The Age of Abundant Thinking: When Master's-Level Intelligence Becomes Tap Water
One Hour, Zero Security Background: How I Used AI to Build a Framework That Predicted a $292M Hack
Three weeks ago, when I watched Drift Protocol get hit by the Lazarus Group for $285M, a framework forced its way into my head.
So I asked an AI to write it up — in about an hour — and then used it to flag fifty-something at-risk DeFi protocols in a public report.
Twelve days later, one of those flagged projects — Kelp DAO — was actually drained. 177M in bad debt cascading into Aave V3. It became the largest DeFi exploit of 2026 so far.
Full technical breakdown: Kelp DAO $292M Vulnerability — Flagged 12 Days Before
The report ended up being shared by a lot of people in the industry — the founder of Chaos Labs, the host of Unchained Podcast, TechFlow, and 20+ other outlets in both Chinese- and English-speaking markets. For someone with no formal security background — someone who's just observing mechanism behavior — that propagation path itself was a signal: this industry is missing a "first-principles, fast risk-repricing" point of view.
The framework and the specific incident are complicated, and I'm only laying out the surface here. What I really want to talk about isn't the case itself — it's the shift in era that the case quietly reveals.
From Animal Power to Electricity: How Scarce Resources Get Replaced
In the agricultural era, animal power was the scarce resource. Humans raised oxen and horses to produce the basic unit of labor.
After the Industrial Revolution, electricity replaced animal power. The energy yield of electricity dwarfed any ox or horse, and animals quietly exited the labor market.
But in the early industrial-electric era, a new scarce resource emerged: human-level thinking. The wage structures, education systems, and professional hierarchies of modern society are essentially built on one assumption — human-grade thinking is scarce, and therefore valuable.
After Large Language Models: Thinking Is No Longer Scarce
In the post-LLM era, human-level thinking is no longer scarce.
After listening to Reid Hoffman interview Ivan Zhao (co-founder of Notion) on the Possible podcast, one idea crystallized in my head: we are living through the transition from a scarcity-of-thinking era to an abundance-of-thinking era.
In that episode Ivan Zhao used a metaphor I love — he treats computation as a material, like steel or steam. When the supply of that material grows without limit, the whole industry around it gets reorganized.
Human-level thinking can now be allocated and consumed the way electricity is allocated through wires.
Opportunity and Challenge: Master's-Level Intelligence Becomes Tap Water
My own read: we are now in the late LLM era, where master's-level intelligence is starting to become abundant. That brings opportunity, but also a challenge.
The biggest challenge: what we used to consider scarce is suddenly abundant, and neither individuals nor society have updated their mental models.
Bachelor's and master's-level knowledge accumulation isn't unimportant — but it increasingly resembles tap water: turn the faucet, and it's there. It's an infrastructure layer now, not a scarce asset, and it can no longer directly define a person's value.
So what do humans actually do in this era? My answer: become someone who can drive intelligence-grade agents.
Not compete with them — orchestrate them, dispatch them, and connect them to the right problems. That skill is fundamentally different from "having a master's-level body of knowledge."
And Intelligence Abundance Will Only Push Higher: AI4Math Is the Real Test
This era keeps moving. And I believe the real test isn't writing code, isn't writing essays — it's doing math.
I've been watching 25-year-old Carina Hong lately. She did a math/physics double major at MIT, a neuroscience MSc at Oxford, then went on to a math + law dual PhD program at Stanford — before dropping out to found Axiom Math, which just closed a 1.6B valuation.
What she's doing is AI for Math — building an "AI mathematician" that can automatically verify code through mathematical reasoning. Google DeepMind's AlphaFold already pushed this kind of machine intelligence into biology; Axiom Math is trying to push it to the frontier of research-level mathematics.
Why Math Is the Real Test
If we tier machine math capability, the rough levels look like this:
| Level | Capability | Where We Are |
|---|---|---|
| L1 | Arithmetic, high school math | GPT-3 (2020) basically had it |
| L2 | Math Olympiad (IMO) | 2024 silver → 2025 gold (DeepMind AlphaProof / Gemini Deep Think) |
| L3 | Undergraduate / graduate level | Maturing through 2025–2026 |
| L4 | Research-level math (frontier-pushing) | Axiom Math, Terence Tao, etc. |
| L5 | Autonomous discovery of new theorems / open problems | The endgame (Riemann, P=NP…) |
L4 is the real test for three reasons:
First, math is the lingua franca of all science. If AI can move freely at the frontier of research-level math, it can produce original contributions in physics, chemistry, cryptography, chip design, drug discovery — because all of those disciplines eventually have to land in math to be rigorously verified.
Second, math's self-verifying property is the root solution to AI alignment and AI safety. Once you commit to formal proofs (Lean, Coq), the result is 100% machine-verifiable — AI can check its own correctness. Which means future AI-written code can be proven by AI to be "provably bug-free." That's exactly Axiom Math's path: using math to auto-verify code. If this works, the problem my DeFi safety framework solves today goes from "probabilistically flag suspicious projects" to "mathematically prove this contract is safe." That's a phase change.
Third, this is the breakthrough point for "PhD-level intelligence." What a PhD does, fundamentally, is create new knowledge at the frontier. Math is one of the few domains where new knowledge can be 100% verified — so once AI breaks the math research ceiling, it has broken the PhD-level intelligence ceiling. The rest is just cross-domain replication.
In other words, if Axiom Math succeeds, machine cognition doesn't just reach PhD-level — it learns to self-expand at the top of the PhD pyramid. That's a phase change, not a quantity change.
And future architectures will keep pushing today's master's-level AI agents to PhD-level, then beyond — across every domain, not just one.
When that phase change arrives, my one-hour DeFi safety framework will look, in retrospect, like an appetizer.
The Questions We Should Actually Be Asking
I have to be honest about something here: most people are not actually participating in this revolution.
Master's-level problems have been solved. Next is PhD-level, then beyond-PhD-level. But the actual producers of this technology are a tiny set of companies, labs, and pools of capital — most people only get to receive the consequences of the revolution in their daily life, work, and consumption. They are not at the table where the direction of the revolution gets defined.
Software that used to require a whole team can now be built by one person + AI; research that used to take a lab years can now be moved meaningfully forward by a single person in months. The question isn't "which jobs will disappear" — it's "where will this seismic shift in social structure take us?" When the resource that used to be most scarce — thinking — becomes tap water, the entire pricing logic of the economy gets rewritten. Just like when electricity made animal labor obsolete.
If most of us aren't at the table, the least we can do is ask the right questions. There are two that matter more than "will I lose my job":
Question 1: How Do We Govern PhD-Level AI?
When AI starts producing original results across math, biology, physics, chemistry — every domain that used to require a PhD — alignment stops being just an engineering problem and becomes a political one.
The concrete sub-questions:
- Who do these AIs actually serve? The current honest answer: paying enterprise customers, militaries, advertisers, and the small group of people with capital to participate in training — not the broader public whose data was used to train them and whose lives are most reshaped by them.
- How does the public participate? Anthropic and OpenAI have both worked with the Collective Intelligence Project to run Alignment Assemblies — a randomly selected group of ordinary citizens, given full expert briefings, voting on AI behavior boundaries. This is one of the most underrated democratic experiments of our time.
- Should there be a "Public AI"? Like public schools, the postal service, or municipal water — a publicly owned, publicly governed AI infrastructure as a counterweight to purely commercial AI monopolies.
- The international coordination vacuum. As of early 2026, roughly 90 countries have published national AI strategies and 33 countries have passed AI-specific legislation — but there is still no enforceable international framework. We are in roughly the same position as the early years of failed nuclear governance.
None of these are purely technical questions. But they will determine who this technology ultimately serves.
Question 2: Beyond Marketing, What Can AI Actually Do For Humanity?
Back to applications. Most AI products on the market today are essentially marketing-layer: better customer service, faster content generation, more precise advertising, smoother workflows. These are good things — but they are only "good things." They are not change-the-human-condition things. In fact, by Q1 2026, 95% of enterprise AI pilot projects had failed to deliver real ROI. A vast amount of compute is being burned on marginal improvements.
In Machines of Loving Grace, Anthropic's CEO Dario Amodei lists five domains where AI could actually change the human condition:
- Health — accelerating cures for cancer, Alzheimer's, rare diseases; he predicts AI could compress 50–100 years of biomedical progress into a single decade.
- Mental health — bringing high-quality therapy to anyone with a phone.
- Economic development — letting the poorest countries directly leapfrog into developed-world healthcare, education, and infrastructure.
- Peace and governance — verifiable, transparent government decisions; eroding the information advantage of authoritarian regimes.
- Labor and meaning — redistributing the gains from productivity so humans have time to do things that matter.
Any one of those five matters infinitely more than "yet another SaaS subscription."
And back to my own little example — the DeFi safety framework that flagged Kelp DAO 12 days before the $292M hack. Does that count as a real contribution to humanity? I'd say half of one. It does let an individual without a security background reprice risk in capital markets, which is good. But the bigger arena is healthcare, education, scientific research, public governance — the domains that genuinely need to be repriced and reorganized.
I believe the truly valuable work in this era is not competing with AI on master's- or PhD-level capability — it's answering the two questions above: how to govern it, and how to point it at problems that actually matter.
The story about a one-hour framework that predicted a $292M loss 12 days in advance is just a small side-shadow of a much larger transformation.
The real story has not yet been written.
Further reading:
- Kelp DAO $292M Vulnerability — Flagged 12 Days Before
- Reid Hoffman × Ivan Zhao on Possible — "Notion's Reboot: Ivan Zhao on AI as a New Material" (March 2026)
- Terence Tao at OpenAI Academy — "AI Is Ready for Primetime in Math and Theoretical Physics" (March 2026)
- UDN Daily News — 25-year-old Carina Hong founds Axiom Math to build "AI mathematicians"
- Dario Amodei — Machines of Loving Grace
- Brookings — How public AI can strengthen democracy
- Collective Intelligence Project — Alignment Assemblies
