Co-Existence

At a Glance

Ethan Mollick (Wharton School professor and bestselling AI author) discusses with Scottish Mortgage fund manager Lawrence Burns the evolution of AI from chat to reasoning to agents, and how businesses and society can coexist with it. Mollick argues that AI is advancing faster than expected, but the biggest mistake businesses make is assuming there is a "standard answer"—in reality, no one truly knows where the future is headed. Companies need to simultaneously adapt and experiment, rather than passively waiting for solutions.

Theme 1: AI's "Jagged Frontier" — Rapid Progress but Uneven Distribution

Mollick introduces the concept of the "Jagged Frontier," meaning AI far exceeds expectations in certain areas while still falling short in others, and progress is not evenly distributed.

Historical context and data support:

Three years ago, AI could not even count the number of the letter "r" in "strawberry"; yet "a week ago, it was solving millennium-old mathematical problems, throwing the math community into crisis."
The pace of progress has exceeded Mollick's own expectations: "I thought human adoption of AI would take time... but even wearing the skeptical hat of a business school professor, progress has been faster than I imagined."

Persistent shortcomings:

Long-form writing: AI can produce decent text, but "for long non-fiction documents or reports, it easily loses the main thread and confuses the conversational context."
Lack of diversity: AI essentially performs tasks like "a very smart person," yet many scenarios require diversity — "with the right prompts, AI can generate more diverse ideas than a single human, but still less diverse than a group of humans." Different models (Chinese models, Grok, Anthropic, OpenAI) are "strikingly similar" in political attitudes, for reasons that remain unclear (possibly training data, reinforcement learning, or some "Platonic" convergence).

Theme 2: Four Key Questions for Enterprise AI Adoption

Mollick proposes four critical questions to assess a company's AI maturity, arguing that merely claiming to be "AI-first" is far from sufficient.

The four questions:

1. What have you stopped doing because of AI? — If the answer is nothing, "there is a problem"; for example, market research reports can now be generated on demand, and companies should reconsider whether they still need to pay for them in the old way.

2. Are you doing anything "impossible"? — AI enables you to achieve goals that were previously unattainable; "you should have at least one attempt that you believe could transform the industry."

3. Are you betting on the curve improving? — "You should also be building something that currently doesn't work, betting that the curve will improve over time."

4. Do you have a "leadership-laboratory-crowd" structure? — Leadership is genuinely committed to AI and establishes reward systems; the crowd uses the best tools to experiment with use cases; the laboratory conducts 24/7 AI development (not just technical, but also business and organizational aspects).

Mollick emphasizes: No one knows the best use cases, "if you don't have a leadership-laboratory-crowd structure, you may need to fix that."

Theme 3: AI’s Reshaping of Professions and Companies—The Bottleneck Keeps Shifting

Mollick argues that AI will not simply “replace” jobs, but will continuously shift the location of value bottlenecks, requiring companies to adapt dynamically.

O-ring theory analogy:

Using the Challenger space shuttle as an example—“Everything worked except for one O-ring, but that small failure caused a catastrophic disaster”
Similarly, there are “multiplier effect” small failures in work: bottlenecks determine the value of work, and these bottlenecks shift as AI capabilities evolve

Examples from programming:

In the past: Writing clear code was the primary value of programmers
Now: “The best engineers I know almost never write code—some don’t write a single line,” but they excel at “directing AI systems, writing test sequences, and defining requirements”
Projection: This new bottleneck (supervising AI) “may also be resolved, and then the bottleneck will shift again”

Mollick warns: Do not jump to conclusions too early—“If you fire your marketing staff because AI can write marketing copy, a year later you’ll find that when I smell Claude in any reading material, I won’t engage with it as much as before”

On “adaptability”:

Burns suggests that “the only thing you can bet on is adaptability.” Mollick agrees but adds: “This is not just about adaptation—no one knows the exact answer… We also need to invent this future, not just react”
Companies need to combine “experimentation and adaptation,” actively creating variation and selecting winners from it

Theme 4: AI Regulation — Risks Exist, But Don't Let "Existential Risk" Distract You

Mollick believes regulation is inevitable, but current discussions are overly focused on "existential risk," neglecting more pressing real-world issues.

Assessment of AI Companies' Motivations:

"I think their concerns about risk are genuine... but I'm not sure their proposed solutions are always genuine — because these solutions would materially impact their businesses"
Self-regulation "may happen to some extent, but it has its limitations"
Geopolitical factors (the US-China race) complicate government regulation

Mollick's Core Concerns:

"If all energy is concentrated on existential risk... even if we stop developing AI more advanced than Astra and Fable 5.1, the next five years will still see massive changes"
Overlooked risks: AI's impact on employment, its effect on financial markets — "we need regulatory frameworks to address these issues"
"Regulation is oversimplified as 'controlling the frontier,' when we haven't even fully utilized the frontier's capabilities"

Regarding the Regulatory Impact of Anthropic and OpenAI:

Mollick believes the anxiety of AI researchers (especially safety researchers) is genuine — "they didn't want to release GPT-2 initially, thinking it was too dangerous"
But their solutions "aim to solve problems while not hindering their own success as companies"
Citing research such as Tetlock's: the probability of catastrophic risk by 2030 is about 0.5% — "not zero, but more likely a major cyberattack than human extinction"

Theme 5: The AI Investment Cycle — Possibly Not a Typical "Boom-Bust"

Mollick remains open to whether AI will inevitably undergo a historical "boom-bust" cycle, arguing that the current situation differs from historical analogies.

Historical analogies (canals, railways, fiber optics):

Classic pattern: Overinvestment → bust → second-generation companies gradually consolidate using the wreckage of first-generation firms
Mollick is uncertain whether this will repeat: "I'm not sure it's wise to bet that Anthropic and OpenAI will disappear — even in a crash, Google won't vanish"

Differences:

"Most companies report getting ROI from generative AI… and they are using very primitive systems"
The era of agents has not yet been factored into the data, and integration takes time
"These systems are useful, and the long-term operating costs are not high"
This may be more like "Industrial Revolution 2.0" rather than a repeat of "Industrial Revolution 1.0"

Burns adds (Scottish Mortgage perspective):

"We are surprised by how many companies firmly say, 'Our tokens have generated good returns' — this is positive ROI"
The shift from "experimental AI" to "AI will have a material impact on revenue or the cost base" — this gives them "a great deal of confidence in the sustainability of demand on both the compute side and the model side"

Mentioned Positions

Position	Guest Stance	Key Data
Anthropic	Neutral (risks and business incentives coexist)	Founders genuinely concerned about risks, but solutions also serve commercial interests
OpenAI	Neutral (same as above)	Once considered GPT-2 too dangerous to release
Google	Not explicitly stated	"Even in a crash, Google won't disappear"
Grok (xAI)	Not explicitly stated	Political stance is "strikingly similar" to other models

Judgments Worth Remembering

1. The "Jagged Frontier" Law (Mollick): AI far exceeds expectations in some areas, remains inadequate in others, and progresses unevenly — "AI excels at some things you anticipate, struggles with things you don't expect, but advances very rapidly across many domains."

2. The litmus test for enterprise AI maturity is "what has been stopped" (Mollick): "If you haven't stopped doing anything because of AI, there's a problem" — true adoption eliminates old processes, not merely adds new tools.

3. AI's diversity deficiency is structural (Mollick): Different models are "strikingly similar" in political attitudes, for reasons unknown — "You can increase diversity through AI committees and personality prompts, but there remains an inherent similarity that is hard to separate."

4. The value bottleneck keeps shifting (Mollick): Using programming as an example — "The best engineers almost never write code anymore, but they have become excellent supervisors of AI systems" — this new bottleneck "may also be resolved, and then the bottleneck will shift again."

5. "No one knows the exact answer" (Mollick): This is Mollick's core assessment of the current AI phase — "I interact with frontier models, government officials, CEOs... but what you should know is: no one knows anything for certain."

6. Existential risk discussions may divert attention from real-world risks (Mollick): "Even if we stop developing AI more advanced than Astra and Fable 5.1, enormous changes will still occur in the next five years" — employment and financial market impacts require regulatory frameworks.

7. AI's ROI is already visible, and using primitive systems (Mollick/Burns): "Most companies report ROI from generative AI... and they are using very primitive systems" — the agent era has not yet been factored into the data.

8. The AI investment cycle may differ from history (Mollick): "This could be a repeat of the 'Industrial Revolution 2.0' rather than the 'Industrial Revolution 1.0'" — it does not necessarily go through the typical boom-bust pattern.