The Road to AGI May Have Been Wrong From the Start

By Tecent Research Institute; not about China or the US is on a wrong path to AGI, but the whole humanity is pushed to run a wrong race.

This is an article about AGI that I wish I could have written. However, I am more glad that it is written by someone inside the Chinese tech industry, which shows that many debates and concerns are indeed beyond borders. It deserves a life beyond the Chinese internet — so I’ve translated it here, without the author’s permission (much as the Chinese internet translates my own work).

There is a chunk where he talks about education’s role that reads like a product pitch, and the ending is weighed down by this pitch-like emphasis on education. Unfortunately, that is the most optimistic view in the whole article, because it is the only part about what individuals can do about AGI. But the bulk of it is worth reading, and worth sharing.

For those who won’t make it to the end (which is itself a mistake), I’ve pulled out a few points in the article that speak to questions I get asked often, and the answers given in this piece are better and probably felt more authentically “Chinese.”

• How does China think about AGI?

  I’ve tried to answer this before. Of course, there is no single China view on AGI, because China does not have only one brain. Compared to my “what AGI means to China,” this article offers a more outward-facing reflection on the AGI debates happened in the US and its implications for China.

• Can China win the “AI race” with its abundant electricity?

  While China’s electricity structure is more sustainable, the author argues this doesn’t change the fundamental math. As compute demand grows exponentially, energy supply can only grow linearly. If scaling LLM is the end of the game, nowhere on earth could satisfy the appetite for compute.

• Do Chinese people feel what AI populists in the US are feeling?

  Not so much, but not because they are not “thoughtful.” There are many reasons some serious people try to reason, but the author gives another — because some things, like data centers, are not exposed to the majority’s eyes. Yet the author warns that the Chinese may soon see all the inequalities and harms that are currently concentrated in the developing world, because nothing is external. The costs being externalized onto the most vulnerable regions today will arrive closer to home tomorrow.

• Will China do a UBI?

  No — and as many others have argued, UBI couldn’t solve the problem, because work is not only about money. Capitalism has shaped society to the point where people attach meaning, fulfillment, daily rhythm, and social identity to their jobs. The author suggests that AI wealth will concentrate into an extremely steep pyramid, with UBI serving less as redistribution than as a mechanism for keeping the displaced majority quiet enough not to revolt.

I like the part where the author references Hassabis’s interview on AI is not a normal technology:

He didn’t directly say ‘we went wrong.” He used a more scientist-like formulation: “the crucial capabilities may not be on the extension of this path.” Translated into plain language: even walking this road to its end doesn’t lead to AGI; we are desperately climbing the wrong mountain.

But reality is this: the mountain is already being climbed, the industry is already burning, hundreds of billions of dollars have already been committed, the global power grid is already beginning to strain — no single player has the ability to get off this ride alone, and no single player has the ability to change direction.

This is the deepest structural predicament of this round of the AI revolution: once the direction is set by storytellers, everyone can only race faster in the same direction. 这是本轮 AI 革命最深的一层结构性困境：一旦方向被讲故事的人定下，所有人都只能在同一个方向上比谁跑得更快。

Although many brilliant minds have argued again and again in different ways that there is no one single AI race that the US and China are participating in, I fear that it is too late. I fear that the story has already become a reality, and we will soon pay the price for a game not started by us.

Picture from the original article.

Full translation by Claude, except the product pitch; some texts are bolded by me. You can also read it in Chinese [original article and archive link].

---

The Road to AGI May Have Been Wrong From the Start

Wang Peng, Senior Expert at Tencent Research Institute

Starting from the Molotov Cocktail at Altman’s Door

In the early hours of April 10, 2026, on Russian Hill in San Francisco, a young man threw a Molotov cocktail at Sam Altman’s front door. Two days later, several other young people drove past the same house and fired shots inside. Two attacks in three days — the first attacker was carrying an “anti-AI manifesto” listing the names of other AI executives; his blog was filled with writings about “AI will cause human extinction.”

This was not an isolated incident. That same month, a city councilmember who supported building a data center had his home hit by more than a dozen bullets; the perpetrator left a note: “No data center.” In a town of 12,000 near St. Louis, simply because the incumbent council had approved a data center project, every single incumbent was voted out overnight on election day. Stanford’s “2026 AI Index,” released in April, shows that the number of Americans who feel anxious about AI has surged to 64% — more than ten percentage points above the global average.

Ordinary people are expressing their dissatisfaction with this AI trajectory through every means available to them. And the most famous spokesperson for this trajectory now lives in a house that requires armed security just to sleep at night.

This isn’t even Altman’s worst predicament this year. Nearly all of his original co-founders have departed. CTO, Chief Scientist, Chief Research Officer — one after another, they’ve said goodbye. And this CEO is still telling the same story.

Telling a story has a price. That price externalizes bit by bit — becoming the Molotov cocktail at his front door, becoming the suddenly strained American power grid, becoming the rock-bottom hourly wages pressed down on workers in developing countries, becoming every country, every industry, every family being pulled into this arms race.

This article wants to articulate a judgment that many people vaguely sense but haven’t fully stated: the large language model path we are now walking toward “AGI” is very likely wrong from the very beginning. It is not a scientific choice but a narrative choice; not a technological inevitability but the locked-in result of capital and geopolitical maneuvering.

Around this judgment, the article answers four questions:

• How did we get here? Who is pushing it? Who is following?

• Where are the physical limits? How close are we to that wall?

• Who pays the price? Why are we — who seem to be “bystanders” — actually not?

• Can the beautiful story of UBI and the “AGI tax” really catch ordinary people when they fall?

This is not fear-mongering. The facts themselves are heavy enough.

The Beginning Was Wrong: How the AGI Narrative Was Manufactured

To understand today’s situation, we need to go back to 2015.

That year, Musk and Altman decided to found OpenAI at a Silicon Valley restaurant. The mission sounded beautiful — “ensure that artificial general intelligence benefits all of humanity” — non-profit, open, with safety written into the charter. But MIT Technology Review’s former senior journalist Karen Hao spent six years interviewing over ninety insiders and published Empire of AI in 2025. The book reveals an entirely different underlying logic: “must be first, or perish.”

This logic is extremely precise. It removes the choice between “doing it” and “not doing it” — if you question it, you’re helping the bad guys. It substitutes “how to do it” with “how big to do it” — since time is short, we must invest more, build bigger models, push harder on compute. It puts external critics in an unwinnable position — any suggestion to slow down is equivalent to standing against humanity.

Musk’s 2015 email to the OpenAI team said that DeepMind was causing him “extreme mental stress.” Note the emotional language — not “interesting,” not “competitive,” but “mental stress.” From day one, OpenAI’s driving force was not curiosity but fear; not science but an arms race.

Fear was the fuel Musk provided; imagination was the coating Altman added. Where he surpasses Jobs is this: Jobs used his “reality distortion field” to sell phones; Altman used the same ability to convince the entire world that we must spare no cost, use the world’s energy and compute to train ever-larger models, because this is the only road to AGI. He convinced not just investors, but governments, media, regulators, and even competitors.

This technological path didn’t have to be this way. Before OpenAI, Google DeepMind was walking a different road — using small, specialized models to solve specific problems one by one. AlphaGo cracked Go, AlphaFold cracked protein folding, AlphaGeometry cracked olympiad geometry problems. Each model targeted one domain, with controllable scale, controllable energy use, verifiable results, and interpretable behavior. DeepMind’s Hassabis had a favorite phrase: “Solve intelligence first, then use intelligence to solve everything.”

But after GPT-2 and GPT-3 appeared, the entire industry was instantly locked in. Google was forced to pivot to large models, Anthropic split off and continued betting on large models, Musk’s xAI is large models too, and even companies holding firm on specialized paths were forced to open internal large-model product lines — otherwise they couldn’t retain capital or talent. This wasn’t the result of technical validation; it was the result of game theory. As long as any single company built a bigger model first, everyone else had to follow — otherwise they’d fall behind comprehensively in valuation, funding, and talent competition. The entire industry entered a Nash equilibrium of “scale or die.”

Hao’s book contains a detail that leaves a striking impression: OpenAI originally wanted to use game AI similar to AlphaGo to attract Microsoft’s investment, but Bill Gates was more interested in the early large language model — still immature at the time — and pushed both Microsoft and OpenAI fully toward large language models. A critical inflection point in humanity’s technological path was decided by a billionaire’s personal preference and a test he gave in a conference room.

Those who determine direction are not necessarily those who understand best. Very often, they are those who tell the best stories.

Hassabis’s Dilemma: A Pure Scientist’s Compromise

Hassabis is almost synonymous with “pure scientist” in Silicon Valley AI circles. He was an international chess master at 12, studied computer science at Cambridge, did a cognitive neuroscience PhD at UCL, worked in game development, and did brain science research. His interest in AI has always been “understanding what intelligence is,” not “replacing humans.” In 2024, he won the Nobel Prize in Chemistry for AlphaFold — the first time in human history that an AI system directly contributed to Nobel-level results. At the 2026 Davos Forum, he announced that the first anti-cancer drug designed by AI would enter clinical trials that same year. This was really the path he wanted to walk: use AI to conquer cancer, conquer Alzheimer’s, conquer problems humanity has been unable to solve for centuries.

But in the past two years, what he says in public interviews has become increasingly contradictory.

In 2024, he stated directly that viewing AI as an “ordinary technology” is wrong — because what it impacts is not physical labor, nor information distribution, but cognition itself. He has cited the concept of “jagged intelligence”: current large models have an extremely uneven distribution of capabilities — able to win gold medals at the International Mathematical Olympiad, yet stumble on elementary arithmetic. He explicitly listed four structural shortcomings of the current path: inability to do long-term planning, no continuous learning (the model is “frozen” after training), no genuine creativity (can solve problems but can’t pose them), and severely jagged capabilities (impossible to predict where things will go wrong). His timeline to AGI is 5 to 8 years — far more conservative than Altman or Amodei.

His conservatism comes not from lack of understanding but from the deepest understanding. With a background in cognitive neuroscience, he knows that many functions of the human brain — long-term memory, goal planning, metacognition, genuine curiosity — cannot automatically emerge just by piling parameters up to a trillion.

Here’s the question: if he so clearly sees the ceiling of scaling, why is he still racing along?

This is actually the most poignant aspect of this whole affair. In interviews, Hassabis reveals the same underlying feeling — he has no choice:

• If DeepMind doesn’t build Gemini, Google’s search business will be upended by ChatGPT

• If Google’s search business is upended, DeepMind won’t get the compute needed to train the next generation of AlphaFold

• If it can’t get compute, the “use AI to defeat disease” work he’s wanted to do all his life won’t be possible

• So he must first do something he doesn’t most believe in (stacking large models) in order to exchange it for the resources to do what he most believes in (defeating disease)

This is a kind of locked-in rationality. The narrative of “must lead or face destruction” that Altman manufactured has locked in not just investors and the public, but genuine scientists too. Hassabis has become a Nobel laureate leading the race on the scaling track, where every day his best strategy contradicts the path he has long believed in.

In an interview in April 2026, the host asked him where the limits of this path lie. He paused for an unusually long moment, then said something to this effect: When we founded DeepMind in 2010, we envisioned a path inspired by neuroscience — specialized, built up layer by layer. The path we’re walking now is different. Some capabilities have indeed emerged, but at great cost, and the most crucial capabilities — genuine planning, genuine creativity, continuous learning — may not be on the extension of this path.

He didn’t directly say “we went wrong.” He used a more scientist-like formulation: “the crucial capabilities may not be on the extension of this path.” Translated into plain language: even walking this road to its end doesn’t lead to AGI; we are desperately climbing the wrong mountain.

But reality is this: the mountain is already being climbed, the industry is already burning, hundreds of billions of dollars have already been committed, the global power grid is already beginning to strain — no single player has the ability to get off this ride alone, and no single player has the ability to change direction.

This is the deepest structural predicament of this round of the AI revolution: once the direction is set by storytellers, everyone can only race faster in the same direction.

The Physical Ceiling: Energy Cannot Keep Up with Compute

This path has another boundary that is harder and closer — physics.

On the compute side, things have already gone wild. Over the past decade, the training compute used by the most advanced models has roughly doubled every three to four months — seven times faster than Moore’s Law. Over the past three years, an even steeper curve has piled on top of the inference side. Once Agents are widely deployed, per-task token consumption jumps from a few hundred to tens of thousands. Every workflow taken over by AI swaps what used to be individual Google searches for inference streams burning tens of thousands of tokens per second.

On the energy side, however, growth remains linear. The IEA’s 2025 report made a judgment now cited around the world: by 2030, global data center electricity consumption will more than double from current levels — equivalent to Japan’s entire annual electricity consumption. The PJM power grid covering America’s East Coast — 13 states, 65 million people — for the first time in early 2026 could not assemble sufficient electricity capacity and urgently needed new power equivalent to more than a dozen nuclear plants; in Northern Virginia, home to the world’s largest data center cluster, the queue time for grid connection has stretched from two or three years to six or seven. Altman and Musk have frequently talked about “nuclear fusion” and “space data centers” in recent years — not out of sentiment, but because there isn’t enough electricity left on Earth, and they need to keep telling these stories so investors will keep spending.

Many will say: America’s power shortage doesn’t matter — China has abundant energy, so we can win. This judgment is half right.

The correct half: China’s electricity structure is indeed healthier. In 2024, China’s newly installed wind and solar capacity exceeded all other regions of the world combined. The lower Yarlung Tsangpo River hydropower project, which broke ground in 2025, will reach installed capacity at the 60 GW level — annual power generation exceeding three times the Three Gorges Dam, equivalent to dozens of nuclear plants materializing at once. Domestic opinion widely reads this as “China’s clean energy trump card prepared for future AI compute”.

The incorrect other half: even so, if we treat “stacking LLMs” as the endgame, the math still doesn’t work out. The reason is simple — compute demand grows exponentially, energy supply grows linearly. The Yarlung Tsangpo project won’t be fully complete until 2033, and at the current pace of compute doubling every few months, those 60 GW may well be consumed by a few first-tier cities’ AI clusters before 2033 arrives. Even adding all of China’s under-construction solar, wind, nuclear, and pumped hydro — with hundreds of gigawatts of new installed capacity per year — it still can’t keep pace with the compute appetite once Agents become truly widespread.

An exponential function versus a linear function — which catches the other first? The exponential always wins. This is not a question of the East-West gap [Zilan’s note: probably refers to the US-China gap rather than the East-West as in “East Data West Compute”] but a matter of math and physics.

There is also Jevons’ Paradox to contend with: declining unit cost does not equal declining total consumption. The price per token has dropped a thousandfold in three years, yet global total spending has multiplied several times — meaning total consumption has exploded by orders of magnitude. Every time inference gets cheaper, it opens up an entirely new category of applications; every time it drops one level, demand rises ten levels. Once everyone has a personal AI assistant and every job has a fleet of Agents running in the background, per-capita inference consumption will be orders of magnitude higher than today.

This is the true limit of this path: physical limits are closing in while narrative momentum continues to accelerate. It’s not that algorithmic breakthroughs are impossible, but that electricity can’t keep up. Electricity prices, data center costs, and training costs will spike first. In the end, only a very small number of players will be able to keep playing.

Who Bears the Cost: An Extremely Unequal Supply Chain

Beyond energy, there are people. The cost of this path has never only been the watts extracted from the power grid — it also includes time and resources transferred from the most vulnerable people and the most vulnerable regions. The two fuels sustaining the AI empire’s prosperity are electricity and people.

Karen Hao spends nearly half of Empire of AI tracing this “hidden supply chain” — from the data labeling factories of Nairobi, to the content moderation centers of the Philippines, Venezuela, and Colombia, to water rights hearings in Santiago, Chile, to protest scenes in Uruguay’s capital. She cites a Stanford researcher’s framing question:

“Don’t ask how AI can do good — ask how AI has changed the power structure. Has it consolidated power, or redistributed it?”

Let me tell you about a detail present in something you use every day, one you may have never thought about.

ChatGPT makes you feel it is “clean, intelligent, and polite” — not because it is innately that way, but because someone in a place you cannot see has already watched the darkest things first, on your behalf.

Between 2021 and 2022, OpenAI outsourced the dirtiest part of ChatGPT’s training — identifying violence, hate speech, extremist content, and suicide promotion — to a company called Sama, located in Nairobi, Kenya. Labeling workers earned $1.30 to $2.00 per hour. OpenAI’s contract with Sama was priced at $12.50 per hour, with most of the $10 gap eaten by management fees and middlemen. Workers had to read hundreds of dark content fragments per day; many were diagnosed with PTSD. In 2023, two former employees petitioned Kenya’s parliament to investigate OpenAI’s worker treatment practices. The union representative’s description: this kind of contract is “worse than modern slavery.”

This is just the tip of the iceberg. Similar invisible labor armies are spread across the Philippines, Venezuela, India, Pakistan, and Uganda. Their hourly wage might not cover a meal, but the models they trained pushed OpenAI’s valuation into the hundreds of billions.

What is more worth noting: the AI these labeling workers trained for survival will eventually turn around and erase their own jobs — call centers, translation, entry-level customer service, text moderation. These are precisely the white-collar entry points that developing countries painstakingly built up over the past twenty years. A headline in Kenya’s Daily Nation was bitterly apt: “We are teaching AI how to put our children out of work.”

Then there is water. Large AI data centers don’t just burn electricity — they burn water, using evaporative cooling for temperature control. A hundred-megawatt data center evaporates each day an amount of fresh water equivalent to the daily water use of a town of ten thousand people. [Zilan’s note: I think the author's framing is somewhat taking certain US existing advocacy at face value, probably because this particular debate (especially the ones arguing against water usage) gets much less coverage on the Chinese internet. The core concern that data centers impose real water costs on water-scarce communities is legitimate and documented. But the specific figures depend heavily on how the baseline is defined, and the piece treats evaporative cooling as universal when many newer facilities use closed-loop or air-cooling systems. I think partly this is because the water/data center debate in China is so little, and partly because he relies too heavily on Karen Hao’s book.]

In 2019, Google announced a $200 million data center in Santiago, Chile, with annual water consumption equivalent to a thousand times the local community’s annual water use, drawing directly from the municipal water supply. The problem: Chile was experiencing its worst central drought in nearly a thousand years, with reservoir levels down to twenty percent of historical averages. On one side, residents rationing toilet flushes; on the other, GPU clusters cooled around the clock with drinking water.

Local residents fought the case in environmental courts and in the national president’s office for five full years. In 2024, Chile’s Environmental Court ordered Google to reassess the impact on the main aquifer; Google announced the project was suspended. An April 2026 report confirmed the original plan was formally scrapped and the entire project sent back to the drawing board.

Almost simultaneously, Uruguay’s capital saw similar protests. In 2023, Uruguay suffered its worst drought in seventy-four years; the capital’s tap water became too saline to drink. The slogan protesters chanted in the streets was blunt: “This is not drought, this is pillage.”

Latin America, North Africa, Southeast Asia — similar community uprisings have occurred in at least dozens of locations. These places share common features: arid climate, cheap land and electricity, local governments eager for foreign investment, and regulatory processes that can be worked around. The public pitch is always “creating jobs” — but once a data center is built and operating normally, routine positions number only in the dozens to a hundred, most of them highly skilled roles that local residents cannot access.

This is the most typical pattern of the AI empire’s interaction with developing countries: promise the future, take the present; promise jobs, take the water.

Some might ask: what good does seeing any of this do? Ordinary people can’t help the labeling workers in Nairobi or solve the water shortage in Chile.

The reason is simple: this supply chain is the foundation of this AGI path, and in a sense, it is also the source of the technology gap between China and the United States.

If we don’t see those young people working overtime in Nairobi at three in the morning, we’ll think ChatGPT’s “politeness” is free. If we don’t see the residents of Santiago fighting five years of lawsuits, we’ll think cloud services come without cost. If we don’t see the villages next to data centers on the Gobi Desert that still occasionally face rolling blackouts, we’ll think AI’s environmental costs can be offset by carbon credits. When costs are hidden, decisions continue to go wrong.

Altman can say “AGI will benefit all of humanity” because he only considers a small portion of “all of humanity” as people; the rest are treated as raw material. He can say “AI will create enough wealth to give everyone a UBI” because he only counts the profits on the Silicon Valley end and not the losses on the other end of the supply chain. Once the full cost is laid out, this story cannot hold together.

There is also a more direct connection — the costs transferred onto the most vulnerable regions today will be transferred onto us tomorrow in new forms. The low wages of labeling workers, the drained water resources, the overloaded power grids — these are not “someone else’s problem.” They are where this path has already completed its stress tests in the softest places first. The next step will inevitably come to us:

• Your child may become a “middle-class labeling worker” — managed by AI while labeling data for the next generation of AI [Zilan’s note: here, I think the argument is weighed down by his pitch on an education platform later. But I keep this line because it makes sense to me nevertheless.]

• Your city may begin debating “data center water priority” versus “residents’ water priority”

• Your electricity bill may quietly double as a nearby compute cluster expands

Seeing other people’s today is seeing your own tomorrow.

Why UBI Cannot Hold

What solution do Altman and Musk offer? Universal Basic Income (UBI). AI will create enormous wealth, redistributed through UBI; people won’t need to work to live decently, and the time freed up can go toward creativity, companionship, and self-actualization. It sounds beautiful.

But economically, this story largely doesn’t hold up.

First layer: Where does the money come from?

Altman’s answer is “tax AI companies with an AGI tax.” Sounds reasonable, but AI profits are concentrated in a handful of companies, all of which have extremely strong cross-border tax avoidance capabilities — Ireland, Singapore, Switzerland, the Cayman Islands, take your pick. More critically: in the current context of great-power geopolitical competition, would the US government really dare to go after OpenAI and NVIDIA — “national strategic assets” — with a big knife? The moment such a tax is seriously proposed, companies can immediately threaten to “relocate to a more friendly jurisdiction.” The so-called AGI tax is, in essence, more of a story told to the public than a real policy.

Second layer: What does the empirical evidence show?

Altman’s own OpenResearch project gave $1,000 per month to a thousand low-income Americans for three years. When results were published in 2024, they showed that recipients experienced short-term improvements in well-being, but their working hours declined, and their savings rates and long-term income showed no significant difference from the control group. Critics publicly argued that “this experiment is not actually a test of UBI at all, and anyone claiming it is either misunderstands or misleads” — because it did not come close to simulating the large-scale unemployment scenario that AGI is supposed to produce.

The reality is: UBI can make impoverished populations live a bit more comfortably, but it cannot solve “the collapse of meaning that follows large numbers of middle-class people finding their cognitive capabilities fully covered by AI.” People do not only need money — they need the sense of meaning, identity, social networks, and life rhythm that work provides. A forty-year-old former financial analyst who discovers that a decade-plus of professional expertise has been fully superseded by a model in two seconds — his pain is not something a thousand dollars a month can solve.

Third layer: Polarization will intensify, not ease.

As discussed above, AI’s costs are transferred onto the most vulnerable countries and people. The same structure will replicate inside every country. AI wealth will grow into an extremely steep pyramid — at the top, a small number of model companies plus upstream chip makers and hyperscale cloud providers; in the middle, a small number of top cross-disciplinary talents; at the bottom, labeling and moderation workers pressed to the survival line; and below that, people fully displaced with no path to reskill, surviving on UBI.

AI is concentrating knowledge production, resource accumulation, and social influence into the hands of an extremely small number of people at a pace unprecedented in history. The true function of UBI in this landscape may not be “sharing wealth” — it may be “maintaining order” — ensuring that the replaced majority have enough to eat and won’t revolt, so that the few at the top can continue to quietly claim the surplus profits AI delivers.

The Molotov cocktail at Altman’s door is just one of the earliest echoes of this imbalance of order.

Human society does, of course, have buffer mechanisms — law and regulation, unions and labor protections, human-machine collaborative design, tax redistribution — and all of these matter. But we should be clear-eyed: these can only slow the slope of the displacement curve, not reverse it. Once AGI possesses the characteristic of “general cognitive capability,” it is unlike all previous technologies. Every previous technology only replaced one specific human capability; people could evade displacement by learning new capabilities. AGI replaces “the capability to learn and apply new capabilities itself.”

This is the first time in human history that we face a revolution with a “capability ceiling.”

One Ordinary Parent Did Something with AI

[Omitted. He mainly talked about his education product, but you can read the original version in the links. ]

Epilogue: We Cannot Go Back, But We Still Have Our Children

[Omitted the last few paragraphs, because they circled back to education, which I think is purely written for his product. You can still find them in the original links.]

Back to Hassabis one last time. Near the end of that interview, the host asked him: “Do you worry about these things at night?” He was silent for a moment, then said — he reserves the time from midnight to 3 a.m. each day to think about these questions. Not reading news, not attending meetings, but pure thinking — thinking about what intelligence is, thinking about where humanity should go.

That detail carries weight.

We no longer have any way to turn back the clock or make the entire industry walk a different road. Energy will be depleted, compute will approach physical limits, and polarization will be more severe than in any previous period in history.