures 4-6)

11
 Fig 1. Prototype of a company profile with token and revenue stats exposed

Fig 2. Prototype of a bounty board with built-in IDEs to work on relevant files

12
 Fig 3. Ranked public timeline of businesses for employer, investment, and product discovery

3.5 From direct democracy to algorithmic governance

Our first five businesses will start out with heavy dependence on human decision-making.
However, we will iteratively test and assign revocable trust to bots and gradually transition to
organization-wide trust in a unified AI solutions.

The digital product businesses explained above introduce the concept of issuing bounties for
software development work, namely making edits to the HTML and JS behind product landing
pages. It is worth noting explicitly that development bounties might also be issued for work on
software products such as data services and web applications. As mentioned previously, a
"Productbot" might be devised to propose, CDXde and compose bounty piecework in certain
contexts. Further, leaving aside immutable DAO components that have already been mentioned,

13
DAO code itself will be open to revision (in ​virtue ​of not being encoded entirely on a
blockchain). These revisions could hypothetically be managed by an analogous bot.
Ongoing work will be put toward developing a maximally expressive action language in which
to precisely formalize a broad range of software development tasks (beyond simple front-end
development tasks described as, say, "edit to frontpage.html"). This will allow AIs to more
effectively reason about work and take on responsibilities ranging from checking the validity of
predefined tasks, to defining entirely new tasks.

As more and more complex businesses prove amenable to this sort of end-to-end AI
management, we hope to close in on our end goal -- autonomous businesses that entirely manage
their own operations, including product development, and iteratively improve their own
operational algorithms by issuing bounties for self-defined coding tasks. This model would be
one step removed, in a sense, from the "Singularity" scenario, in which AIs are able to
effectively improve their own code. The model would cover all cases of self-improvement such
that the AI can (1) recognize that a task needs to be completed (2) accurately define it in
human-interpretable terms and (3) know whether or not it has been completed (though the AI
cannot complete the task itself). Though it's sci-fi today, as we begin to automate sufficiently
complex business structures and products, it will make sense to direct research into this area.

Finally, we hope to direct research into a "Venturebot" that is capable of CDXding Codex's own
investment decision-making. While programming such an AI to make ​profitable​ decisions is a
hard enough task on its own, doing so opens up an even more challenging possibility: an AI that
looks to ​public sentiment​ and makes decisions that are profitable not only to itself and its
shareholders, but also to the businesses that depend on it, the businesses and people who depend
on or are affected by these businesses in turn, and so on down the line, i.e., an AI that cares about
the entire ecosystem in which it is embedded. Many picture AI-driven businesses in a dystopian
light, but we contend that they can be made ​far more ethical ​than the average person placed in
the same position of power, given the right code.

4 An AI-driven VC fund

4.1 Codex and the CDX Token

Having 5% equity in all businesses that launch with it, Codex will collect 5% of all revenue
handled through its platform. CDX token holders are entitled to a share of these revenues when
(1) the accumulated revenue share has exceeded 10 lumens (2) 2 days have passed since the
previous disbursement and (3) they have answered all active ​survey questions​ -- we will issue

14
periodical surveys regarding issues at vote to enforce some minimal participation from any DAO
token holders who have decided to entirely delegate their voting power to other parties.

Token holders are entitled to a share, and are encouraged to share, in Codex's decision-making
through voting on proposals -- these may include on-platform funding proposals, proposals
within on-platform businesses in which Codex is a chief equity holder, and proposals to add
off-platform assets to Codex's portfolio, or to invest in platform development. Proposals at-vote
will be broadcast to the relevant voters and also show up in periodical surveys sent to all token
holders, and will be publicly accessible to ​all​ from our frontend, in order to gauge public
sentiment and check our behavior against it.

4.2 Token and Platform Launch Sequence

With the infrastructure in place to stream revenue to token holders and build basic autonomous
businesses, we will transfer control of the Codex fund to its DAO (DAO #1 on the Codex
platform). The development team will then begin work on our first four businesses -- the Codex
blog, and the blog Financebot, Productbot and Hypebot -- perhaps with funding from Codex. At
this point, Codex will be open for new autonomous businesses, which will in turn be open to
individual investment, and the dividends will flow to token holders. The roadmap will be as
follows:

Testnet development

Backend platform development

Frontend platform development

Launch Codex DAO

Testing and initial business development

Token pre-sale

15
Livenet launch

Token sale

Livenet development

R&D
complex apps and data services, advanced bots: Venturebot, Creatorbot, etc.; expanded action
language, PLD governance protocol, and complex bounties

Platform scaling and expansion

16