Notes from 'High Output Management'

"First, everything happens faster. Second, anything that can be done will be done, if not by you then by someone else."

https://www.amazon.com/High-Output-Management-Andrew-Grove/dp/0679762884

Notes taken from Andy Grove's High Output Management, originally published in 1983:

Page xv (Introduction): The motto I'm advocating is "Let chaos reign, then rein in chaos"

You could almost extract value out of each word in this sentence; 'Let' as in - you are in control and allowing this chaos to reign, 'chaos' as in - things can devolve into absolute disorder and yet you are still in control (allowing it) and on and on. 

Page xvii (Introduction): You need to plan the way a fire department plans. It cannot anticipate where the next fire will be, so it has to shape an energetic and efficient team that is capable of responding to the unanticipated as well as to any ordinary event. 

Here we see the benefits of building a dynamic and flexible business versus a defensible business. Some investors ask/evaluate how the current team/product etc. can be defended against would be attackers. However, as elegantly outlined in the above quote, it's more important to build flexible teams and products that can actively respond to competitive threats.

Nassim Nicholas Taleb (author of The Black Swan and Fooled by Randomness) calls a company (or any other entity) that follows this idea antifragile. The human immune system is an example of an antifragile system: one that improves with every new threat [1].

Page 16: You should guard against overreacting. This you can do by pairing indicators, so that together both effect and counter-effect are measured. 

Page 60: Delegation without follow-through is abdication.

Page 104: You should attempt to determine your customers' expectations and their perception of your performance.

Page 189: [On giving feedback] Level, listen and leave yourself out...The purpose of the review is not to cleanse your system...but to improve his/her performance. 

Page 194: Moving from blaming others to assuming responsibility constitutes an emotional step, while the move from assuming responsibility to finding the solution is an intellectual one.

The Startup Genome

Hidden within the magisterial work, The Gene by Siddhartha Mukherjee was this incredible quasi-equation about the interplay between the code of humanity, the genotype and the physical representation of the code, the phenotype. Basically it is not a one to one mapping, code != reality. This is true since the genotype is influenced by the environment the person is in and by random chance, hence:

phenotype = genotype + environment + triggers + chance

It seems plausible to port this into the realm of early stage companies and can support the discussion around using quantitative tools for private investment. Here we can think of the phenotype representing the eventual outcome of the venture, the genotype the attributes of the company, the environment the competitive and general endogenous financial landscape and the triggers and chance representing their usual meanings.

Building machine learning tools for venture investment selection really only uses the genotype of the deal (and perhaps the environment). This allows us to see the inherent limitations of using quantitative tools in private investment (as Chamath Palihapitiya recently quipped on twitter, to paraphrase: it's hard to use excel to predict the future) since the phenotype is made up of triggers and chance. 

But this hasn't stopped medicine from turbo-charging genomics to do what we can to eradicate disease so shouldn't we do what we can (build databases, ML models, graph theory etc.) to at least understand the startup genome as best we can so to remove as much uncertainty as possible? Obviously I think the answer is yes, which is why we've built and deployed ML models to support our investment decisions. 

Activity in Deep Learning Specialized Hardware

Exponential growth in acronyms: CPU/GPU/TPU/IPU/DPU/ASIC

"If you don't want to boil a small lake, you might want specialized hardware" - Yann LeCun

Best practice hardware implementation of deep learning seems to be using (Nvidia) GPUs for training, but for inference the GPUs tend not to be efficient enough for probable applications. Incumbents and startups alike see this as an opportunity to build specialized inference based hardware (although some are also attempting to unseat Nvidia as leader in training hardware). Below is an aggregation of new hardware approaches for DL.

Incumbents

Alphabet: 

Difficult to even list Google as an incumbent but they have moved into the space. Google built the Tensor Processing Unit TPU "several years ago" originally for internal use. It has been used to power Google tools like Inbox smart reply, street view and voice recognition. The two original engineers that led the development of Google's TPU has since left the company and created Groq, Inc. (more below). 

IBM:

Building a "brain-inspired computer", apparently over the course of the previous 6 years, called TrueNorth. It breaks with the 70 year old Von-Neumann traditional architecture, literally designed from the ground up. TrueNorth is again seemingly purpose built for inference, with a research paper claiming the chip is able to perform inference at 6,000 frames per second per Watt (better than the Nvidia Tesla P4 GPU at 160 fps/Watt). I haven't really heard much buzz around TrueNorth, it remains to be seen how much of a differentiation/impact it can have in adoption.

Intel:

Recently acquired Nervana Systems for a reported $400m+. Nervana is a "full-stack solution that is optimized from algorithms down to silicon to solve machine learning problems at scale." In terms of hardware, Nervana is building Lake Crest, a.k.a the Nervana Engine. Lake Crest is an ASIC custom designed for deep learning (to be released "in 2017"). It quotes interestingly a 10x increase in training speeds. Intel presumably paid a nice price for Nervana since they don't want to miss out on the ML wave (as pundits noted they did on the global emergence of mobile dominance). 

NVIDIA:

Some have claimed that Nvidia's dominance of the early machine learning revolution was a lucky break. In any case it seems to be evident that they have at least dominated the training market. It appears the new Pascal architecture (Tesla GPUs) are the architecture that Nvidia has deployed for inference. With so much activity happening in inference specific DL hardware, I'm not sure that Nvidia will have such a dominant position as they do with training. 

Qualcomm:

The ARM community page documents methods of "tweaking models and codes" to use Qualcomm's SnapDragon chips for efficient inference. Not much has been made (as far as I can tell) on Qualcomm's position as inference player. But it would be very interesting given their dominant mobile position. Yann LeCun did mention in March in a Wired article that Facebook is helping the chip maker "develop technologies related to machine learning". Qualcomm's position in the emerging landscape of training/inference focused hardware for machine learning seems to be an open question and outside relationships with Facebook and others, they don't seem to have made explicit, bold public moves. 

Startups

Not a lot of information is public on these startups since, justifiably for competitive reasons they deliberately keep information from the public eye. Nevertheless, below is a summary of the limited information available. 

Cerebras Systems:

Startup formed in late 2016 that now (June 2017) lists 32 employees on LinkedIn. Formed by serial entrepreneurs Andrew Feldman and Gary Lauterbach. Andrew Feldman previous Co-Founded SeaMicro with Gary Lauterbach (acquired by AMD for $355m) that built low power server technology. Cerebras raised a reported $25m - $30m from Benchmark and others. 

Deep Vision:

Building "ultra low power processor that is build from the ground up" to be deployed in mobile settings; IoT, drones, phones etc. Known investors include Comet Labs, Danhua Capital and Ideal Bulb ventures.

Graphcore:

Building an "IPU - Intelligence Processing Unit". Interestingly, Graphcore is a proponent of integrating training and inference on the same chip. Grapchore has raised $30m from Amadeus Capital, Samsung Strategy and Innovation Center and Foundation Capital among others and is based in the U.K.

Groq:

This is the interesting one. Groq was founded by Jonathan Ross and Douglas Wightman, both former Google engineers. Ross co-founded the TPU unit as part of his '20% time'. Ross then went on to be a 'Rapid Evaluator' at Google[X]. As the story goes, Chamath Palihapitiya spent one and a half years trying to find the secretive TPU group engineers within Google after hearing they had build their own hardware on a Google earnings call. Groq now boasts 8 of the 10 original engineers on the Google TPU project and raised a reported $10.3m from Chamath. Indeed he believes they could be building a "fundamental building block for the next generation of computing"...sign me up.

Wave Computing: 

"A plug and play solution for training neural networks in the data center". Wave's dataflow systems utilize "Dataflow Processing Units, DPUs". It is claimed on the Wave website that it delivers up to 1000x increase in training performance (speed). Wave has raised an undisclosed amount from Tallwood Venture Capital and Southern Cross Venture Partners.

Note: Cognite Ventures has put together a nicely comprehensive, and vertically categorized list of interesting machine learning startups. 

Machine Learning in Private Investment

Short post: The Establishment tools used to 'model' the value of a company (CAPM, DCF, etc.) use values (variance in returns, revenue, etc.) as a way to quantify and prescribe 'value' to a company. People are comfortable with it (despite Mandelbrot, Taleb et. al.) because the quantified values it uses are 'primary' to the underlying asset (company): revenue, variance of returns, growth rate etc.

I think in applying machine learning, we may look at 'secondary' values, so-called 'alternative data;' satellite images, metadata on companies, investor network graphs, hidden relationships. The private investment system is complex but there must be a mechanical, abstract representation of the system that we can use to improve the likelihood of investment success beyond 50-year-old establishment-level financial theory. 

Note: For a great primer on the usefulness of mechanical representations (models) in system representation and the application of data (machine learning) in solving problems read Peter Norvig's classic 'The Unreasonable Effectiveness of Data' or watch here. 

Seed Fund: Number of Portfolio Companies, Ownership and Fund Returns

Return On Invested Capital (ROIC) is a common metric used in describing the performance of an early stage fund. Common ROIC targets (Rt) for seed funds are 3.0x – 5.0x (higher the better). Over N portfolio companies at initial check size (assumed to be constant) i, this is just Rt.N.i (for example a $50m fund at 5.0x ROIC is $250m = 5.0x.50.1m [1]).

It is very difficult to attribute success outcomes to portfolio companies at the earliest investment stage (will this be $3b exit or $300m exit?) so we can look at the portfolio aggregate exit value, or the Implied Portfolio Market Cap on exit (IPMC). That way we don’t need to assign individual success outcomes to specific companies but just look at the overall market value of our portfolio at exit. This can be represented as:

When pr is 0 (no reserves for pro rata) IPMC generalizes to:

This states that adding more portfolio companies generates a higher required portfolio market cap on exit (all else being equal). Looking at an example, say we had a portfolio where our return target is 5.0x, we invest in 30 companies (N = 30), at first check of $10m post money (Vpost = $10m), not executing pro rata (d = 0.5) our required combined exited market cap is $3b. So we would need at least $3b in exited value from our 30 companies to achieve 3.0x ROIC (Note: if we had executed every pro rata opportunity, d = 1, and IPMC = $1.5b).

Indeed we can calculate the marginal increase in IPMC for every new portfolio company added (just set d = 0.5, Rt = 5.0x, Vpost = $10m). In this generalized scenario, adding another portfolio company (independent on initial check size) adds another $100m to the required portfolio market cap on exit [2].

We can use IPMC as a way to 'sanity check' the construction of our portfolio. Will we see/have access to enough high quality companies to generate a portfolio with $3b in combined exited value?

It is also helpful in the initial portfolio construction process. The below chart shows the IPMC for different targeted ROIC scenarios at different levels of aggregate ownership at exit. (based on a $50m fund). We can see that as ownership decreases, IPMC increases exponentially (this is also bad given the IPMC is tied to underlying portfolio company performance which is modeled as a power law). 

Notes:

[1] This is the simple case where there are no reserves for pro rata

[2] It should be noted that in the power law environment this marginal contribution model makes a little less sense since the fund returns are not made up of a little return from many companies but in fact the opposite: a large return from a small number of companies. I still believe this framework is of strong value; even if returns come from just one company, this can still be represented in this model.)

Optimizations and Likelihoods of Fund Returns

Returns from a VC fund are driven by the exit outcome of the underlying portfolio companies and the funds’ ownership of these companies at exit. For seed stage, it is common to model the company level returns using a power law. The first sentence above can be represented as the following (leaving the representation of the power law aside for the moment):

Ownership at exit can also be represented as:

As is clear, the fund returns are just the sum of contributions from exited portfolio companies; the exit value of company n multiplied by ownership at exit of company n. The interesting thing to note here is that if exit ownership is reduced linearly (either through smaller initial check size or greater pro rata based dilution) for returns to be the same, the size of the exit En (in $) must compensate and be higher. But this is exponentially less likely since we are in a power law environment. A large En (the ‘unicorn’ scenario) is the scarce element in this power law environment and (generally) not the ownership at seed.

Structuring a Seed Fund

There are factors that structurally limit the size of a seed VC fund like deal flow, allocations and check sizes and pro rata participation. Thoughts on each of these below, with the size of the fund at the end.

• Step 1: Deal Flow and First Check Sizes
• Step 2: Pro Rata
• Step 3: Fund Size

Step 1: Deal Flow and Ownership

Deal Flow and Portfolio Companies

• The number of companies you will have in your fund may be influenced by the following:
• Where/how you get your deal flow
• Whether you have a sector focus and how specific this is
• GP threshold of deal quality
• There are funds that will do 4 deals a month and others that will do 4 deals a year
  • The number of companies in your portfolio is the first limiting factor in fund construction

First Check Sizes

• Together with the post money valuation (pre-money valuation + money raised in round), the funds’ first check into a company will determine the funds’ initial ownership
• This is critical since it sets the upper limit on ownership in the company (assuming no super pro rata rights) and will likely be diluted (assuming the seed investors do not participate in late stage pro rata rounds even if they have pro rata in the first place)
• Lead investors in the seed round generally end up with 15% to 20% of the company
• Raising $1.5m on a $8m post with lead investor deploying $1.2 leaving $300k for angels
• You should also determine whether the fund will lead seed investments (sometimes taking a board seat, guiding/advising the company on strategy, supporting hiring, guiding the company to Series A through networks)
• This is a key differentiator for micro-VC seed stage funds since generally you do want an active lead investor getting the company to series A, if this is not your fund you will own significantly less than the lead and this changes fund economics (deploy $300k to a $8.0m post, own 3.75% versus 15%)

Step 2: Pro Rata

• But ownership in the company at exit is also heavily affected by whether the fund participates in pro rata rounds
• Many times seed investors will not be given pro rata rights and in some instances will not be provided these rights even if they were given at the time of seed (in some cases this can be positive if a very strong Series A investor is leading the A, therefore giving the company a higher potential exit opportunity)
  • An illustrative example (real raising amounts/valuations will change):
    • Let’s say the fund invests $1.0m as part of a $1.5m round at a $10m post valuation. Here the fund owns 10% of the company

• The table on the bottom left shows the subsequent funding rounds, how much the company is raising, how much the funds pro rata allocation would be to maintain 10% and the pre and post money valuations to work out dilution
• The bottom right table shows the effect of this dilution if the fund did not participate in any pro rata, if the fund Series A pro rata only, and if the fund did Series A and Series B
• Note: it is unlikely that the seed fund would be able to participate in Series C rounds since here the fund allocation would be $10m, also the Series A, B, C investors would be aggressive on gaining as much ownership as they can)
• Dilution calculation: If the fund did not participate in the Series A:
• Fund owns 10% of the pre-money $30m, $3m
• But company raises $5m cash, so we now own $3m in a $35m company ($3m/$35m = 8.6%)
• From the table, we see taking no pro rata we get diluted to 4.5% ownership, while participating through the B we get diluted to 5.9%
• In a billion dollar exit this represents a $14.5m difference
• We have effectively exchanged $2.5m to retain 1.4%
• If we did not invest in the A or B, we would have $2.5m to deploy to other seed companies
• At $1m per check this would be 2.5 more companies
• To make up this $14.5m through 2.5 other seed checks (assuming 1 would become a winner), we would need that company to realize an exit of at least $330m
• Math as follows: In that other seed company our 10% ($1m into a $10m post) would be diluted to 4.5%, this 4.5% would have to cover $14.5m, 4.4%*$330 = $14.5
• So if we did not participate in the A or B pro rata rounds we would need 1 out of the next 2.5 companies to realize an exit of $330m+ in order for this scenario to be better than participating in the Series A, Series B pro rata rounds (assuming you knew at Series A if it is likely to be successful)
• Having a $1b+ and $300m+ realized exit in 4 companies may be difficult given the power law in seed stage returns

Step 3: Fund Size

So now putting it all together we can generate the size of the fund that fits with our strategy

Number of companies:

• Say we could identify 1 company per month that crosses the threshold of quality (sector, team, market etc.) that’s 12 companies per year with a 2 year investment period on a micro-VC fund that’s 24, say 25 companies
• At $1m per check that’s $25m in primary investments
• Assuming we have pro rata in all companies, and 60% of companies survive to Series A that’s 15 companies of which the fund will choose to participate in 10 of these companies’ Series A, that’s $5m (assuming we own 10% and they raise $5m at A)
• We already know that in some instances we will not be able to participate our full pro rata amount given deal dynamics, so this seems to be upper bound
• Say half of these 10 companies survive to series B, that’s 5 companies and the fund would participate in 3 of them, that’s $6m (assuming we still own 10% and they each raise $20m)
• So total fund size through 25 primary checks at $1m each over 2 years, and Series A pro rata in 10 companies at $5m and Series B pro rata in 3 companies at $6m we get total fund size of ~$40m

Returns:

• 6.0x ROIC on a $40m fund is $240m
• In 25 companies, say 5 are winners at blended 5.5% ownership this implies success case should be $900m
• This seems unreasonable
• Assume 1 company runs to $5b, at 5.5% ownership this would return $275

 

Fund Construction Summary:

• $40m fund, approximately 60% to primary seed checks, 40% reserved for pro rata
• 25 companies at seed, participate in 10 Series A rounds and 3 Series B
• Will generate 6.0x ROIC with 5.5% ownership (diluted after Series B) with 1 out of 25 companies that exits at $5b (3.0x ROIC with 1 out of 25 companies exiting at $2b)
• Even if fund could not maintain (or does not get pro rata rights) will be diluted to 4.5%, of $5b exit this is still ~6.0x ROIC

Will need to prove:

• High quality deal flow
• Ability to write $1m seed checks into $1.5m seed round
• Ability to maintain pro rata through Series A

Human Learning vs. Machine Learning

Machines may undergo extremely accelerated learning cycles given the inherent inter-contentedness of their network. By this I mean when a human learns the optimal strategy for a particular event it is a single point of learning affecting that one individual. When a machine learns the optimal strategy it will be seamlessly distributed to every other machine in the network. 

There are very clear, inherent limitations in human learning. You need to "teach" (verb) me your understanding. It takes time to get you from 0 to 1. Once one machine gets to 1, every machine has got to 1. 

Imagine every human benefiting from the collective intelligence of every other human on the planet every day for the entire history of humanity. No rote learning, repetition or time delay through new cognition. Instant understanding, unlimited progress. 

The A3 Thesis: Automation, Augmentation and Autonomy

At CSC our goal is to identify and support companies that will experience significant increase in demand over the next 10 years (and then scale into the next 50). In support of this goal we have developed the A3 Thesis: Automation, Augmentation and Autonomy, which helps direct our investment decisions even at Seed. The A3 Thesis is deliberately broad and has applications in almost every industry vertical. It is focused on data generation, processing and performance enhancement for the industrial and commercial application of machine learning.

The A3 Thesis represents the large scale direction of technological innovation we believe will occur over the next 10+ years. In understanding how this technology will be adopted and deployed we rely heavily on Carlota Perez’s model: Technological Revolutions and Financial Capital, what Fred Wilson has called USV’s ‘bible’. Since we cannot see a deterministic future we attempt to learn from the past; the Perez model of technological deployment is based on the interplay of technology and capital in the last 5 technological revolutions.

It can be tempting to think current investment opportunities rely heavily on a third piece, namely timing. However, it’s more appropriate to look at the ‘window of opportunity’ as Bruce Gibney, Founding Partner at Founders Fund describes. Arguably Google’s window of opportunity has been open for the last 18 years. We focus more on transformational, thesis driven companies that present robust, long term, sustained value creation versus short term, exploratory and exhaustive applications of current technology.

A3: Direction

In The Rise and Fall of American Growth, author Robert Gordon proposes
 (in detail) that the technological advances of 1870 to 1970 (what he calls the ‘special century’) cannot be repeated and recent productivity declines will continue unabated. Gordon is not alone in questioning productivity stagnation. Morgan Stanley recently released a report showing U.S. and OECD productivity at or near decade level lows. The White House even referenced this debate in their recent report asking whether this is just the next chapter of automation or whether “AI will affect the economy differently than past waves of automation”.

Our distinction between the 3 A’s is an important one. Automation refers to labor replacement from machines for non-complex tasks traditionally completed by a human; it is low level value add. In the technological context, augmentation implies enhancement. It may also (and usually does) enable new solutions to be created for the complex problem being executed. Marc Andreessen described in an interview with Tim Berners Lee that with machine learning we’re seeing an “entirely new kind of product that wasn’t possible before”. This is where the value lies and forms the central theme of the A3 Thesis. It is what Joi Ito and the MIT Media Lab calls ‘extended intelligence’ or William Ross Ashby in 1954 called ‘intelligence amplification’; the deliberate combination of human and machine intelligence. Autonomy is still the most distant. It is complete human removal from a complex task while maintaining or (probably) enhancing performance.

Here at CSC we are what Gordon would call ‘techno-optimists’. We believe not only that the opportunities presented by the development and application of machine learning will influence and increase productivity and growth but also do so quicker than before given ready access to information and open data. Associate Professor of CS at Stanford Dr. Fei Fei Li recently remarked that advances in computer vision based on deep learning have the potential to set off a “Cambrian explosion” of innovation, Klaus Schwab, Founder and Executive Chairman of the World Economic Forum has already called machine learning the ‘Fourth Industrial Revolution’ and Head of AI at Baidu and the leader in deep learning applications Andrew Ng calls “AI the new Electricity”.

We believe augmentation (human performance enhancement through machine learning) is ready to influence almost every major industry in the industrialized economy. In 2012 Geoff Hinton’s team built neural nets that beat humans in the ImageNet Competition and DeepMind recently conquered the game Go. The value driver in machine learning applications is enhancingthe completion process of a task, be it time, cost, efficiency etc. We believe deliberate industrial and commercial applications will be key echoed by Satya Nadella and Microsoft who are building AI to solve the “most pressing problems of our society and our economy”.

A3: Deployment

We now turn to the deployment of the technology and the relationship between technological development and financial capital. As a primer on the Carlota Perez model, technological revolutions have 4 main phases; Irruption, Frenzy, (Crash) [1], Synergy and Maturity (for more resources on the Perez Model see [2]).

The Irruption phase (first) and Maturity phase (last) of the former revolution generally are concurrent. In Maturity there is a focus on M&A and oligopoly building and idle capital is searching for new opportunities even abroad (foreign investment from idle capital also occurs in the Frenzy phase, the difference being in maturity it is production seeking in Frenzy it is speculative).

In Irruption “huge successes, incredible rates of growth and even more incredible profit margins become potent magnet for still looming quantities of idle capital”. The difference in our current context (the new exploratory applications of the internet, on-demand etc.) is that the profit margins have not been anywhere near “attractive”. For this to be realized, it requires the installation and deployment of A3, specifically augmentation. Take Uber, once “autonomous” cars are widely deployed the incremental cost of the cars reduces to maintenance (of the hardware and software), improving the profit margin dramatically.

Frenzy brings financial innovations; new financial instruments (the CDS, the ‘private IPO’, secondary markets etc.) making money from money on idle capital and the expectation that all innovations will be equally successful (the exhaustive misappropriation of the technology).

So where are we today? This influx of idle capital and innovation of new financial instruments and increased M&A activity support the idea we are in the Maturity phase of however you characterize the previous 30 years. Marc Andreessen also identified the lack of new technology saying “my thesis is that we’re not in a tech bubble — we’re in a tech bust. Our problem isn’t too much technology or people being too excited about technology [Frenzy]. The problem is we don’t have nearly enough technology [Maturity]”.

We do believe however, that overlapping this we are now in the Installation phase of the machine learning revolution [3]. Deep learning models have been around for nearly 25 years. But the combination of GPU processing power and open access to data from the Internet means the machine learning/deep learning revolution has now begun to provide economic opportunities and just last month, Jen-Hsun Huang called the 2012 ImageNet competition the “Big Bang” of the GPU based deep learning revolution. So if the life of a VC fund is 10 years, what are the opportunities today that will experience an increase in demand (to the point of realizing transfer of value from enterprise or individual)?

Opportunities

There are three main elements we believe in successfully deploying the potential of machine learning:

1. Data Assets

2. Data Utilization

3. Industrial and Commercial Application

NVIDIA, the leader in GPU driven deep learning lists them in a similar way as Training, Datacenters and Devices. These three elements can be also read chronologically: to have a successful application of machine learning (in a specific industry for example) there needs to be effective processing and analysis on appropriately and robustly attribute-tagged data assets. The huge leveragability of machine learning lies in the fact that the same processes can be used to train models for any different industry application. For this we need robust, attribute tagged data assets. We believe this is where the most attractive opportunities lie currently; building the infrastructure of machine learning in the generation, classification, storage and processing of large data assets [5].

Applications are high beta; either that specific application has product market fit and is adopted or it still needs time to effectively find a willing market. We believe the appropriate time to invest in applications layer of machine learning is Post-Frenzy (the timing of which is extremely difficult to forecast). We are excited by companies that focus on data asset creation.

Thematic Investments

There are some opportunities that we see that fall outside this thesis. We believe given the right set of conditions these present compelling investment opportunities. Examples of out thematic investments includes those in Virtual Reality, Consumer focused companies, New Media and Marketplaces to name a few.

Constant Evolution

The A3 Thesis was created to support our investment decisions. It is our probabilistic view on the direction of technological innovation and will therefore evolve over the course of the fund through the discussions with founders. The evolution of the A3 Thesis is not due to a lack of conviction but rather our attempt out optimization, a kind of supervised learning for our own investment model. We are patient capital, prioritizing long term value creation and look forward to learning from everyone involved in creating and contributing to the deployment of the machine learning revolution.

info@cscventurecapital.com

[1] It’s clear to see how the Perez crash could be created. Schumpeterian ‘creative destruction’ of traditional employment will most likely lead to socio-political unrest and an even increased focus on augmentation opportunities (rightly and wrongly) from speculative capital. A recent report from Oxford University found that nearly 47% of all current employment in the U.S. is threatened by the application of machine learning and Larry Summers recently wrote in the Washington Post that he estimates nearly 1/3 of all men 25–54 in the U.S. could be out of work by mid-century. With the pace of augmentation increasing, the U.S. extending ZIRP driving more speculative, idle capital and productivity stagnating the Perez crash could in fact come sooner than expected.

[2] Carlota Perez Model primers: AVC post: http://avc.com/2015/02/the-carlota-perez-framework/, AVC post: http://avc.com/2011/10/required-reading-for-the-carlota-perez-interview/, Fred Wilson Carlota Perez Interview: https://www.youtube.com/watch?v=hmesHdCcXn4

[3] In fact, an argument could be made that the entire electronics revolution has formed the installation phase for the machine learning revolution. It has allowed us to create, store and process actionable data and develop systems to augment human limitations in solving complex problems.

[4] SoftBank recently announced the launch of a $100bn fund dedicated to technology investment in conjunction with the Saudi Arabian sovereign wealth fund and many public investment firms have been entering the private technology investment asset class.

[5] An analogy can be made here to the process of extracting commercial value from exploratory oil and gas drilling. The quality of the oilfield is unknown until after the exploratory drilling. If oil is found it is extracted, refined and sold to readily available consumers. In our machine learning context, data is the oilfield. But the major difference here is we have control over how ‘valuable’ that asset can be (how well attributes or features are tagged in the data). In effect we can create our own high quality oilfields. However, we do not have readily available consumers; this comes in the Synergy phase and is the applications layer of the technology. Ironically, there already companies applying machine learning to oilfield exploration.

Decision Making in a Low Information Environment:

Applications in Angel and Seed Investing

In December of 2000 the NASDAQ had fallen 51%. In an effort to help understand why, researchers from the Berkeley International Computer Science Institute and Rutgers authored a paper examining the effects of learning with limited information, focusing on strategic interactions “on the Internet.” However, these academic models can also help in understanding the decisions today between participants in another low information environment; angel and seed investors.

The paper examines the monopoly, synchronous and asynchronous Cournot and Stackelberg competition environments. We will consider all here (since it is difficult to label the ‘correct’ model, if there even is one). The model that is most applicable (potentially) to angel/seed investing is the asynchronous Stackelberg environment (indeed it is shown in the study that asynchronous environments tend towards Stackelberg competition. Here there is a leader (large VC’s participating in Series A, B), follower (Angel/Seed funds) structure where the leader knows ex-ante that the follower observes leader actions).

Lesson 1: Players are responsive and react quickly to changes in environment even if their current near-equilibrium payoffs are not affected

Since angel and seed investing is a low information environment, players may find it hard to attribute high payoff events with their corresponding investment actions. As a result actions that provided a high ‘equilibrium’ payoff state are potentially unknown so when they observe experimentation by other players they may perceive the other player to have more direct (higher payoff) information. The importance of building a strong negative feedback loop is examined later.

Lesson 2: Instead of independent probabilities of experimentation subjects appear to enter autocorrelated ‘experimentation phases’

This is a very surprising result and may provide some insight into how tech bubbles are formed around these ‘autocorrelated experimentation phases’. Autocorrelated because they follow a repeatable pattern in each returning occurrence of the experimentation phase.

Lesson 3: The greater variation a player can have in the payoffs of opponents the greater the inherent instability of the system where the opponent is more likely to experiment

It could be argued that venture capital follows a Stackelberg competitive environment; there is a leader and a follower and the leader knows ex ante the follower will respond to his or her actions. Here the leader is the institutional VC and the follower is the angel. The angel is incentivized to align with actions of the institutional VC since the angel’s investment is potentially dependent on future funding from a VC. In this system the variation is extreme from bankruptcy to IPO leading to, as the paper finds, greater instability and higher rates of experimentation.

Lesson 4: Players can be led to confuse experimentation by opponents with changes in underlying payoffs leading to a cascading effect of experimentation

Cascades of experimentation in the startup investing environment doesn’t sound good. But it does allow us to have greater insight into the competitive pressures leading to tech bubbles. One VC ‘experimenting’ in a new internet startup is observed by all others as a change in the underlying payoffs each could be expected to gain leading to cascading experimentation from all.

Understanding the competition in a low information environment could provide valuable insights into the creation and sustaining nature of tech bubbles. Since it is very difficult to connect high payoff events (10x investment) with the direct causes of this success (team, market, execution, tech?) players are incentivized to experiment (all the way to NASDAQ -75%.)

Structure for reviewing potential investment: Input is information on new investment (the pitch), B outcome of previous investment decisions, A is structure for filtering investments based on inputs (diligence etc.), output is investment decision

So what can we learn from this? In my view it highlights the value in building a robust negative feedback loop for investment decisions. Here we take as ‘Input’ the information on the current potential investment as well as the information on competitive experimentation (from other angels/VC’s). We also take ‘B’ the result of our previous investment decisions and feed both through ‘A’ our decision model for the current investment and arrive at ‘Output’ our investment decision (which the result will then be fed back to inform the next potential investment).

Without doubt this system is employed by a great number of successful angels and VC’s. Formalizing its structure can be beneficial so as to be balance ‘new’ competitive experimentation information with ones own investment theses and diligence structures (to avoid cascading experimentation).

Most importantly this formalized negative feedback loop allows the angel or VC to learn what decisions/theses/ideas/diligence allowed them to arrive at a successful payoff (and which didn’t). The issue here is the latency between input decision and output (payoff result) which should be the subject of another post since this one is almost TL;DR. Takeaway: Build formalized negative feedback loops for your investment decisions.

*There may be situations in the angel/seed application here that do not correlate exactly to theory (Cournot and Stackelberg competition) however this application is designed to illustrate concepts for discussion rather than add to the academic literature

Investment Opportunities in Artificial Intelligence

Disclaimer: Much of the following post will be proven wrong by the relentless judgement of time. Indeed, there may be something to learn from where we were wrong (in January of 2017). This post is designed as an exploration tool rather than a descriptor of fact.

There are numerous investment opportunities provided by the awakening of AI. Both in seed stage and 30 year plus public companies. This post is an attempt at structuring our current explorations and understandings of the investment opportunities afforded by the development of (current, narrow) AI. This forms the basis of our A3 Thesis.

The Three Pillars of AI

It looks like the earliest documented mention of the three pillars of AI was from Wired magazine in October 2014, these being:

·         Appropriate Algorithms

·         Data Availability

·         Computational Power

1.       Algorithms

Per Wired (and numerous academic journal articles) the algorithms that power much of the hyper AI tech of 2017 originated much earlier (the 50’s, the 80’s etc.) and have just been waiting. Geoff Hinton, now infamously sparked a resurgence in neural nets in the 2012 ImageNet competition with again now ubiquitous ‘deep learning’ approach. It seems to us that the commercial value in algorithm development may be in specialized use cases (outside traditional image recognition etc.) but it hard to imagine such opportunities today. Also much of this value lies in the prodigious minds of the human capital of the company or University; hard to monetize. It appears much of the value in AI algorithm development has been commoditized.

Opportunity: Large public companies (FANG’s); recruitment and retention of key human capital (large public tech may have the means necessary to provide utopian research environment for academic talent).

2.       Data

Incredible opportunities for early stage technology companies in data, specifically ‘data generative assets’. Any company that creates differentiated, deep, clean, useful data through sensors or incentives or any other means possible. Ginni Rometty has called data the new natural resource of our time. In fact, it is even better; companies can create their own data (effectively creating their own oil).

Opportunity: Startups: industrial sensors, new business models that incentivize people to track data, satellites for more pervasive, cleaner data, collaborative business models etc. are all attractive.  

3.       Computation Power

Given the capital intensity of developing deep learning-specialized GPUs we had originally thought the best investment opportunity here was with NVIDIA. Indeed, if you had invested at the beginning of 2016 it would have been a remarkable investment (+260% and best performed in the S&P 500). However, recently we came across Cerebras Systems, a startup run by a team who created SunMicro and sold it to AMD. Cerebras is funded by Benchmark proving that perhaps there is room for a proven team, specialized focus and deep pocketed, long term, committed investors. 

Opportunity: Market neutral NVDA/INTC; NVIDIA has been aggressive in its adoption of deep learning based AI even as early as 2014. Intel has missed the mobile revolution and risks being left behind in deep learning. Very hard to bet against Intel. More to come.