What if we are the bubble?
Gil Dibner

Everyone's worried about the "AI bubble." We are all asking if the massive cash poured into generative models and automation tech will actually pay off. Is the hype real? Will companies like OpenAI and Nvidia deliver productivity fast enough to back up their insane valuations? These are certainly important questions, but they are perhaps orthogonal to the real issue: The bigger, more pernicious bubble may turn out to be human labor itself.

The human labor bubble. To understand the economic landscape and identify where the next pockets of opportunity lie, we need to talk about us (humans) and what we do all day. In a lot of cases, an honest assessment of knowledge work would reveal the "human labor bubble": the systemic overvaluation of routine, low-impact tasks that consume massive payrolls without generating proportional economic value. This idea is not new. In their influential 2014 work, The Second Machine Age, Andrew McAfee and Erik Brynjolfsson showed how digital platforms and smart machines are quickly taking over jobs previously believe to require people, especially office and admin work. Specifically, McAfee and Brynjolfsson focus on jobs such as routine data entry, basic legal discovery, and warehouse inventory management that are easily automatable by machines - and they wrote this ten years before AI was everywhere. Even more extreme is the theory (popularized by anthropologist David Graeber in his 2018 book, Bullshit Jobs: A Theory) that millions of people are occupied in roles they themselves think are pointless—like excessive middle management, low-value administrative roles, or professional 'flunkies.'

The tech startup landscape is not immune to this. In the ten years leading up to the ChatGPT moment, startups were often judged (and valued!) by how many people they hired. Large ZIRP-era venture rounds led to massive hiring which (viewed as a success in and of itself) led, in turn, to more venture rounds. I’m pretty sure it doesn’t take 200 engineers to build a simple CRUD application, but that’s what we as an industry were doing. Overall, these patterns suggest that a big chunk of our total economic output is just busywork caused by bloated company structures and a failure to reason critically about the value of human labor.

Spotting local human labor bubbles. Rather than fear the destruction of jobs by AI, it would be more constructive to embrace the idea that vast swathes of human knowledge work are systematically overvalued and have been for a very long time. The faster we can find them, the faster we can drive true productivity increases. This is a crucial point that is often overlooked in the current discourse on AI. We tend to be overly focused on inputs (“we bought a lot of GPUs!”) and demand (“we managed to rent out all those GPUs!”), when in fact we should be focused on outputs (actual increases in productive capacity per dollar of cost or unit of human labor). There is no better place to deploy an AI-powered solution than where a local human labor bubble exists. One way to diagnose where this “labor bubble” exists by looking for the following indicators:

• Basic data entry and classification work.

• Text or content generation problems.

• Humans acting as human APIs (performing simple, repetitive information lookups).

• Other simple cognitive tasks where complex human judgment, holistic reasoning, trust, or relationships are not required

A new lens on the AI opportunity. For me, this has become a more helpful lens on where the AI opportunity actually lies. Every day, we are meeting founders who are working to chip away at the foundations of the human labor bubble we’ve been steadily accumulating for years. There is appeal, no doubt, in the attempts to automate away entire job functions or professions such as lawyers or radiologists. But there is perhaps even greater appeal in a straightforward, bottom-up approach to identifying pockets of replaceable expensive human toil: the thousands of emails that must get categorized and responded to, the thousands of simple calculations that must get made repetitively, the thousands of pieces of data that must be fused from two different data sources over and over again, and the thousands of pages of text that must get entered into forms just to make the economy run.

We may be years away from automating and replacing human judgment with AI - but we are already deep in the process of replacing millions of hours of repetitive work that can easily be performed by the AI models we already have readily available at reasonable costs. Deploying such systems in real production environments is non-trivial. It requires exception handling, human-in-the-loop workflows, guardrails, security, hallucination management, deterministic taxonomy handling, hybrid semantic-vector queries, complex retrieval, and many other intricacies. It’s not as simple as unleashing an LLM on every problem. These sorts of problems are precisely the right balance of complexity and possibility that might enable defensible enterprise businesses to emerge. Over a reasonable time horizon, eliminating the human labor bubble is the single biggest AI opportunity we face. In most segments, the race to do that seriously has just begun.

The AI-Native Enterprise Playbook
Ten real-time observations on a rapidly evolving playing field

No More Painting by Numbers
It’s the end of the “SaaS playbook.

The Age of Artisanal Software May Finally be Over
Every wave of technological innovation has been catalyzed by the cost of something expensive trending to zero. Now that’s happening to software.

Founders as Experiment Designers
David on why founders should run everything as an experiment.

ENTERPRISE/TECH NEWS

FDEs - Following Palantir’s lead, more tech companies have been hiring ‘Forward Deployed Engineers’. While Palantir took the idea from the military over a decade ago and it makes up a large percentage of it’s overall work force, it is only recently that it’s reached mainstream coverage. In Financial Times this week they crunch the numbers on how the job title has seen over 1000% increase in usage in job posts since January 2023, with OpenAI building an FDE team and Cohere CEO quoted on how important the strategy is for customer alignment. 

For Profit - Microsoft and OpenAI announced a new agreement (or rather a restructuring of it’s earlier deal) which allows OpenAI to shift away from its nonprofit structure toward a public-benefit corporation, while Microsoft retains significant IP rights and exclusivity (extending through models and research up to about 2030-32). Satya Nadella discussed on TBPN the skepticism around his decision to make the initial (and in hindsight extremely modest) investment into OpenAI, recalling that Bill Gates said “Yeah, you’re going to burn this billion dollars.” (Bill Gates finds himself needing to row back more than one earlier proclamation this week)

HOW TO STARTUP

Prodigies - In ‘Your Age isn’t a Personality’ Nikunj Kothari talks about the drive for attention when seemingly every product looks the same, and why early attention doesn’t always translate to long-term success: "This is what AI did to building: when everyone can build, nobody cares that you built. So founders optimize for the only thing left that feels hard: getting noticed. The $100K launch video. The choreographed Product Hunt. The personal brand. All chasing the same dream: that one viral moment that takes you from $200M to $2B overnight. They tell themselves distribution is everything now. Building is commoditized, so audience becomes the moat. Get enough views and your X followers become customers. Sometimes it even works. For a month. Then reality hits: attention doesn’t compound. Retention does." Nikunj lived through this at a well-funded start-up where cash didn’t equate to product-market fit: “You’ll also die. Just with more money.”

HOW TO VENTURE

ARR Expectations - In Pat McGovern’s weekly radar he suggests it is time for a reset on early-stage revenue discussions, with VC’s selecting only unrepresentative outlier outcomes to discuss as ARR expectations. “My view: building a durable, high-growth business isn’t always going to allow for a $0 → $5MM Year 1 ramp. VC’s - let’s all take a breath. And founders - just tell the truth - there is nothing worse for trust than having to walk back and reclassify revenue the bulk of your ‘revenue’ on call #2 once the questions start flying.”

Rethinking Thinking - Charles Hudson, founder of Precursor Ventures writes about how building AI tools for his personal workflow and teaching AI to think like him has made him self-reflect on how he thinks. “The most significant benefit hasn’t been time saved or better decision support. It’s been that teaching Claude to be a good and useful partner forced me to be explicit about how I make decisions about founders, markets, and companies. I’ve always had a model in my head for how I weigh information, but I’d never really articulated it to anyone - including myself.”

Blue Energy announces partnership with Crusoe to develop a nuclear-powered data center campus in Texas.

Portchain has saved 800k metric tonnes of CO2 emissions.

Belidor
Founding Full-Stack Engineer (Paris)

Firebolt
Account Executive (Bay Area, Hybrid)

Paradime
Chief of Staff (NYC, London)