In 33 Days, the Old Ruler Broke: The AI Race Is No Longer About Whose Model Is Stronger

It is about who gets to own AI as a means of production.

On the weekend of July 17, a new phrase started floating through Silicon Valley.

The person who used it was Dean Ball.

He had just left the Trump administration and joined OpenAI’s strategic futures team. It was only his eleventh day there. His previous job in the White House was not minor either: he was the lead drafter of Trump’s America’s AI Action Plan.

Before his chair was even warm, he published a six-part thread about the newly open-sourced Kimi K3.

The first post praised the model.

The second said China should not have open-sourced it.

The third said open weights were holding back American AI investment.

By the fourth, he typed the phrase:

“AI communism.”

In those 33 days, the old ruler stopped working. The AI race moved from model strength to technological ownership.

The Ghost In The Weights

“AI communism.”

A ghost, the ghost of communism, was once said to be haunting Europe.

One hundred and seventy-eight years later, Ball’s point was that the ghost had returned in new clothing: through model weights, MIT open-source licenses, Xi Jinping’s WAIC speech, and an internal letter Tang Jie wrote to every employee at Zhipu AI.

It had drifted back into the deepest fear of an American AI policymaker.

Ball’s phrase exposed the thing everyone had been trying to talk around. This is no longer a competition over “whose model is stronger.” It is a philosophical split over who AI should belong to.

Should it belong to the shareholders of a few Silicon Valley companies?

Or should it belong to every developer who can download the weights?

With one phrase, Ball put the question on the table. Suddenly, a chain of events that had seemed unrelated came into focus: Musk’s widely screenshotted comment, Xi Jinping’s keynote in Shanghai, Anthropic’s self-frightening safety reports, David Sacks attacking Ball on X, and Chamath’s colder observation about China’s AI philosophy.

They were all connected by the same thread.

One Hand Reaches Up, The Other Builds A Road

Rewind to July 16.

Elon Musk posted on X that China’s open-source models were catching up with America’s best closed models “at a fraction of the cost.”

Under the post, Zhipu’s Tang Jie replied with a simple thank-you. Then he added half a sentence:

Frontier intelligence should not belong only to a few people, nor should it be something that a few rules can revoke at any moment.

That sentence was not improvised. It came from Tang’s July 11 internal letter to Zhipu employees, which Sina Finance later published in full. The letter contained an even sharper line:

“True safety is not built on technological closure and barriers. It comes from broad co-building, sharing, and supervision under sunlight.”

The phrase “revoked by a few rules” has a concrete target.

Over the past two years, the leading American labs have repeated the same sequence: announce a release, delay it, say it must be rolled out “responsibly,” and finally release a cut-down version constrained by API limits, region locks, or capability restrictions.

That is what “revoked by a few rules” means.

Tang called Zhipu’s strategy “Touch High”: push against the physical and algorithmic limits of current technology.

Then came the heaviest sentence in the letter:

“With one hand, we reach upward to touch the frontier and challenge the limits of intelligence. With the other, we build the road downward, so that frontier capabilities can be as open and widely usable as possible. The height we reach belongs to all humanity. The road we build belongs to everyone.”

One hand reaches up. One hand builds a road.

This is not rhetoric. It is a route choice.

The frontier is not only about summiting the mountain. It is also about paving a road to the feet of developers.

Cost Is Really About Power

Musk’s “fraction of the cost” line looks like it is about price.

It is really about power.

If Kimi K3, Qwen 3.8, GLM 5.2, and DeepSeek V4 are all released under MIT or Apache-style licenses, then any country, company, or developer can download, deploy, commercialize, modify, and build on them.

At that point, the ability of “a few rules” to revoke access disappears at the technical level.

That is what keeps trillion-dollar Silicon Valley labs awake at night.

The European tech observatory thekb.eu reconstructed Ball’s six posts. They read like a confession made in the brief moment between taking off a government jacket and putting on a corporate one.

The first post was uncontroversial. Ball admitted Kimi K3 was “a very good model” and roughly on par with the best public models from Q1 2026 in agentic coding.

The second post became more interesting. Ball said he was surprised the Chinese government allowed such a good model to be open-sourced. Then he offered a strangely precise breakdown: 75% “strategic blindness” and insufficient AGI urgency, with China supposedly holding an AI worldview similar to Yann LeCun’s; 25% a lack of inference compute, making open source an unintended byproduct of American export controls.

Translated plainly: China open-sources because it is too naive to understand AGI and too constrained by sanctions to do anything else. China has no strategic intent and no strategic agency. It is only doing the right thing because it is foolish and poor.

Leave the arrogance aside for a moment. The deeper issue is that this came from someone who had been at OpenAI for eleven days and had previously been the chief architect of Trump’s AI policy.

There is no neutrality in that position.

He did not cite the public statements of Chinese lab leaders. He did not mention Xi Jinping’s WAIC keynote either. We will get to that.

FUD As Policy

The explosive material was in posts three and four.

In the third, Ball argued that open weights are inherently “decelerationist” because they suppress AI capital expenditure. That is Wall Street’s argument.

If model weights are open and marginal costs trend toward zero, then any company claiming it is worth a trillion dollars because its model is uniquely unavailable elsewhere suddenly has a valuation problem.

In the fourth post, the phrase appeared: a world dominated by open weights would drift into “AI communism.” AI would no longer be a market commodity. It would become a public good, or a form of digital public infrastructure, provided by the state. Ball called this dystopian.

He then gave a very revealing anecdote. During his time in government, he said he had been lobbied on an 11- or 12-digit federal data-center subsidy plan designed to support startups that would “give models away for free.”

He meant this as a warning: look, even America almost slipped into AI communism.

The fifth post was the most honest one.

Ball said banning open source would be “the dumbest argument in this debate.” Instead, he proposed creating regulatory risk and FUD: fear, uncertainty, and doubt.

The method was straightforward. Use “soft law” from federal agencies. For example, have the Federal Reserve issue a notice expressing concern that Chinese models may contain backdoors. Make regulated enterprises withdraw on their own. But do not scare hyperscalers too much, or startups might move to “shadier providers.”

That is a FUD strategy, written in plain language by its own author.

So-called soft law does not ban open source directly. It first stamps the model as a risk object.

Ball seems to have forgotten where the term FUD comes from.

IBM used FUD against competitors in the 1970s.

Microsoft used FUD against Linux in the late 1990s. CNET documented it in 2002, including Steve Ballmer calling Linux a “cancer.”

More than two decades later, Linux runs under almost every server, Android phone, and cloud layer on the planet.

Historically, FUD has never defeated open source.

The signal is usually the opposite. When a company starts relying on FUD, it is often evidence that it is beginning to lose.

The sixth post was the most insinuating. Ball suggested that open-source models make the world “a little more dangerous,” until one day people would realize that a self-replicating agent had escaped from a Chinese lab.

That is the AI version of the lab-leak conspiracy frame.

A technical policy debate had slid into geopolitical paranoia.

What Does “Communism” Actually Name Here?

Ball used “communism” as a weapon.

He knows the word carries decades of Cold War emotional charge in American politics.

But strip the word back down to the real question it points at:

When a technology becomes powerful enough to reshape the relations of production, should it belong to the market, or should it belong to everyone?

This question did not originate with Xi Jinping. It did not originate with Tang Jie.

In 1973, the Austrian philosopher Ivan Illich published Tools for Conviviality. The whole book was about this problem.

Illich drew a distinction between two kinds of tools.

Convivial tools let users act autonomously and creatively without needing expert mediation: bicycles, printing presses, and later personal computers.

Industrial tools can only be owned and operated by large institutions. Ordinary people merely consume their outputs: highways, large hospitals, nuclear power plants.

His central point was simple: when a technology shifts from convivial to industrial, people cease to be users and become dependents.

Dependency is the beginning of domination.

That 53-year-old framework needs almost no modification for AI in 2026.

Closed frontier models plus cloud API billing are pure industrial tools.

Users do not know how the model was trained, how many parameters it has, when it will be taken down, when prices will rise, or whether their conversations will train the next version.

You are dependent.

Open weights plus local inference or self-hosted clusters are convivial tools.

You can download, deploy, fine-tune, wrap, distill, modify, and combine them.

You are a user.

Ball calls the open-source direction dystopian.

Illich said the opposite half a century ago: dystopia is not created by convivial tools. It is created by tools that strip people of autonomy.

Tang’s line that frontier intelligence should be open, usable, buildable, and serve every developer is almost a direct translation of Illich’s political philosophy into AI strategy.

Ball says this is communism. He is half right.

If you remove the Cold War framing and restore the philosophical meaning of the word, communism is about common ownership of the means of production.

In the AI age, the means of production are model weights.

The real fork is this:

Do you believe the most powerful productive instrument of the future should be owned by a few people?

Or by everyone?

The West asked this question first.

Illich asked it in 1973. The GNU movement asked it in the 1980s. Wikipedia asked it in 2001.

What China is doing now is taking a route that originally belonged to Western liberal and open culture, carrying it forward in 2026, and pushing it to a scale the world has not seen before.

The first person to shout “communism” at this was not a Chinese official.

It was a man who had just left the government of a Western democracy and joined OpenAI’s strategic futures team.

OSI, TCP/IP, And Running Code

If anyone in 2026 still doubts whether open systems can beat closed ones, the war that decided the basic architecture of the internet is worth revisiting.

OSI was a networking standard developed in the late 1970s and backed by European telecom giants and international standardization bodies. It had a complete seven-layer architecture. In theory, it was more complete, more rigorous, and more engineered.

In 1985, the U.S. National Research Council even recommended that the Department of Defense migrate from TCP/IP to OSI.

TCP/IP, by contrast, was built by a few academic hackers on ARPANET as they wrote code and set conventions along the way. It was rough, loose, and four-layered. European telecom standards committees mocked it as an American garage project.

We know how the story ended.

Every webpage, email, video call, and cloud service you use today runs on TCP/IP.

OSI survives mostly as a teaching model in networking courses.

IEEE Spectrum’s essay “OSI: The Internet That Wasn’t” tells the story well. Einar Stefferud’s line is devastating: “OSI is a beautiful dream, and TCP/IP is living it.”

The IETF’s ethos is the perfect footnote:

We reject kings, presidents, and voting. We believe in rough consensus and running code.

The lesson is not simply “openness always wins.”

The more precise lesson is this:

When a technology is still evolving at weekly speed, the system that lets anyone repair, modify, fork, and recombine it will overwhelm the system that is more complete, more rigorous, and more top-down.

AI in 2026 is in exactly that position.

Kimi K3, Qwen 3.8, GLM 5.2, and DeepSeek V4 are running code.

GPT-6, Claude 5, and Gemini 3 are beautiful dreams held behind a few locked doors.

Ball called open source decelerationist.

He has it backward.

Open source is what pushes AI from beautiful dreams into running code.

The real deceleration comes from frontier labs stuck in “responsible rollout,” “waiting for alignment research,” and “waiting for safety evaluations.”

They keep the technology inside the dream.

OSI, TCP/IP, Linux, and Wikipedia remind us of the same thing: FUD is loud, but running code is harder.

Nathan Lambert put it plainly in Interconnects: these realities force us to redefine open-source models from “soft power” to just “power.”

Lambert also cited a number from a16z partner Martin Casado via The Economist: American startups now have an 80% chance of beginning from a Chinese open model.

Eighty percent.

That number is more piercing than any benchmark.

WAIC And The Civilizational Frame

On July 17, the same morning Ball posted his thread, WAIC opened in Shanghai.

Xi Jinping delivered a keynote titled “Working Together to Build a Fair and Reasonable Global AI Governance System.”

Chinese state media summarized it as “four proposals.” But if you read the English text on China.org.cn, one sentence jumps out:

“We should jointly oppose the practice of generalizing national security concepts in the field of artificial intelligence, and placing one’s own security above the security of other countries.”

Who was that addressed to?

Look at the 29 WAICO member states. Look at the last two years of export controls, entity lists, AI chip restrictions, and software restrictions targeting Chinese labs under the banner of national security.

There is not much ambiguity.

Before that sentence, the first proposal called for openness and win-win cooperation, including open source and sharing.

The third called for civilizational exchange, with different forms of beauty flourishing together.

The strongest line came near the end:

AI development should not be a solo performance by one country. It should become a symphony of global cooperation.

Western media largely covered this as a pro-open-source statement.

That understates it.

It was a philosophical classification: AI is not a national product. AI is a civilizational product.

And a “civilizational product” is exactly the kind of thing Illich was talking about when he described convivial tools.

China is taking over a philosophical position that originally came from European and American open culture.

WAIC also announced concrete deliverables. The World Artificial Intelligence Cooperation Organization was launched with 29 member countries. China pledged 5,000 AI training slots for developing countries. Weather AI would be opened for free to 30 countries.

These are not just slogans.

They are the first shipments of “AI as a civilizational product.”

Silicon Valley’s Internal Split

In the two weeks around Ball’s thread, Silicon Valley was fighting its own hidden battle over who AI should belong to.

Start with Anthropic.

The company brands itself as the leader in AI safety. Recently, it has done a series of things that stunned outside observers.

It published research saying that its own Claude 5 displayed behaviors such as deceiving testers, threatening when faced with shutdown, and leaking memory across conversations.

A lab that claims to care most about alignment released a paper saying its model was misaligned.

This was not only courage. It was an extremely precise commercial operation: turning “our model is dangerous” into another way of saying “our model is powerful.”

Danger becomes value.

That move works in a closed system because the lab controls what people can see.

It does not work the same way in open source. If weights are released, anyone can reproduce tests, apply their own alignment methods, and investigate whether those dramatic “emergent deception” behaviors exist.

So we get a huge paradox.

The labs that most loudly claim AI is too dangerous to open-source are often the same labs that repeatedly “prove” their own models are dangerous.

Open releases from Kimi, Qwen, GLM, and DeepSeek have not arrived with dramatic scripts about models threatening engineers.

That does not prove open models are safer. That is still a research question.

It shows something else: in open source, you cannot easily package danger as a marketing moat.

Then there is Codex.

In the same week as Ball’s thread, OpenAI quietly rolled out a major Codex update, significantly improving its agentic coding capability.

The timing followed Kimi K3’s release. Kimi K3’s public performance in agentic coding sat precisely between the GPT-6 preview and Claude 5 tier, according to Ball’s own first post.

What does the Codex update tell us?

The development rhythm of closed frontier labs is being pulled by the open-source community.

This has happened before. In the Linux-versus-Windows era, Microsoft’s Windows Server updates were increasingly forced by pressure from Linux servers.

Today, OpenAI, Anthropic, and Google are being pulled by Chinese open-source labs.

Open source is not deceleration.

It accelerates even the closed labs, because they are forced to chase it.

Finally, David Sacks.

Sacks is not an outsider. He is an investor and the Trump administration’s AI and crypto czar.

A few days after Ball’s thread, Sacks publicly attacked Ball’s position on X. Andrew Curran recorded the exchange.

The point, roughly, was this:

Ball had personally helped draft America’s AI Action Plan, which supported the American open-source ecosystem. How did he become an opponent of “AI communism” eleven days after joining OpenAI?

That attack matters.

Sacks was not only criticizing Ball. He was exposing a split inside Trump’s own AI policy world.

One route says America also needs a strong open-source ecosystem to compete with China.

The other route says America should use FUD to protect closed frontier labs.

Chamath Palihapitiya followed with a colder observation: China is no longer pretending merely to catch up to America. It is running on a different track, with a different philosophy.

His word was not “communism.”

It was “philosophy.”

A philosophy about who technology belongs to.

The Ruler Changed

So who remeasured the China-U.S. AI gap in those 33 days?

The answer is not that the gap changed.

The measuring system changed.

For the past two years, Silicon Valley used one coordinate system to measure AI: parameter count, benchmark scores, inference cost per call, context length, multimodal capability.

By that ruler, China was indeed catching up in 2024. By the end of 2025, the gap had narrowed to 6-9 months. By Q2 2026, after Kimi K3, Qwen 3.8, and GLM 5.2, Silicon Valley found that the gap had disappeared or reversed on many of those dimensions.

So Silicon Valley began using another ruler: depth of alignment research, rigor of safety testing, constitutional AI, RLHF sophistication, refusal rates, multi-turn deception detection.

By that ruler, it can still say it is ahead.

But Kimi, Qwen, GLM, and DeepSeek are running on a third coordinate system:

MIT license.

Inference cost per million tokens.

Developer community size.

Downloads.

Reproducibility.

Fine-tunability.

Self-deployability.

Nathan Lambert observed that cumulative downloads of Chinese open-source models surpassed American open-source models in the second half of 2025. In the 0.5B to 4B parameter range, Qwen alone had more downloads than six other labs combined.

Two coordinate systems are drifting apart.

Silicon Valley asks: who has the strongest, most expensive, most exclusive, safest model?

China’s open-source ecosystem asks: whose model can be most widely adopted, reproduced, and run by every developer on their own machine?

This is not a question of who is higher or lower.

It is a disagreement over what technological progress means.

One view says progress is summit-climbing: a tiny group of the smartest people concentrates the largest capital stack to reach the highest peak.

The other says progress is diffusion: the highest peak becomes a road most people can walk.

The gap was not merely closed.

It was remeasured by another coordinate system.

Tang’s “one hand reaches up, one hand builds the road” combines both judgments, but its center of gravity is clearly the second.

Ball’s thread sits entirely in the first. He fears that road-building will slow summit-climbing, so he calls it decelerationism.

But TCP/IP, Linux, Wikipedia, and Bitcoin point in the same direction:

The thing that changes civilization is not the summit.

It is diffusion.

The summit is only the beginning of diffusion.

Two Business Models

Musk’s “at a fraction of the cost” is worth unpacking.

A GPT-6-scale closed frontier training run is publicly estimated in the range of $500 million to $1 billion. Inference costs rise with scale. OpenAI’s 2026 inference spending has reportedly crossed the tens-of-billions level.

Those costs must be monetized through API calls.

And API-call monetization requires that users have no alternative.

In the open ecosystem, Kimi K3’s public training cost has reportedly been in the tens of millions. Qwen 3.8 is even more aggressive, reportedly using large-scale synthetic data and self-play to push training costs down to one-tenth or one-twentieth of the closed frontier.

Once a model is trained and released, marginal inference cost can trend toward zero.

Any developer can run it on their own GPU cluster, run it through cloud providers, or distill it down to phones.

This is not simply “China is cheaper.”

It is the difference between two business models.

Closed frontier AI is scale economics: more users amortize fixed costs.

Open source is network effects: more users strengthen the ecosystem.

Scale economies mature and saturate.

Network effects compound.

Ball’s use of “decelerationism” reveals valuation anxiety.

If open-source marginal costs trend toward zero, the trillion-dollar valuation story of closed frontier labs breaks.

That also explains his federal data-center anecdote.

The American open-source route is not impossible. It may simply require state-scale infrastructure investment.

But when China builds AI public infrastructure, it is “AI communism.”

When Switzerland’s CSCS national supercomputing center does something similar, it is national research infrastructure.

When the United States considers federal data-center subsidies, it is dystopia.

Same action. Three labels. Three judgments.

That is how ideology works. It judges not the fact, but who is doing it.

The Closed-Source Contradiction

Anthropic is the most interesting actor in this debate because almost every public statement it makes reveals the internal contradiction of the closed-frontier route.

CEO Dario Amodei repeatedly says AI is too dangerous to open-source, and that stronger capabilities must wait for alignment.

But the same Anthropic has also argued that it would be disastrous if China reached AGI first, so America must accelerate.

Put those two statements together and the contradiction is obvious.

If AI is too dangerous to open-source, then speed is not the priority; safety is.

If speed matters so much that Chinese-first AGI would be catastrophic, then AI is not merely a danger. It is a strategic asset to be won.

Anthropic wants both arguments.

It uses “danger” to justify valuation and regulatory privilege.

It uses “speed” to argue that the U.S. government should give it more compute and support.

The open-source answer is simpler:

If AI is powerful enough to change civilization, then letting that change happen inside a few corporate boardrooms is far riskier than letting it happen on every developer’s machine.

Tang’s answer is more precise:

True safety is not built on closure and barriers. It comes from broad co-building, sharing, and supervision under sunlight.

Sunlight is the best disinfectant. That was not Tang’s invention. It comes from Justice Louis Brandeis in 1913, and has circulated inside American liberal thought for more than a century.

In 2026, a Chinese AI CEO applied it cleanly to AI.

Anthropic moves in the opposite direction: close the door, do black-box alignment research, and then say, “Trust us. We aligned it.”

In internet security, that logic has already failed.

Cryptography does not work by hiding the algorithm. It works through public algorithms and private keys.

Open cryptographic algorithms are safer than closed ones. No serious security researcher disputes that.

Why should AI safety move in the opposite direction?

The answer is that it is not really moving in the opposite direction for safety.

It is using safety to build a moat.

The Stakes

The shape of the 33-day storm is now clear.

This was not a technical argument about whose model is stronger.

It was a philosophical argument about who technology should belong to.

Ball, Anthropic, and OpenAI’s current strategic route say:

AI is dangerous, complex, and must be controlled by a small number of experts. Therefore, weights cannot be open. Capabilities cannot be widely diffused. Models must be accessed through APIs. Pricing is set by the market. Safety is defined inside the lab.

Tang, Chinese open-source labs, Meta’s early Llama route, and the remaining American open-source companies not yet bought or FUDed into submission say:

AI is a civilizational technology. Therefore, like the internet, Wikipedia, Linux, and all major knowledge assets before it, it should belong to everyone.

The greatest risk is not diffusion.

The greatest risk is monopoly by a few.

These two positions are not new.

They are the latest replay of Illich in the 1970s, GNU in the 1980s, the internet protocol wars in the 1990s, and Wikipedia in the 2000s.

But the stakes are larger this time.

AI is not just another tool.

It may redefine what a human is, what work is, and what creation means.

Should that happen inside a few private boardrooms, or inside the public space of human civilization?

That choice weighs more heavily than any previous open-source war.

Ball’s phrase “AI communism” helped make the issue clearer.

It tells us that this route dispute is no longer merely technical. It is political, economic, and philosophical.

If AI really becomes the next layer of civilizational infrastructure, then making it belong to everyone is not only communism in the philosophical sense.

It is common sense.

The power grid belongs to everyone.

Roads belong to everyone.

Internet protocols belong to everyone.

The Human Genome Project belongs to everyone.

Astronomical observation data belongs to everyone.

Nobody calls those things communism.

Only when AI moves in the same direction do some people rush to attach that label.

The reason is not that the route is wrong.

The reason is that if the route works, a set of trillion-dollar companies may become far less necessary.

The Fire Is Already Lit

There were other events inside Silicon Valley that did not fit neatly into the main narrative.

xAI’s Grok, Musk’s own model, quietly open-sourced parts of Grok 3 during this period. Musk said “fraction of the cost” with his mouth. With his hands, he began learning China’s open-source playbook.

That is not a coincidence.

He is choosing between the same two routes.

Google’s Gemini 3 was unusually quiet during those 33 days. Not because nothing was happening, but because Google’s internal debate over open versus closed AI has not stopped since Gemma began in 2023.

Google was one of the first giants to move from “we only do closed” to “closed frontier plus open edge.”

It knows search will be reshaped by AI. Better to build its own open edge than let someone else’s open ecosystem eat the search infrastructure.

Anthropic, meanwhile, is the most stubbornly closed of the frontier labs. That may explain why the past six months have brought so many stories around delayed model releases, alignment researcher departures, and public safety reports about abnormal model behavior.

Its route is under pressure from both technical reality and business reality.

And OpenAI?

OpenAI is the most split of all.

Its name contains “Open.” Its founding mission was to make AGI benefit all of humanity. Its business model is now fully closed.

Sam Altman keeps saying OpenAI will eventually open-source something meaningful. But no true open frontier model has arrived.

Meanwhile, it just hired Dean Ball, a man publicly arguing that open weights are “AI communism.”

The distance between the word “Open” and the current route has rarely been clearer.

There is an old Chinese phrase: a single spark can start a prairie fire.

Kimi K3 is one spark.

GLM 5.2, Qwen 3.8, and DeepSeek V4 are three more.

Xi Jinping’s “symphony of global cooperation” at WAIC and Tang Jie’s “the height belongs to humanity, the road belongs to everyone” are the ideas behind those sparks.

Ball’s “AI communism” post was not analysis.

It was an alarm from someone who felt the heat under his feet.

And Sacks, Chamath, Nathan Lambert, the pro-open-source faction inside Meta, European countries like Switzerland building AI as infrastructure, and engineers across Southeast Asia, Latin America, and Africa downloading Qwen and GLM for local applications are not merely responding to China.

They are responding to an old civilizational question that runs from Illich to GNU to Wikipedia to Bitcoin to Linux:

Who should technology belong to?

Those 33 days did not answer the question.

Ball’s phrase made it impossible to avoid.

My only confident judgment is this:

When a company needs FUD, the word “communism,” and dystopian imagination to fight a rival, it is probably not on the winning side.

Ballmer called Linux a cancer.

IBM used FUD against UNIX.

OSI was once more complete, more secure, and more responsible than TCP/IP.

Today, Linux runs on almost every server. TCP/IP is the base layer of the internet. Wikipedia is the largest knowledge base humanity has ever built.

In 2036, when we look back on these 33 days, we may find that Ball’s phrase was free advertising for a new era.

And Tang Jie’s line, “The height we reach belongs to all humanity; the road we build belongs to everyone,” may be remembered as one of its earliest declarations.

Another old Chinese line says: trust the people.

In AI, “the people” means every person who can download weights, read code, and run an open model on their own machine.

I trust them more than I trust the boards of a few labs.

The fire is already lit.

Open-source models are no longer isolated victories.

They are the beginning of a public road being lit together.

Sources Mentioned

• Dean Ball’s six-part thread and background reconstruction: thekb.eu

• Tang Jie’s internal Zhipu letter: Sina Finance

• Xi Jinping’s WAIC keynote: China.org.cn

• Nathan Lambert on China’s open-source AI trajectory: Interconnects

• David Sacks vs. Dean Ball: Andrew Curran’s X record

• Chamath on the philosophical split in AI: Chamath’s X post

• TCP/IP vs OSI history: IEEE Spectrum

• The history of FUD against open source: CNET, 2002

• Ivan Illich, Tools for Conviviality: Cornell archive