Musk agrees with Will Depue’s assessment: AI capabilities will not slow down, and recursive self-improvement (RSI) will arrive in less than 24 months.

Author and source: 0x9999in1, ME News



TL;DR

• Musk agrees with Will Depue’s assessment: AI capabilities will not slow down, and recursive self-improvement (RSI) will arrive in less than 24 months. I agree with the first part, but reserve judgment on the second.
• But there is a mismatch in this conversation: the mainstream argument against valuing frontier labs has never been that "the models don't work," but rather that "the models work well, but you can't collect payment."
• Elon Musk's line, "AI two years ago was old enough to go into a museum," is actually the best line for short sellers. It speaks to the rate of asset depreciation, not a moat.
• The capability curve is backed by hard data: the "task time span" measured by METR approximately doubles every seven months, and has accelerated even further over the past two years. This is currently the most important metric to monitor.
• The issue with RSI is not about possibility, but about its definition being too vague. Narrow-domain self-improvement has already occurred (AlphaEvolve is empirical evidence), but the full-domain closed loop is still hindered by computational power, energy, experimental cycles, and evaluation.
• Elon Musk’s timeline carries historical baggage. FSD has been touted as "next year" for a decade, and AGI was predicted for 2025 and 2026. The direction is right, but the timing is often off.
• My conclusion: The ability to go long does not equate to going long on any specific balance sheet. These two things must be bet on separately.

They are not answering the same question.

Let me clarify this first.

Will Depue’s original statement includes a qualifying clause: this bearish view “is quite reasonable if you ignore AGI.”

This premise already acknowledges half of the short seller's logic.

Most bearish observers of Frontier Labs’ valuation do not believe the model will stagnate. Their arguments center on other factors: unviable unit economics, excessively high capital expenditure intensity, underestimated depreciation cycles, competition eroding pricing power, and enterprise adoption rates far below expectations.

None of these rely on a slowdown in AI capabilities.

So, Musk’s statement that “AI in two years will blow people away” is, as a judgment of capability trends, likely to be correct. However, as a rebuttal to bearish arguments, it doesn’t actually hit the mark.

That was a solid punch, but it hit nothing but air.

Let me put it more plainly: the technology curve and the cash flow curve are not the same. Historically, they’ve often moved in opposite directions. The faster fiber is laid, the cheaper bandwidth becomes—and the worse off the companies laying it fare. In the late 1990s, U.S. long-haul fiber capacity surged, causing prices to collapse; Global Crossing and WorldCom both ended up bankrupt. The real beneficiaries of that infrastructure boom? Companies building streaming and cloud services a decade later.

Technology won. Those who invested in technology didn’t.

This is not discouraging AI. This is a reminder: there are many missing steps between “AI will become extremely powerful” and “this valuation is reasonable today.”

Second, the metaphor of "going to the museum" is actually evidence of a short position.

I acknowledge the insight in Musk’s statement.

Two years ago, in mid-2024, the leading models were GPT-4o, Claude 3.5 Sonnet, and Llama 3.1 405B. They lacked substantial reasoning-time computation, reliable long-chain tool usage, and could not perform autonomous tasks lasting hours.

Today, it really seems like a specimen.

But please note what this statement actually describes.

It describes: AI models as assets that depreciate rapidly.

The economic lifespan of cutting-edge models may be only 12 to 24 months. Once the next generation emerges, the marginal value of the training compute, data, and labor costs invested in the previous generation rapidly declines. You cannot rely on it to generate revenue for a decade.

And what about the hardware supporting all of this? Major hyperscale vendors typically depreciate GPUs and servers over five to six years in their financial reports. Is this assumption too long? It was publicly questioned in 2025, and Michael Burry directly called attention to it in November of that year: if the actual economic lifespan is shorter, then reported profits are overstated.

I don’t want to defend short sellers, but the logic needs to be evened out: if a model is already outdated after two years, why should the steel and silicon that carry it be allowed to depreciate over six years?

The more true Musk's statement is, the more acute this issue becomes.

Even more problematic is pricing. Epoch AI’s tracking shows that the cost of achieving the same level of inference performance can decline by an order of magnitude annually. This is fantastic for users but represents a relentless squeeze on gross margins for companies trying to monetize their models.

Your product improves dramatically each year, while your pricing is cut significantly annually.

This is the source of the mechanism known as “racing to capability, struggling for profit.” It’s not poor execution—it’s structural.

Three: Capabilities Won't Slow Down: This point is backed by data in favor of Musk.

Now let me state the parts I agree with.

If we only discuss it based on "feelings," we'll never reach a clear conclusion. So I prefer to look at a specific metric: how long a model can independently complete a task.

The 2025 study by METR revealed a robust trend line: the length of tasks that models can complete with 50% success, measured against the time required by human experts to perform similar tasks, approximately doubles every seven months. For samples after 2024, the slope is even steeper.

By the second half of 2025, the strongest models have entered the "two-hour range" on this metric.

Please consider the meaning of this curve: exponential growth appears gradual at first but suddenly accelerates later. It takes only two to three doublings to go from two hours to a full workday.

Other benchmarks emerged simultaneously. In September 2025, OpenAI launched GDPval, evaluating models using real-world professional deliverables, with state-of-the-art models already approaching industry-expert performance on certain occupational tasks. In July 2025, experimental models from multiple labs achieved gold-medal-level performance at the International Mathematical Olympiad. These are not marketing slogans—they are verifiable public events.

Therefore, I believe the statement "frontier progress has slowed" does not hold up in terms of capability.

What has truly slowed down is something else: users' threshold for surprise.

We are shocked every three months, until shock becomes routine. This subjective numbness is easily mistaken for objective stagnation.

At this point, Musk is clear-headed.

Four: RSI Under 24 Months: Which Half Is True, and Which Half Is Belief

This is the part I disagree with the most in the entire discussion.

It's not because it's too optimistic, but because it's too vague.

There are at least three versions of the RSI with completely different scales.

First version: AI improves its components within a narrow domain.

This has already happened. Google DeepMind’s AlphaEvolve, announced in 2025, served as proof: it automatically discovered a more efficient algorithm, reducing the number of multiplications required for 4×4 complex matrix multiplication from 49 to 48, breaking a record that had stood for over fifty years, and recovering approximately 0.7% of global computing power in Google’s internal data center scheduling.

This is a real, verifiable case of AI improving AI infrastructure.

Version 2: AI significantly takes over its own development process.

This is currently in progress. Coding agents are already handling a significant portion of engineering implementation within the lab, with code reviews, experiment scripts, and data pipelines being automated.

Third version: An autonomous acceleration loop free from human bottlenecks.

Also known as "intelligence explosion."

The first two versions were used to argue for the third, which is the most common slippery slope in this type of prediction.

Why is it difficult? The bottleneck isn't in algorithmic inspiration at all.

The pace of AI research is limited by the physical time required for experiments. A large-scale pre-training run can take weeks, and you must wait for hardware, compete for power, and queue for access. Even if you increase a researcher’s thinking speed by a hundredfold, the experiment queue won’t speed up by the same factor.

Beneath that lies energy and supply chains—advanced packaging capacity, HBM supply, transformer delivery cycles, and grid connection wait times—all of which move on a quarterly and annual scale. Around 2025, grid connection wait times in multiple regions across the U.S. have already reached multi-year levels.

Silicon wafers can work overtime; substations cannot.

Another less-discussed constraint is evaluation. For AI to improve itself, it must first know whether it has actually gotten better. When tasks stretch to dozens of hours and answers no longer have a single correct solution, determining what “better” means becomes a research challenge in itself. Reward signals become noisy, and models may optimize for the evaluator rather than for actual capability.

So my assessment is: Within 24 months, we are very likely to see significant progress on the second version, with AI's share of the development process continuing to rise. However, I believe the probability of the third version—the true recursive loop that fully breaks free from human pace—emerging within 24 months is significantly less than 50%.

This is not technological pessimism; this is engineering reality.

Five: Musk's timeline requires a discount factor.

This isn't an attack on people; it's a calibration.

His track record in predicting trends has been remarkably accurate: electric vehicles will prevail, reusable rockets are feasible, satellite internet can succeed, and large-scale GPU clusters are the critical variable. The fact that his Colossus cluster was deployed at an unprecedented scale of 100,000 GPUs in 2024 within an extremely short timeframe demonstrates his deep understanding of the essence of the compute race.

However, his judgment of time has consistently been overly optimistic.

Fully autonomous driving "will be achieved next year"—a claim made since 2015. He publicly predicted that AI would surpass any individual human around 2025 and that AGI would emerge around 2026. Today is July 31, 2026, and these predictions have clearly not materialized.

Best practice: Multiply the schedule he provided by two to three.

There’s another important point to clarify: he is not an impartial observer.

xAI is heavily betting on this path, requiring continuous massive funding and market belief that the curve will continue to steepen. It is entirely normal for an individual to express strong confidence in the direction they are betting on, but the recipient should account for this conflict of interest in their weighting.

This does not affect the potential correctness of his judgment; it only affects how much confidence you should place in it.

Six, the implementation on the enterprise side is where the real gap lies.

In August 2025, a study by a team associated with MIT sparked widespread discussion: 95% of enterprise generative AI pilot projects failed to deliver measurable improvements in profit and loss. Later, many questioned the methodology of this study, raising concerns about its sample size and metrics; I do not recommend treating it as definitive evidence.

But the pain point it touches is real.

Between model capabilities and enterprise value lies an entire set of unglamorous, often neglected tasks: data governance, access control, process reengineering, audit trails, accountability assignment, and failure containment.

A model that can win a gold medal can't enter a bank's core process—the bottleneck is never intelligence.

This also explains a seemingly contradictory phenomenon: usage is exploding, yet returns are difficult to quantify. Google’s monthly token processing volume rose rapidly from the quadrillion level to the quintillion level in 2025, and OpenAI’s weekly active users reached approximately 800 million by autumn 2025. Usage is unquestionably taking off.

At the same time, according to multiple media reports, leading laboratories are still expected to incur significant net losses through 2025 to 2026, with profitability generally projected around 2029. Meanwhile, the combined annual capital expenditures of ultra-large manufacturers have been further increased from the hundreds of billions of dollars level in 2025.

Usage growth addresses "someone is using it." It does not automatically address "who is making money."

Between these two things stands pricing power—and that pricing power is being eroded by five or six competitors with similar capabilities.

Seven: Don't just listen to the conclusion—watch the scale.

Debating whether AI will slow down is of limited value, as both sides define "slowdown" differently.

I recommend focusing on three measurable factors.

First, the duration of the task. Continue tracking the METR curve. If the time span for a 50% success rate extends beyond eight hours within the next year, Depue’s assessment gains strong support. If it repeatedly fluctuates within the two- to four-hour range for over a year, it suggests we’ve hit a structural ceiling.

Second, reliability, not peak performance. Look at the time span at 80% or 90% success rates. Enterprises buy stability, not occasional brilliance. The rate of improvement in this metric directly determines the slope of commercialization.

Third, unit economics, not total revenue. Look at gross margins and whether the rate of decline in unit economic costs can outpace the price war. See if any company has truly achieved positive operating cash flow.

Three scales, none of which require trusting anyone's predictions.

Eight, Conclusion

My position is simple.

I agree with Musk’s assessment about capability. Models from two years ago already feel like specimens; two years from now, today’s most advanced models will likely seem clunky. This curve hasn’t bent yet.

But his response to the valuation质疑 was misplaced.

What the bears are really asking is: When models depreciate every eighteen months, when the price of equivalent intelligence drops by an order of magnitude each year, and when four or five companies simultaneously possess nearly identical capabilities, who will pay for those hundreds of billions in capital expenditures, and how will they recoup them through pricing?

This question cannot be answered by "AI will be stronger."

I remain skeptical about the RSI part. It’s not because it’s impossible, but because it requires computing power, electricity, experimental cycles, and an evaluation system to all align—and these don’t follow exponential growth.

Finally, let me say something that might not be popular.

The technological revolution is real, and so can be the bubble—these two things can absolutely happen at the same time, and historically, they often do. Railroads were real, but railroad stocks crashed. The internet was real, but the Nasdaq fell by 78%.

Being bullish on technology does not mean being bullish on every balance sheet.

Place separate bets. This isn't hedging—it's being smart with your money.

As for whether AI will be "mind-blowing" two years from now, I tend to believe it will.

I won’t bet on who’s still at the table back then.

Source reference

1. METR, "Measuring AI Ability to Complete Long Tasks," METR Research, March 2025.
2. Google DeepMind, "AlphaEvolve: A Gemini-Powered Coding Agent for Designing Advanced Algorithms," May 2025.
3. OpenAI, "Introducing GDPval," OpenAI Blog, September 2025.
4. Epoch AI, "LLM Inference Price Trends" and "Trends in Machine Learning Compute," Epoch AI Data Insights, 2024–2025.
5. MIT Media Lab / Project NANDA, "The GenAI Divide: State of AI in Business 2025," August 2025.
6. Goldman Sachs Research, "Gen AI: Too Much Spend, Too Little Benefit?," June 2024.
7. Elon Musk and Will Depue, posts on X regarding Frontier Lab valuations and recursive self-improvement, July 2026.
8. Reuters and The Information, reporting on OpenAI and Anthropic’s revenue, losses, and profitability timelines, 2025–2026.