Import AI 452: Scaling… | Jack Clark from Import AI Summary

1. Key Themes

Theme 1: AI Cyberoffense Is Scaling as Fast as AI Capability Generally

Frontier AI models are demonstrably better at offensive cyberattacks with each generation, and the improvement curve is accelerating.

"Across frontier models released since 2019, the doubling time is 9.8 months. Restricting to models released since 2024, it steepens to 5.7 months. The most recent frontier models in our study, GPT-5.3 Codex and Opus 4.6, sit above both fitted trendlines, achieving 50% success on tasks taking human experts 3.1h and 3.2h respectively."

Clark's framing is stark: AI is an "everything machine," and capability gains in beneficial domains automatically extend to harmful ones. The policy surface area grows with every model generation.

"AI that is especially good at helping you find vulnerabilities in code for defensive purposes can easily be repurposed for offensive purposes."

Theme 2: AI-Native Startups Are Structurally More Competitive

A large-scale field experiment shows that startups which learn how to map AI across their production process dramatically outperform peers — not just in output, but in capital efficiency.

"Treated firms complete 12% more tasks, are 18% more likely to acquire paying customers, and generate 1.9x higher revenues compared to control firms."

The capital efficiency finding is especially significant for investors:

"Treated firms report just over $220,000 less in capital demand relative to control firms, a 39.5% decrease (p < 0.05), with no corresponding increase in labor demand."

Theme 3: AI Automation Is a Rising Tide, Not a Crashing Wave

MIT research across 3,000 tasks and 17,000 worker evaluations finds AI capabilities are expanding broadly and simultaneously across job types — gradual, but ultimately pervasive.

"Rather than arriving in crashing waves that transform a certain set of tasks at a time, progress typically resembles a rising tide, with widespread gains across many tasks simultaneously."

The timeline is concrete and near-term:

"Most tasks are projected to attain AI success rates of 80%–95% by 2029 at a minimally sufficient quality level."

Theme 4: The Mapping Problem Is the Bottleneck, Not the Technology

The INSEAD/HBS study reframes the competitive AI advantage: it's not about access to tools, it's about knowing where to apply them inside the business.

"The bottleneck is not the technology — it is the managerial challenge of discovering where the technology creates value within a firm's production process. Teaching managers and entrepreneurs how to solve the mapping problem may be at least as important as ensuring they have access to the technology."

2. Contrarian Perspectives

Contrarian 1: Experts Expect Smarter AI But Not a GDP Revolution — And That Tension Is Unresolved

Despite widespread expectation of rapid AI capability progress, a major cross-cohort forecasting study finds muted economic impact predictions. This runs against the narrative pushed by frontier labs.

"The most surprising finding is that all the surveyed groups expect AI systems are more likely to make moderate to rapid progress in coming years rather than slow progress, but that the impacts on GDP will be relatively minor, adding ~1 point (relative to 2025's 2.4%) by 2030."

Clark raises but doesn't resolve the key question: is this a bearish signal on AI, or a universal human failure to model exponentials?

"Is this discrepancy a bearish signal on AI progress, or is it indicative of the fact that humans are universally bad at truly modeling exponentials? It's hard to say."

Evidence behind the forecast: 69 economists, 52 AI industry and policy experts, 38 highly accurate forecasters, and 401 members of the general public were surveyed from mid-October 2025 to end of February 2026. Economists put only a 14% probability on AI causing major GDP increases in the short term.

Contrarian 2: Open-Weight Models Will Democratize Offensive Cyber Capability on a Short Timeline

The gap between closed-source frontier models and open-weight models in cyberoffense is shrinking fast — meaning nation-state-level attack capability could become widely accessible.

"Our most recent open-weight model, GLM-5, lags the closed-source frontier by 5.7 months, suggesting that frontier offensive-cyber capability may diffuse into open-weight form on relatively short timelines."

This undercuts the assumption that capability controls on frontier labs are sufficient to manage AI-enabled cyberwarfare risk.

Contrarian 3: AI's Disruption Will First Appear as a New Class of Leaner Competitors, Not Mass Unemployment

The framing of AI risk as mass job displacement may be missing the more immediate mechanism: new AI-native entrants destroying incumbents through superior capital efficiency before labor displacement becomes visible.

"It surely implies that one of the ways we'll see AI first show up in the economy will be the emergence of a new class of competitive firms that are more efficient with capital (in part by employing fewer people) than the firms they displace."

3. Companies Identified

Company	Description	Why Mentioned	Quote
Gamma	AI-powered product development startup	Case study in AI-enabled product iteration at scale	"Used AI to detect usage patterns and generate product variants directly, enabling a single PM to continuously ship features that would previously have required an entire team."
Ryz Labs	Software/product development firm	Case study for AI-parallelized product development	"Founder writes a Product Requirements Document and feeds it into multiple AI coding tools simultaneously, building the same idea multiple ways rather than betting on a single approach."
FazeShift	Process automation startup	Case study for eliminating human steps in financial workflows	"Showed how to automate an accounts receivable process by using AI to skip over the human steps."
Ranger	Early-stage startup	Case study for bootstrapping with AI to improve fundraising outcomes	"An illustration of how to use AI to bootstrap a startup, get initial traction, improve margins, and then raise money later when the business is more mature, which allows them to raise at better rates."
OpenAI	Frontier AI lab	Technical partner in the INSEAD accelerator; models evaluated across cybersecurity benchmarks	Referenced as onboarding partner providing API access and frontier model access to startups.
Manus	AI tooling company	Technical partner in the INSEAD accelerator experiment	Referenced as onboarding partner alongside OpenAI for the startup field experiment.

4. People Identified

Person	Description	Why Mentioned	Quote
Jack Clark	Author of Import AI; co-founder of Anthropic; former OpenAI policy head	Author and analyst; explicitly acknowledges his own position as a "frontier lab" voice with potentially biased predictions	"Studies like this are hard to reconcile with the panicked and sometimes breathless-seeming provocations about AI-driven societal change that come from frontier labs (including myself!)."

5. Operating Insights

Insight 1: Expose Your Team to Concrete AI Use Cases, Not Just Tools

The INSEAD/HBS experiment shows performance gains came specifically from teaching founders how other companies had reorganized around AI — not just giving them API access. Founders who got $25K in tools without the use-case workshops underperformed.

"Each additional AI use case prompted by treatment leads to 0.85 more completed tasks and approximately 26% higher revenue."

Tactical implication: Run structured internal workshops featuring real AI transformation case studies from comparable businesses. The ROI on this is measurable and large.

Insight 2: Use AI to Raise Less Money at Better Terms

AI adoption doesn't just make startups more productive — it changes the fundraising calculus by reducing how much capital is needed before achieving proof points.

"Treated ventures achieve faster growth without proportional increases in labor or capital, consistent with a reduction in the costs of experimentation and scaling seen in earlier technological waves."

Tactical implication: AI-native startups should reach customer and revenue milestones with less capital, then raise — compressing dilution and improving leverage at the term sheet.

Insight 3: Treat "Force Multiplier" as Literal, Not Metaphorical

Founders in the experiment reported that AI didn't replace expertise — it amplified it, collapsing the cost of work that previously required outsourcing or larger teams.

"One treated founder reflected: 'In just a few hours I was able to produce what previously cost $1,000 from an outsourced dev team.'"

6. Overlooked Insights

Overlooked Insight 1: Economists Favor a "Manhattan Project" for AI Development as a Policy Intervention

Buried in the GDP forecasting section, the surveyed economists express support for a large-scale nationally coordinated AI development effort — and are notably cool on UBI, compute taxes, and job guarantees.

"The surveyed economists like modernized unemployment insurance and a large-scale AI development project (manhattan project) as interventions, and are a lot less keen on job guarantees, taxing compute, or universal basic income."

This signals a potential policy direction that could concentrate AI investment and capability in state-backed institutions — a material consideration for investors assessing geopolitical AI risk.

Overlooked Insight 2: AI Capability Progress Is Already Measurable in Work-Hours, Not Just Benchmarks

The MIT and Lyptus research both independently converge on using human expert time as the unit for measuring AI capability progress — a more economically legible metric than traditional benchmarks.

"Between 2024-Q2 and 2025-Q3, frontier models went from achieving a 50% success rate on 3- to 4-hour tasks to 1-week tasks, and achieving a 70% success rate on 1-minute tasks to 1-hour tasks."

This "task-hours" framing may become the standard way investors and operators assess AI substitution risk across job categories and business functions.