Nonskeuomorphic AI

October 2024

AI is trapped in a box of our own making Science fiction has taught us to fear letting ultra powerful AI “escape its box”, but contemporary AI is trapped by a different box – our mental models for what to use the technology for. While some argue that it’s just a matter of more “OOMs” or CapEx spend before we unlock radically new capabilities, I’d argue that the currently available technology is already transformative given the right application mindset.

Why now for medium native AI

Existing models are sufficiently intelligent

The atomic units of knowledge work are typically not IQ-bound. Consider the drudgery of most work you do; I suspect you'd agree having a higher IQ wouldn’t make them substantially easier. Rather, discipline, focus, parallelizing yourself, and access to the right information probably matter a lot more. Top AI models already perform at par (or better than) humans on many benchmarks, especially where there’s significant data available to train on. The leap from where we are to beyond has a lot to do with giving models more autonomy and self-learning capabilities. Model agency is not requisite for significant value creation. We have the infrastructure and there are no chicken and egg problems

Unlike with bandwidth for the internet, we already have relatively sufficient infra to serve GPT-4o / Claude Sonnet level inference at reasonable scale. Unlike with internet/web2 networks, AI has significant value in a single player context and doesn’t require overcoming cold-start / chicken and egg problems. AI has an existence proof of utility in the fact that we already deeply understand that intelligence is useful. The real bottleneck: confusion & skeuomorphism Our main bottleneck isn't insufficient intelligence or infrastructure - it's confusion about what we should be building with AI. We're trapped in skeuomorphic thinking, viewing machine intelligence through the lens of human intelligence. In McLuhan terms, the first things we’re building with the new media (machine intelligence) are things we already build in the old media (human intelligence). The television just arrived and we’re still just televising radio broadcasts… we’re putting the yellow pages on the internet.

How did we get here? Well, there are many participants in the broader technology ecosystem that reinforce this skeuomorphic way of thinking about machine intelligence:

Researchers referenced the Turing test (human imitation at its core) as a longstanding goal post in the field for meaningful machine intelligence. Economists and think tanks forecast AI's disruption by reasoning about the labor force and what % of jobs are automate-able at increasing capability thresholds. Founders similarly make arguments about the value of their AI businesses with labor automation logic: “okay X many people currently have this job... they cost Y per year.. if we can automate that work then our market size is on the order of X times Y annually.” Investors create speculative sky high valuations (the short-term scorecard for company success) for companies that fit this metaphor, and you get a feedback loop to early seed investors/accelerators/founders that these are good companies. Silicon Valley thinkpieces abound with analogies to AI/agentic workers that can do work just like a human co-worker, take its recent favorite AGI bull (Leopold’s Situational Awareness) I expect us to get something that looks a lot like a drop-in remote worker. An agent that joins your company, is onboarded like a new human hire, messages you and colleagues on Slack and uses your softwares, makes pull requests, and that, given big projects, can do the model-equivalent of a human going away for weeks to independently complete the project. You’ll probably need somewhat better base models than GPT-4 to unlock this, but possibly not even that much better

The result is a solidification of the right way to think about AI's proliferation: AI will do what humans did. This result is not without consequence, our collective understanding of the future shapes how we build it, invest in it, what we contribute to, what founders choose to found, etc.. As capital (early stage in particular) becomes more abundant, cultural narratives and metaphors wield more force in what people decide to work on.

Labor automation is clearly a tempting mental model, and there will be some very successful companies built around it. But, consider a few downsides:

Its relative obviousness and legibility lends itself to intense competition. “We are the only AI company automating X job in Y vertical”, starts to sound a lot like “we are the only British restaurant in Palo Alto.” If you offer a business a way to do something that they already have a way of doing (paying humans, often inexpensively, who are super flexible and high-context), it has to be significantly better on some axes and highly reliable to justify risking what companies value most: predictability of growing profits. Many will not be sufficiently high ROI in a one to one comparison to existing globalized, technology augmented human workflows. Imagining medium-native machine intelligence applications It is common parlance in technology to speak of “AI-native”, or before it, “cloud-native” and “mobile native” as a way to identify upstarts in relation to their “non-native” incumbents. When I evoke “medium-native: here, my meaning goes much further. It’s not enough to simply “build with AI” to be medium-native, it’s to build in a way that takes advantage of its inherent strengths compared to what came before (human intelligence).

Instead of focusing on how AI is the same as human intelligence we should be asking how it is different. If we build applications around the things AI is uniquely suited to doing, we are very likely to come upon green-field territory for new services and products. While higher risk (there’s no pre-existing solution to anchor off of as evidence for demand), green-field can be a proxy for fast growth because if you’re right about the utility of a thing, no existing customers have a solution in place. This is category creation, and it’s the type of thing that can spread by word of mouth very quickly.

Below is a short list of illustrative examples of thinking in the skeuomorphic paradigm vs in the medium-native paradigm for machine intelligence.

Skeuomorphic AI Medium-native AI Do what humans did before, but automated Do a thing you never could have had humans do "AI Programmer" Socket: Every single time a codebase is changed, recursively scan all dependencies and intelligently analyze them for security issues and vulnerabilities. Patina: Spontaneous and customizable software for everyday consumer use "AI Salesperson" Clay: Intelligent data enrichment at massive scale "AI Coach" Tenor: Create infinite realistic and personalized simulations to practice difficult/emotional audio/video conversations in real time "AI Recruiter" Mercor: A website where anyone in the world can instantly get their skills assessed through a flexible AI interview and matched with relevant companies. "AI Scientist" Elicit: Search through and analyze many research papers simultaneously and at superhuman speed. What the medium-native AI applications have in common is leveraging the unique affordances of machine intelligence - namely, scale, its ability to be everything, everywhere all at once. All of these applications pass the following criteria: "what could I never fathom asking a human to do for me"?

While I won’t deter you from building the next great AI-laborer-(insert code for a job from the bureau of labor statistics that suggests a 100B+ market size), consider instead the road less traveled (so far). I personally expect it to generate the defining AI application companies of the coming decade.