Trust is All You Need — a thesis for investing at technology’s frontier

The complexity around the evolution and adoption of emerging technologies makes building and investing in companies at the frontier both exciting and difficult. The market for generative AI is a perfect candidate to examine, replete with competitive deals, eye-watering valuations, and complex dynamics that develop and change rapidly as new technical capabilities are commercialized in a matter of days or weeks rather than years.

Ultimately, our mandate at Golden Ventures is to understand emerging technologies on the basis of what, how, and where value will accrue over the long term. The excitement around Generative AI is palpable, but at the same time, it’s incredibly difficult for those assessing the technology (from an investment, industry or company-building standpoint) to ascertain the realization, durability and resilience of this value.

Over the past year, we’ve developed a core investment thesis that has helped guide our thinking as we contend with this “new normal”. In our view, it applies not just to AI, but to every new technology that surfaces.

Our thesis starts with an attempt to answer a question many are asking:

Is generative AI approaching a “peak of inflated expectations”?

Gartner’s Hype Cycle for Generative AI (2023)

Gartner certainly believes so, and I would tend to agree.

The market research company’s “Hype Cycle for Emerging Technologies” has gained popularity for its visualization of the lifecycle of nascent technologies from conception through maturity as a measure of market expectations (y-axis) vs. time (x-axis). The plot’s famous shape denotes inflated expectations early on in the lifecycle of a technology, which tend to fall sharply as a result of failures to deliver value via technical limitations and key unaddressed risks, before rising again — with its tail between its legs — delivering on a more achievable promise of value than at the outset.

So while Gartner has forebodingly placed Generative AI at the peak of inflated expectations as of their latest update in August 2023, how has this manifested in the market?

On one hand, AI’s “iPhone moment,” catalyzed by the launch of ChatGPT in November 2022, has produced a period of unbridled excitement for the potential value that AI can deliver across a myriad consumer and business applications, with 1/3 of large enterprises self-reporting the regular use of generative AI in at least 1 business function.

Notably, generative AI has supercharged product-led growth for a number of companies building software applications and platforms not necessarily designed with LLMs in mind. One key example in our portfolio is Composer, a platform for retail investors to simply build, test and execute complex trading strategies typically reserved for the largest HFs:

Before 2023, Composer’s workflow was accessible to more technical users, but not a broader audience; the platform worked more as a flow chart that produced programming instructions in a custom coding language.

By getting early access to the GPT-4 API, the company was able to — in 2 weeks — create a natural language interface that enables users to create, backtest and execute trading strategies within their platform by simply performing natural language queries and automating the entire workflow completely.

This allowed Composer to realize its vision of giving retail investors a seamless way to go from natural language query → trade generation → backtesting → execution; the company has grown 25% m/m and processed over $2B of trades since monetization began a year ago.

On the other hand, there’s a more complicated story playing out in the capital markets:

Private valuations for AI companies have outpaced their non-AI counterparts at every stage, largely due to a combination of excitement for generative AI paired with excess unallocated capital owing to reduced VC investment activity in 2023. My friend Ryan Shannon recently did a great write-up outlining the market forces contributing to the idiosyncratic state of the market.

This has also led to a 6x YoY increase in the level of VC funding for generative AI companies in 2023. But nearly $20B of this funding (83% of total value) can be attributed to astonishingly few companies — 7 to be exact:

1. OpenAI ($10B)
2. Anthropic ($6.85B)
3. Inflection ($1.3B)
4. Aleph Alpha ($500M)
5. Mistral ($527M)
6. Adept ($350M)
7. Cohere ($270M)

This high concentration of funding into mega-rounds for 7 companies has caused ripple effects in a market of over 389 generative AI deals and 2,500 AI deals in 2023. This phenomenon is reminiscent of the dynamics in scarcity-driven, illiquid markets for art, cars and watches, where price-setting for the entire market often happens in auction settings between 2 to 3 prospective buyers (the 2017 sale of the most expensive wristwatch — Paul Newman’s Rolex Daytona — and its immediate downstream price impact is a particularly salient case study).

At the same time, the majority of funding has been captured by foundation models, most of which besides OpenAI ($2B ARR by end of 2023) have little revenue to speak of.

Fantastic analysis from my friend Kelvin Mu on the dichotomy of funding vs. revenue in generative AI

To date, the only meaningful realized value has accrued to compute infrastructure that underpins LLM usage (read: NVIDIA). The link between funding and revenue still remains largely to be seen at all layers of the stack.

Lastly, understanding sustainability of long-term value is complex and not yet broadly understood. New narratives are constantly emerging as a result of the hastening path from research to commercialization. When product releases and updates from a single company like OpenAI regularly send shockwaves across the market that potentially put business models and entire categories of applications and tools at risk, it can be a humbling experience that causes us to rethink:

1. What are truly defensible business moats?
2. Do technical advances benefit incumbents or new entrants?

What is becoming increasingly clear is that in many cases, AI can enhance an application’s value proposition and product-market fit, but it doesn’t broadly create it.

Confidence vs. Competence

Now, back to the Hype Cycle. There’s a reason it looks familiar: it’s essentially a market-scale representation of the Dunning-Kruger Effect, a Psych 101 favorite that maps the evolution of confidence in one’s own abilities in a given domain as they gain competence. The saying “I know that I know nothing” is derived from Plato’s account of Socrates’ wisdom, and it’s no wonder that this represents the core ethos of the Socratic method — you need to be humble(d) to learn.

A couple observations from comparing the two curves:

1. Gartner’s “expectations” y-axis is analogous to Dunning-Kruger’s “confidence” y-axis, but refers to the market’s confidence in a given technology, rather than confidence in oneself
2. Dunning-Kruger uses “competence” as its x-axis, while Gartner uses “time” as its x-axis. Time is used as a proxy for the progression of a market’s competence in implementing or realizing value from a given technology

To gain more insight into the hype cycle phenomenon, we actually need to understand the interplay between a market’s confidence in adopting a technology and its competence in implementing a given technology. These are both dependent variables (y-axis) plotted against time (x-axis).

We know what confidence looks like against time, but how should we plot competence? One of the leading theories of how innovative knowledge proliferates through a market is Everett Rogers’ Diffusion of Innovations model from 1962:

Rogers’ “Diffusion of Innovations” model plotting adoption (blue) against market share (yellow)

Rogers postulated that ideas and technology spread via information flows that mirror the spread of an epidemic, a sigmoid or S-curve (in yellow) — initially slow as innovators and early adopters start to use the technology, then accelerates rapidly as the majority catch on, and finally plateaus as adoption reaches saturation. Based on the 1/3 large enterprise adoption rate, the market for Generative AI is somewhere between early adoption and early majority.

Trust is the market value of innovation

The interplay of confidence and competence in the market for innovation is the basis for value creation with respect to a given technology. In other words, these two variables dictate the market value of innovation, which can be conceptualized as “trust”. Trust is the foundational ingredient that facilitates exchanges of value between two unrelated parties, which is especially true in the market for innovation, where trust is the prerequisite to the adoption of a new technology.

Trust is constrained in the following ways:

1. Confidence: buyers / consumers must ascribe significant value to technology adoption; sellers market the technology to convince buyers they can access this value
2. Competence: sellers must be able to deliver the promised value to buyers

In building a combined view of “confidence” and “competence” derived in the last section, we can build a richer understanding of Gartner’s Hype Cycle that I’m calling the “trust curve”:

The Trust Curve, a superposition of Gartner’s Hype Cycle and Rogers’ Diffusion of Innovations Model

This chart illustrates the value created by a given technology over time as a function of confidence (black) and competence (purple), yielding a resulting trust curve (blue).

There are three key insights from the chart that give us the roadmap for sustainable value creation via trust:

Realization → Durability → Resilience

Realization of value:

• The “trust curve” (blue) is defined as the minimum of confidence and competence at any given time. The area under the curve (green) represents the realized value of a given technology over time
• High market confidence early on in a technology’s lifecycle yields significant demand and initial value creation, but roadblocks to value sustainability arise as key issues and risks become identified

Durability of value:

• The areas (shades of red) between the curves represent value erosion due to over- or under-confidence
• Overconfidence leads to unsustainable value creation due to unmet expectations, while underconfidence precludes adoption and leaves potential value unrealized

Resilience of value:

• As a technology matures, the plateau of sustainability can only be reached as confidence and competence align
• Any significant drops in confidence must be recoverable at the risk of all future value being lost

Based on this chart, we can formalize the relationship between trust and value as:

Trust = min (confidence, competence) = market value of innovation

In further examining each part of the value roadmap (alongside the narrative of AI’s evolution from “Classical AI” to generative AI), 3 “pillars” of trust emerge…

Realization of value → Accessibility

Market value can only be created by a technology insofar as it’s adopted. But how can you convince people there’s value in adopting a technology in the first place?

The answer, made clear over the past year, is that demand follows accessibility.

For a given technology, buyers or end users have domain or task-specific knowledge about what they want to accomplish but limited knowledge on the inner workings of the technology used to accomplish them. Sellers, in turn, recognize the pain points that can be resolved through the use of technology.

Accessibility creates and capitalizes on a shared surface area of knowledge (known-knowns) between buyers and sellers to catalyze demand — a mutually understood narrative of how value can be delivered through a technology’s implementation.

Prior to generative AI, “Classical AI” operated largely outside of the realm of public consciousness. These systems operated largely in the background, supporting processes and decision-making. They were specialized for specific tasks like pattern recognition, data processing, and executing predefined instructions, without directly generating new content or ideas. Their functions were crucial yet mostly invisible to the average consumer, encapsulating tools like predictive algorithms, chatbots, and automation processes.

The accessibility boom catalyzed by generative AI tools such as ChatGPT has transformed AI from a backend optimizer to a front-facing creator, with the ability generate new content by learning the patterns and structure of their input training data. An explosion in consumer and business adoption has ensued, with human feedback intelligently leveraged to fine-tune models at speed and scale through the embedding of RLHF into chat interfaces.

But adoption is only sustainable on the basis that a given technology makes good on its promises.

Durability of Value → Reliability

Though accessibility is a key catalyst for building market demand, the question that follows is: once people decide to show up, how can you ensure longer-term retention?

Retention follows reliability, which tackles the knowledge gaps and information asymmetries (known-unknowns) owing to market and technical dynamics that may not be immediately apparent to buyers or users, but which must be anticipated and managed by sellers to retain demand over time with a long-standing value proposition.

The importance of reliability in building sustainable value can be understood by inspecting the initial divergence of confidence and competence early on in a technology’s lifecycle:

• When a technology is adopted, it’s expected to work as anticipated under real-world conditions with few surprises. Yet overconfidence is inevitable early in the lifecycle of a technology, leading to an unsustainable “peak of inflated expectations” as confidence far exceeds competence
• This divergence in confidence and competence can mean that value created during this period is ultimately unsustainable, leading to a precipitous fall into the “trough of disillusionment”, where trust is broken and confidence reaches all-time lows (e.g. early implementations fail to meet expectations, and previously unknown technical roadblocks / risks arise)
• Worse yet, this dynamic can lead to an extended period of underconfidence, which leaves unrealized value on the table, as potential buyers sit on the sidelines, despite significant improvements in the technology

In a macro sense, we’re talking about building consistently performant systems that can scale with adoption, complete with the necessary guardrails to defend against key risks and points of failure. But the micro definition applies in equal measure with respect to competition — durability also means ensuring that there is a defensible moat to the value provided by a single company.

The emergence of ChatGPT has catalyzed an explosion in demand for AI across business and society writ large. Ever greater promises of value are correlated with offloading human cognition to computation. As AI encroaches on more critical applications, so does the need for guardrails across key areas such as performance, scalability, security and privacy — which were hard enough to pinpoint in simpler times.

Breaches of trust that have tempered market confidence include ChatGPT’s large-scale data breach in mid-2023 (100,000+ accounts compromised), and reliance on “hallucinated” output in federal court cases. Coincidentally, yet-unsolved “data quality” issues have hampered enterprise LLM implementations, spawning significant research and proposed solutions for grounding AI output — recent candidates include advancements in model fine-tuning, the emergence of LLM security as a category, and the integration of RAG pipelines and specialized knowledge graphs.

At the same time, existing incumbents and upstarts have found themselves needing to redefine the concept of business defensibility as the playing field is leveled. When advanced AI capabilities are available to all, what gives a value proposition staying power?

Resilience of value → Accountability

The ethos of disruptive technologies is that the state of play could fundamentally change at any time. This can be more of a bug than a feature given that transformative technologies are often associated with negative externalities and tail risks such as black swan events; these so-called “unknown-unknowns” have the potential to destroy value much more easily than it can be created.

As a given technology matures, “the plateau of sustainability”, where sustainable value persists over time, can only be reached when there are minimal gaps between confidence and competence in the market. For this to occur, two things need to be true:

• Both buyers and sellers need to maintain a mutually beneficial relationship as long as possible in order to maximize collective value
• Breaches of trust cannot be irrecoverable

This puts the onus on innovators to build with resilience in mind, which is driven principally by accountability.

The need for accountability is in part an acknowledgment of Murphy’s Law — things will always go wrong, and to varying, sometimes unforeseeable, degrees of impact. This makes it essential to identify and hold the right parties accountable and incentivize them to act in good faith, to ensure that all market participants are well-protected.

Beyond the guardrails required to satisfy buyer and end user expectations of the direct applications of generative AI, its wider social impact is just beginning to become felt.

Already, thorny attribution and ownership issues around IP have begun to be prosecuted, with a specific focus on the data used to train LLMs — NYT has sued OpenAI over its use of copyrighted content, while Anthropic has been sued by Universal Music Group over lyric distribution. The current models for attribution and ownership have little utility at the current speed and scale of AI.

And yet, all of this could be superseded by the spectre of AI alignment and its broader implications on control and safety. The risk isn’t merely about minor misunderstandings or errors but about the potential for AI to pursue objectives that, while logically derived from their programming, diverge from human interests and safety. While the alignment discussion tends to veer to a philosophical debate regarding existential risk, the emergence of high-quality deepfakes across image, video and audio formats has already led to devastating spearphishing attacks and instances of electoral fraud, highlighting the dearth of solutions to authentication and verification.

The 3 Pillars of Trust

We now have a way of neatly mapping sustainable value creation to three underlying pillars of trust, as follows:

Trust = f (accessibility, reliability, accountability)

These can be visualized as a hierarchy, whereby everything ultimately rolls up to value creation:

Accessibility is the direct link to value realization — it’s what ultimately connects technology to buyers and users. But accessibility is truly the tip of the iceberg; the remaining pillars (reliability and accountability) are not typically perceptible to the demand-side of the market, but are critical to ensuring that any value realized is ultimately sustainable.

These pillars of trust are the foundation on which we are building our understanding and mapping the value of emerging technologies and the companies that create and leverage them. They give us a roadmap that illuminates what we need to focus on from an investment perspective, and is a framework we believe is equally important to the builders bringing emerging technologies to market.

With regards to frontier technologies like generative AI, it’s helped us to prioritize areas of interest from a technical research, vertical market and business model perspective, which we’re excited to share and further enrich through ongoing engagement with the global community of company builders, academics, industry veterans and investors.

What’s Next?

Over the coming months, we plan to enrich and build on this foundation by publishing content focused on the application of the thesis and its underlying pillars across 4 categories:

1. Memos for our new and existing investments in frontier technology companies and their wider impact
2. Overviews of leading-edge technical research and development becoming commercialized (in collaboration with AMII and other leading AI research institutions)
3. Explorations of specific markets significantly impacted by AI (including discussions with portfolio and ecosystem companies)
4. Cross-disciplinary explorations (read: armchair philosophy) and other special interest topics