• Reason to collaborate: to be the master of your own fate (English Dubbed) — targeting senior officials

• Reason to collaborate: to create a better future (English Dubbed) — targeting junior officials

What can I say? Those ads are pretty cringe. Besides the content, it just doesn’t feel very professional to be recruiting international spies on a social media platform. There’s clearly a mismatch of the “brand identity” and the “marketing channel,” but I guess anything goes these days if the White House has a Lo-Fi MAGA playlist.

Anyway, I digress. Although the CIA stunt hasn’t gone viral on the Chinese Internet, commentary has started to surface. While most are ridiculing the awkward accent in the ads, Hu Xijin, the former editor-in-chief of the Global Times—a media outlet affiliated with CCP —called the campaign “an action that is almost openly hostile to China.” Hu is more of a clown on the Chinese Internet at this point, but he did make another interesting point — instead of looking at these ads as a sincere olive branch, they are more likely designed to sow distrust within Party ranks by stoking fears of internal espionage.

In other words, this is more of a psyops than a recruiting attempt.

While this is laughably unlikely, it sent me down a rabbit hole into psyops history. Psyops has been around forever, probably since when we first learnt to kill one another — there’s a legend that Alexander the Great commissioned giant armor built for eight-foot-tall “soldiers,” and left them behind to terrify his enemies into thinking that his army had giants.

While psyops can take many forms — from fake giant armors to cringeworthy ads — they all try to tap into basic human instincts like fear, self-preservation, love, and even jealousy. Peering back in psyops history, the constant theme has been scaling the reach of influence on the back of new technologies. As global tensions mount again, where does the battle for hearts and minds go?

Scaling psyops over the air

Scattering fake made-to-order giant armors is hardly scalable for the modern warfare. As time goes on, each leap in technology expanded the reach, impact and creativity of psyops campaigns. If we take a look at the early history of scaling psyops, for a very long time it had been a battle over the airspace.

Over the air 1: Hot Air Balloon

One of the earliest examples of using balloons to drop propaganda leaflets was in 1854, when the French tried to persuade Russian soldiers to stay out of the Crimean War. Later, during the 1870 siege of Paris by the Prussians, the French launched 66 unguided balloons over enemy lines carrying 10,000 leaflets, urging German troops to recognize that the French were no longer ruled by emperors and suggested that the needless bloodshed could end if the German simply joined the French.

Over the air 2: Airplane

As aviation technology advanced, more efficient and guided psyops campaigns became possible. The invention of airplanes in the early 20th century turned leaflet drops from a weather-dependent gamble into a somewhat precise tactic. During World Wars I and II, planes became the workhorses of aerial propaganda, scattering millions of leaflets behind enemy lines in a matter of hours. Both the Allies and Axis powers used them extensively to demoralize troops, encourage surrender, and provide instructions for safe defection. The British and Americans even developed specialized leaflet bombs, like the Monroe bomb, which could deliver hundreds of millions of leaflets over enemy territory.

Over the air 3: Sound waves

While planes scaled up psyops by a magnitude, they are still limited by borders, weather and enemy activities. Sound waves do not care. By World War II, propaganda over the airwaves had become a cornerstone of psychological warfare, with both overt and clandestine stations broadcasting propaganda and misinformation.

While loudspeakers blasted on at the edge of conflict, clandestine radio stations were able to penetrate enemy line and often masqueraded as legitimate enemy broadcasts. During Vietnam War, in just 1966 alone, the Air Force dropped over one billion leaflets and broadcasted hundreds of hours of propaganda, which led to the defection of more than 15,000 Viet Cong soldiers.

Scaling psyops over Internet

While airborne tactics defined much of the 20th century’s psychological warfare—from floating pamphlets to booming loudspeakers—psyops continued evolving alongside communication technology. As societies shifted from physical to digital infrastructure, so too did the methods of influence. The rise of the internet didn’t just change how people connected; it redefined the terrain on which psyops could operate. No longer constrained by geography or weather, psyops entered a new phase.

Why email isn’t a good psyops platform

In addition to large scale propaganda happening in the airspace, there were also more creative attempts to slip messages behind enemy lines. One of the most well-known psyops campaigns was Operation Cornflake in WWII. On February 5, 1945, operatives dropped eight mailbags in a bombed German mail train — each one filled with fake letters and newspapers. These counterfeit mailbags were carefully designed to look authentic, complete with real German addresses, legitimate business return addresses, and postage stamps that closely resembled official ones—some subtly altered to show Hitler’s face as a skull.

Creative as it was, it’s extremely hard to scale given all the effort going into forging the materials. So when communication shifted from physical mails to emails, it seemed like psyops would follow too. Email is easy to automate, personalize, and mass-distribute—so why didn’t it become a staple of psychological warfare? You could argue that it’s due to the fact that we haven’t had major scale conflicts that would require a digital campaign since the invention of email in 1971, or you could argue that other social media platforms simply have more reach than email.

However, email is fundamentally flawed for psyops delivery. Spam filters and cybersecurity tools make it challenging for these messages to even reach their targets. Unlike a poster on a wall, a radio broadcast, or even a social media post that can be passively consumed and shared, emails require the recipient to actively open and read them. If you’ve ever been on a GTM team, then you know that many people simply don’t open their emails. Plus, most people are naturally skeptical of unsolicited messages in their inbox—especially from unknown sources—which makes it hard to gain trust or spread influence at scale.

TikTok: The perennial new frontline

This brings us to TikTok—arguably the most potent modern vehicle for influence operations. In fact, one of the key motivations behind the proposed TikTok ban is the platform’s potential to conduct large-scale psyops on U.S. users.

TikTok currently boasts ~170 million active users in the U.S., far surpassing X and on par with Instagram. When it first faced a potential ban, TikTok urged users to contact Congress in protest. The move backfired, obviously. What was meant to feel like civic engagement came off as mass agitation. While there are precedents of companies rallying their users in political matters, TikTok’s ties to the Chinese government made this appeal far more troubling.

If the concern ended there, the threat might be more manageable. In a time of crisis, the U.S. could, in theory, block TikTok at the ISP or CSP level. But psyops is not a one-off risk, as there’s growing evidence of TikTok’s systematic influence operations.

A 2023 paper from Rutgers University’s Network Contagion Research Institute, titled A Tik-Tok-ing Timebomb: How TikTok's Global Platform Anomalies Align with the Chinese, examined how TikTok manipulates content visibility. The study compared the distribution of topics on TikTok with similar content on U.S.-controlled platforms like Instagram, and found that critical content about China is heavily suppressed.

Outro

The battle for our hearts and minds continues—but it no longer floats by balloons or blasts through a speaker. Today, it scrolls through our feed. As global conflict increasingly plays out in the digital realm, psyops is shifting from broad propaganda to more subtle, persistent influence. Future operations may not aim to shock or persuade in obvious ways, but to slowly shape perception through everyday content in our little echo chambers. Platforms like TikTok are only the beginning. As AI-generated content, deepfakes, and hyper-personalized influence campaigns become more accessible, the most effective psyops may not feel like psyops at all—it may just feel like the Internet is working as usual.

Should AI companies pay artists for training on their work?

Well, that depends on what you mean by should — legally, ethically, or philosophically? Either way, it’s a question I rarely hear at Bay Area networking events. No wonder, of course, questions like this get easily drowned out in the fireworks of progress.

While there are AI startups like Moonvalley or Beatoven that exclusively train on licensed data, the rest of the technosphere does not share the same view. Both OpenAI and Google argue that it should be free to train on copyrighted data. More specifically, Google argues that the current U.S. law can be interpreted to support its claim. OpenAI, on the other hand, raises national security concerns, claiming that copyright laws could cause the U.S. to fall behind in the AI race — since it believes that Chinese AI companies won’t respect copyrights.

Artists clearly disagree, as hundreds of Hollywood figures like Guillermo del Toro and Paul McCartney signed an open letter in protest of OpenAI and Google’s stance. However, the world had long moved past the niceties of open letters, as there has been a series of litigations between copyright holders and AI companies such as

• New York Times v. OpenAI (text data)

• Getty Images v. Stability AI (image)

• Recording Industry Association of America v. Suno AI (music)

There’s clearly a rift between the tech world and the artists, so where do we go from here?

At the inaugural HumanX conference, I had the opportunity to sit down with two key executives from the music industry: Andrea Czapary Martin, CEO of PRS for Music, and John Mottram, Chief Strategy, Communication & Policy Officer. PRS for Music is one of the largest music societies in the world, responsible for collecting and distributing royalties for over 42 million musical works on behalf of 175,000 members — including artists like Billie Eilish, The Beatles, and Charli XCX.

In this blog post, we’ll explore

1. Perspectives from PRS for Music

2. The current legal landscape

3. What the future could look like, legally and ethically

Perspectives from PRS for Music

PRS for Music is the perfect company to speak on this subject. On one hand, it has represented musicians’ interests for over 100 years; on the other hand, it is one of the most tech-forward organizations in the industry. Throughout my conversation with Martin and Mottram, I learned that they are genuinely open to the adoption of AI in creative processes, and they even hosted a panel about Making Music with AI so their members can better understand the potential of this technology.

To PRS for Music, AI is a wonderful tool that unleashes the creativity of musicians and can be their copilot in producing even better work. This is supported by a recent survey conducted by PRS for Music, which showed that ~60% of musicians are considering using AI in the future, while 30% are using it already.

However, despite PRS for Music’s willingness to collaborate with AI companies, the goodwill has not been reciprocated. While AI music companies like Suno AI have publicly admitted that they are actively training on copyrighted music, PRS for Music has never been approached by any AI companies to seek permission.

PRS for Music believes that AI companies should license music for training, and they are not alone. Alongside other key industry organizations like — Motion Picture Association, Getty Images and the Associated Press — PRS for Music is a part of the Creative Rights in AI Coalition, which believes that

1. AI developers should obtain permission to train on copyrighted data

2. AI developers should disclose the source of their training data

3. A robust content licensing market is essential for the growth of both the creative industry and generative AI

While the coalition is advocating for greater collaboration between AI developers and copyright holders, some parties — as we’ve mentioned earlier — are already taking the fight to court.

The current legal landscape

Fair use or over use?

Most of the ongoing legal disputes hinge on the interpretation of fair use, which I went over in my last blog post about the Ghiblification trend.

In general, four key factors guide the conversation around fair use

1. The Purpose and Character of the Use

2. The Nature of the Copyrighted Work

3. The Amount and Substantiality of the Portion Taken

4. The Effect of the Use Upon the Potential Market

In the case of AI music model training, Suno argued that training is fair use because

• The Purpose and Character of the Use: Suno AI used the musical work to train an AI model as opposed to play it in public, which makes it a transformative use.

• The Amount and Substantiality of the Portion Taken: While Suno trains on an extensive corpus, very little of the original material appears in the output.

In contrast, the plaintiff — Recording Industry Association of America (RIAA) — argues that training is not fair use due to The Effect of the Use Upon the Potential Market, since a flood of AI generated music could severely impact musicians’ livelihoods.

Some argue that fair use, codified in 1976, is overdue for an update, especially given the rapid advancements of AI. In other words, AI has pushed copyright laws into uncharted territories, and the ambiguity of fair use makes it extremely challenging to reach a clear verdict. So, again

Should AI companies pay artists for training on their work?

As we look ahead for the answer, it might benefit us to look around first. After all, both AI and music are global industries, so what do the laws and regulations look like in other regions?

Look around before looking ahead

Let’s examine two examples at opposite ends of the spectrum — Japan and EU.

In 2018, Japan amended its Copyright Act to include Article 30-4, which created a broad exception for the use of copyrighted works for "information analysis" — including AI model training even for commercial purposes. Note that, Article 30-4 does include nuances that limits applications that generates similar work or “unreasonably prejudices the interest of the copyright holders.”

While Japan is much more lenient on AI model trainers, EU is more on the side of the creators. The EU AI Act allows copyright holders to opt out of AI model training, and the extraterritorial reach clause mandates that the opt-out preference must be honored globally as long as

1. EU data is used

2. The AI model is deployed within EU

While this seems promising in theory, it’s nearly impossible to prove that a specific song is used in training unless AI companies proactively disclose their data sources. In other words, how can we enforce a law without the ability to reliably detect violations? In response, Mottram smiled and said

The system is very obviously broken at the moment, and all governments need to come together to make it work.

Looking ahead

During the interview, Mottram suggested that the current industry consensus is that the license should not be a one-off purchase. Instead, the license should include a recurring component since, once the musical work is ingested, the AI model can generate music and potentially draw on the original work in perpetuity.

GEMA has recently proposed a licensing model that includes an upfront and recurring fee. However, the recurring component related to subsequent use remains vague, as it is challenging to detect whose work has been used in the AI generation process. PRS for Music did not comment on the GEMA model, but emphasized that AI could be a powerful tool for musicians, and they are keen to work out the difference to ensure the mutual prosperity of AI developers and musicians.

Another proposal advocates for statutory remuneration rights, which allows AI companies to pay a predefined fee to right holders without prior authorization. However, the model lacks a recurring component and it would still require negotiation, as the predefined fee would likely vary from artist to artist.

One model that I’ve been mulling over is inspired by the AI music startup Lemonaide, which offers a library of music models trained by popular producers like KXVI and Mantra. This kind of artist model makes it much easier to track and attribute license, and it even opens the door to MoA (mixture-of-artist) models that could spark new musical styles. In addition, it allows musicians to properly credit the source of their inspiration. In a previous blog post about the AI x music stack, I wrote about my experiments with Suno —

But maybe that 15-second spark came from a real musician somewhere — someone I might have heard in a random coffee shop, a jazz bar, an elevator or just on YouTube, yet there was no way for me to trace back and credit that person.

Outro

We looked around before we looked ahead, but it’s also important to look within.

Law is the lowest bar of our moral standard — just because something’s legal (for now) does not mean it’s right. I do not believe that massive AI training on artists’ work without permission or compensation is right — most artists already make very little, and it is unfair that AI companies are raking in millions at the same time, which would not be possible without the artists that poured hearts and souls into their work.

This is likely a symptom of today’s AI learning paradigm, where massive consumption of data is needed. Yet we all know that you don’t need to hear every song in existence to make music that resonate with the world.

Perhaps a better way forward is by building an AI that actually understands music theory, instead of chewing everything up and regurgitate inspiration as a service.

Progress is good, messy, and can often come at the cost of somebody else. However, as a modern society, we should do our best to address the interest of the affected party. I continue to believe in the potential of AI in creative processes, and I look forward to many more uncomfortable discussions ahead.

Vote Left Nothing Left

Trump is Right

Turn Right Everything Right

Next to the Duilian, dressed in red, is our protagonist. Her name is Nanyang Li and she holds many identities — a U.S. citizen, a magnet engineer, a Hoover scholar, a mother, a wife, an individual of her own will, and as you can see, a firm believer in President Trump — but perhaps her most well-known identity is as the daughter of Mao’s secretary.

Both Li and her father are outspoken critics of CCP, especially on its idolization of powerful figures like Mao. Therefore, I was very surprised to see her unwavering support for Trump, yet another powerful figure. I was intrigued by the juxtaposition and wanted to understand how she thinks.

This post is not about politics, ethics or even truth.

It is about rationality, the rationality of who I consider a critical thinker. In order to first understand how Li’s views were shaped, we must start from the beginning, in her most formative years.

The life of Nanyang Li

Nanyang Li was born in 1950, a time of both upheaval and hope. The Chinese Civil War had just ended, giving birth to the People’s Republic of China. Both of Li’s parents were active CCP members at the time, and they had been a part of the revolution since their teenage years.

At the time of her birth, Li’s father was only 33 years old but had already become the Minister of the Publicity Department in the Hunan Province. His political ascent scaled rapidly, soon he would become the Minister of Water Resources in 1952.

In 1958, Li’s father was hand-picked by Mao to be his personal secretary. At just eight years old, Li couldn’t have imagined how quickly her life would be turned upside down. Just one year later, after speaking out against Mao’s disastrous Great Leap Forward — a policy that led to the death of more than 10 million people — Li’s father would spend the next two decades in exile between solitary confinement and labor camp.

Li’s mother, on the other hand, was devout to CCP and worshipped Mao —so much so that she turned in her brother and her husband, in order to demonstrate her unequivocal allegiance to the Party. Even after divorcing Li’s father, when his name was cleared and returned to power in 1984, Li’s mother still wrote to the Party and forced him into retirement. Li’s mother hated her husband for bringing shame into her family. In a letter she wrote in 1976, she summarized their 20-year marriage with a single phrase:

It was a class struggle.

While Li’s father was in exile, Li’s mother was consumed by anger and routinely took it out on Li. During the Cultural Revolution, Li was bullied in school due to her father’s “crime.” When Li’s mother found out, she sneered and said

Looking back, Li said it felt like sinking into a vast ice crevice. From that point on, her heart closed itself off to her mother, forever. Throughout her childhood, love was conditional. Affection? As foreign as freedom.

Growing up in this kind of family, the sentence Li feared the most was when people said “you’re starting to look more and more like your father.” Li knew how much her mother despised the comparison. For Li’s mother, what had started as enthusiasm to serve the revolution had escalated into fervent loyalty to Mao. Over the years, it had putrefied into bitterness, then hatred, a hatred so large that it overflowed onto her children. In order to shield herself from her mother’s resentment, Li eventually wrote a letter to formally denounce her father in 1967. It read

A turning point for Li was the 1976 Tiananmen Incident. Appalled by the violent measures taken against civilians, Li’s belief system was shaken. For someone raised in a household of unwavering loyalty to the Party, the incident cracked the foundation of everything she had been taught to believe.

She then spent two years searching for her father, and she finally learnt the true story about him in 1978. Through his firsthand accounts, Li came to understand the depth of betrayal he had endured, and she also learned things about Mao that the newspapers had never dared to print. Over the next decade, she watched her country continue to wrestle with its own identity — hope and fear, reform and regression. Eventually, Li would leave China for the U.S. in 1990 after witnessing events that further disillusioned her.

Li & Trump

Decades after Li first set foot on the U.S. soil, it is now 2016 and the presidential election was captivating everyone’s attention. When Trump first came onto the political scene, Li was suspicious of Trump’s qualification because he was more of a businessman or a media personality than a politician. However, Li was gradually won over by Trump, and she was not shy about showing her firm support on Twitter, posting things like

Given that Li had experienced systematic indoctrination before finally realizing the harm of worshipping powerful figures, her passionate support for Trump must have been a deeply thoughtful decision.

After going through her online footprint, one plausible theory presents itself

Li believes that Democrats is becoming more and more like the CCP that she so vehemently despises, and Trump is the only antidote.

On 10/21/2023, she posted

The Democrat has become the CCP, and the GOP is as muddy as Kuomintang. Only Trump and MAGA grass root haven’t given up.

Li drives even more parallels between the CCP history and America’s current political climate. For instance, she commented that

• The BLM movement is the same as Mao’s strategy to divide and control, turning people against each other in order to solidify control over the masses.

• The Green New Deal is the same as planned economy because the government is in charge of it.

• Obama’s concern for a divided media makes him the same as the CCP. In an article, Li said that “Obama wishes he would be respected as Mao Zedong whom he is so fond of. He hopes the Democrat holds the power as the CCP does perpetually.”

• The treatment towards Trump and his supporters is deeply unjust. For instance, She was sympathetic to the Jan 6th rioters, believing that these rioters were suffering the same injustice that her father went through for acting out against the state. She elaborates on her points in the following video.

These claims are clearly a far stretch at best. For instance, as a quick rebuttal on her comment about the Green New Deal — if any government-led program were to be construed as planned economy, then effectively every country in the world would be on the brink of communism.

One cannot blame Li for her biases. After all, after decades of trauma she went through, even the faintest echo of communism can ring deafeningly loud.

After Trump’s first term, Li was even more fond of Trump. She mentioned that she was impressed by how Trump delivered on 80% of his promise, and she believed that the remaining 20% was botched due to bureaucracy and political infighting. Despite multiple sources (BBC, Poytner Institute) that disputed the validity of those stats, Li still says that she “agrees with every single one of his (Trump) policies.” This includes U.S. foreign policies as well — on the subject of Ukraine War, Li was moved by Trump’s humanitarian effort in terminating the war in order to prevent further casualty. But by that logic — if human life were to trump everything, should China have surrendered to Japan during the invasion?

Many have criticized that Li’s loyalty to Trump is not so different from the blind idolization of Mao. Still, despite repeatedly likening Democrats to the CCP, Li refused to compare Trump with Mao:“it is not fair to compare Trump, a democratically elected leader, with Mao. This analogy is logically incoherent and lacks common sense.”

At this point, it is clear that what’s “logically incoherent and lacks common sense” is Li’s own set of views. However, her firm belief has rendered her oblivious to her own fallacy, and this led to me digging deeper, in search of other rationale.

In a recent interview, I believe I found the answer. When the host asked Li why she supports Trump, she cited anecdotes that assured her of Trump’s upstanding character.

1. In 1986, a farmer committed suicide so his family could benefit from his life insurance. However, the insurance company discovered this fraud and refused payout. Trump found out about the story and paid off the family’s debt.

2. In the winter of 1995, Trump’s limo had a flat tire. After a stranger helped out, Trump asked what that person wanted. The stranger said “just some flowers for my wife.” A few days later, a big flower basket arrived on the man’s doorstep, with a card from Trump saying “We paid off your home mortgage, Marla and Donald.”

Whether these stories are true or not is beside the point. What matters is how they became the anchor of Li’s belief. Li went on saying that she was impressed by how Trump is able to keep all his children close under a big family, despite having gone through three different marriages.

“He must be a great father,” Li said, “My family has such little love, the humanity that Trump showed really moved me.”

Outro

Here we examined a mind — a mind that has been poisoned then detoxified, a mind that is strong and fragile, a mind that knows some truths but blind to others — a mind not so different from ours.

The traumas caused by the CCP and her parents are the two cornerstones of Li’s mind. It doesn’t matter whether Trump was the right choice or not, it doesn’t matter if there are opposing opinions, it doesn’t matter if there’s contradicting evidence, it only matters because it feels right to her.

To Li, Trump is everything that she longer for: a disruptor of a society she believes is slipping into communism, and a loving father that she never had. It is both ironic and saddening to witness her intellectual relapse, but as Maya Angelous wonderfully put

I've learned that people will forget what you said, people will forget what you did, but people will never forget how you made them feel.

Unless you’re living under a rock, you’ve heard of Browser Use. Basically, it’s an open source dev tool that allows AI systems like Manus to surf the web through browser interactions. It has just announced a $17mm seed round, and developers love it as you can tell from the ~54k stars on GitHub. I mean, why shouldn’t they? After all, Browser Use boasts 89.1 on the Webvoyager benchmark, crushing AI powerhouses like Anthropic and OpenAI.

However, quietly emerging from the same YC batch is another contender: General Agency, whose state-of-the-art model currently scores 93 on Webvoyager, yet it hasn’t received the same recognition as Browser Use. How come?! Quickly checking my maths…but 93 is clearly higher than 89.1 right? With those questions, I sat down with Harvey and Mo, the founders of General Agency to learn more about their company.

Before we dive in, to see a sample side by side comparison between General Agency and Browser Use, when they are both prompted with

Find out the price of The Jay, Autograph Collection SF on March 9th, 2025. Go to http://booking.com and click through all the steps to give me the final price after taxes and fees.

General Agency was able to retrieve the most accurate pricing by actually attempting to reserve the room, whereas Browser Use prematurely stopped and only reported the pricing without taxes and fees. When I asked if there is a set of real world tasks that General Agency would beat Browser Use at, Mo said

“Any task requiring piping data through API integrations. One of our customers is using our system to interface with Intercom, pull data from there, browse the internet using that context, and also sending emails. Most importantly, we can do this with basically 99.9%+ reliability.”

Despite achieving the state of art performance on web browsing benchmark, General Agency has a bigger vision. Just as its named suggested, General Agency is not satisfied with just being a dev tool, and web browsing is simply an element of a bigger picture. General Agency wants to be your trusted teammate that helps you with the busywork, and in order to do that, it developed a new architecture called the Neurosymbolic Self-Model.

The compute graph is the orchestration layer, consists of multiple LLMs, that can reason, plan, and prioritize. However, what sets the architecture apart is its use of a procedure graph, which allows the system to continuously learn without constant fine-tuning. While the knowledge graph helps with ingestion and retrieval, the procedural graph captures state-action pairs, effectively creating a “world model” for the agent. This enables the system to remember and adapt to unexpected outcomes—something an LLM alone cannot do—by retaining episodic replay without modifying model weights.

In terms of GTM, General Agency is first setting its eyes on business process outsourcing (BPO), which you can learn more about from a recent interview with Kimberly Tan, Investing Partner at a16z.

In terms of their growth strategy, while the team is keenly aware of the benefits of open source, they’re choosing to block out the noise and overhead that comes with maintaining an open source project.

Overall, I’m excited to see where 2025 will take General Agency and revisit the “rivalry” at the end of the year. Currently, the team is working with design partners in supply chain, finance and consulting firms, so if you are an enterprise looking to automate part of your BPO stack, talk to the team to learn more!

Don’t think I need to spend too many words on what happened, but here’s a short recap for my friends on LinkedIn. Basically, OpenAI updated their image generation feature, and the demo inadvertently started a trend to Ghiblify everything on Twitter.

Initially, it started with relatively cute use case of turning family pictures to Ghibli style, but then people started applying it to memes. It all went to shit starting there, culminating at The White House’s Ghiblification of deporting an alien felon.

Needless to say, lots of people hated it — you can read a more complete recap of the timeline, background and additional commentary on ’s blog post Everything is Ghibli.

Good Graf!

Everything is Ghibli

OpenAI unleashed its native image generation in ChatGPT on Tuesday. By Wednesday morning, every social feed was drowning in Studio Ghibli-style portraits. (Linkedin, check back next week.) What happened—and why—is another signal of where AI, art, and our attention are headed…

Read more

a year ago · 107 likes · 12 comments · Carly Ayres

In this post, I’ll dive deeper on the legal complication with some additional thoughts.

Is this all legal

As the Ghiblification of Twitter continues, this Twitter user claimed that he received a cease and desist from Studio Ghibli. Although later (thanks to community note), we know that it’s only a growth hack stunt, it does beg the question — is this all legal?

It is clear that OpenAI has likely ingested Studio Ghibli’s work during its training process, but is Studio Ghibli getting paid for this? Likely no, as OpenAI has stated in its open letter to the Office of Science and Technology Policy, it believes that the U.S. should allow AI companies to train on copyrighted materials for free and without permission. In addition, this is all in the name of national security because Chinese AI companies will win the AI race, as they do not care about copyright laws.

However, the legal basis for OpenAI’s center claim is a legal doctrine called fair use.

Fair Use

Fair use, first formally codified in the Copyright Act of 1976, is a legal doctrine in the U.S. that allows for the limited use of copyrighted material without obtaining permission from the copyright holder. It is designed to allow the public to use creative works for purposes such as criticism, commentary, news reporting, teaching, scholarship, and research.

There are a couple key factors to determine whether something is fair use or not

1. The Purpose and Character of the Use

2. The Nature of the Copyrighted Work

3. The Amount and Substantiality of the Portion Taken

4. The Effect of the Use Upon the Potential Market

Note that this is not a check list as laws are up to interpretation. A landmark lawsuit around fair use is The Authors Guild v. Google Inc — in short, authors and publishers sued Google for copyright infringement over its digitization of millions of books for online search and snippet viewing. During the litigation, the following key factors were considered

• The Purpose and Character of the Use: Google scanned the books mainly to provide information about the book rather than for online consumption

• The Amount and Substantiality of the Portion Taken: Only snippets of the books showed up during the search result

• The Effect of the Use Upon the Potential Market: Since only snippets showed up, Google’s action does not significantly impact the sales of the actual books. Moreover, it could contribute to sales as more people discover the books.

Hence, with all these factors considered, the courts ultimately ruled that Google’s action fell under fair use despite its commercial motivation. If we were to apply them to OpenAI

• The Purpose and Character of the Use: OpenAI used Ghibli images in training so people might create images in Ghibli’s style. However, the liability would be on the users — if they Ghiblified their family album, they should be fine as it’s a completely different use case than what Ghibli Studio’s work was intended for. However, if they used OpenAI to produce a Ghiblified anime film, then they are in trouble because it’s the exact same purpose of Ghibli Studio’s work.

• The Amount and Substantiality of the Portion Taken: Not sure, but with Lora you actually don’t need that many images to learn the Ghibli style.

• The Effect of the Use Upon the Potential Market: OpenAI’s usage of Ghibli images do not directly impact Ghibli Studio’s business. The liability again falls on the users. If there’s a new studio that churns out hundreds of Ghibli style film per year, then that would not fly under fair use.

In summary, my take (not a lawyer) is that

OpenAI training on Ghibli images does indeed constitute as fair use, but the users of OpenAI might be subject to copyright infringement depending on the nature of their usage.

Closing Thoughts

I’ve said this many times and I’ll say it again

AI cannot replace taste, it can only amplify taste.

We’ll start seeing magnitudes more of garbage on our timeline, but I’d invite everyone to cut through the noise and see what AI can do when combined with artistry, taste and craft. If you are interested in seeing what’s possible with AI, check out my blog post that showcases the winners of the Austin AI Film Festival.

In the meantime, subscribe and make me feel needed. Stay tuned for my next blog post, in which I’ll share my interview with some music executives and how the music industry is thinking about AI and copyrights.

In loving memory of David Lynch

I first stumbled upon the festival back in December, right around the time I started shaping my thesis on AI x Media Entertainment at Amazon. Fast forward three months — as I dashed into the conference center before Austin turns me into a walking brisket (it was 90°), I still had very little idea what to expect.

Now that the festival has wrapped, I’m happy to say that I had a great time there. Granted, this wasn’t a Vegas-B2B-SaaS-conference-scale event, but at least everybody here actually likes what they are doing.

Yes, that was uncalled for, so before my sass goes out of control, let me start by sharing the three main takeaways of the event.

• People are getting tired of just using AI for overdone themes/aesthetics like Studio Ghibli, cyberpunk, or zombies. Who’s the next Tarantino or PTA of AI? Who will push the boundaries and use AI to create something so refreshing that it’s impossible to deny its creative virtue?

• Again, you can never replace taste or vision, AI is only going to amplify it. If you want to use AI to churn out brain rot, nobody’s stopping you (but shame on you). Use it for something good, to tell a story that you care about, to start a revolution from your bed.

• My field survey shows that people’s favorite AI video models are Luma and Kling and the one capability people desperately want is better prompt adherence on movement and camera control.

One more thing before we dim the light — overall, the festival has received 2000+ films from 25 countries across the world, and only 5 films were selected as the winner. A huge round of applause for the winners and to everybody that participated.

Obviously, all of them have many flaws, but they also offer a glimpse of what’s possible with AI. With that in mind, I hope you enjoy!

And the Winner is…

Best Experimental Award — Remembering by Diego Maclean

Best Music Video Award — CAOS by Giovanni Abitante

Best Documentary Award — Knights without Armor by AI or Die

Best Narrative Award — Alex by Dale Williams

Best Trailer Award — Moonblood by Rosa Lykiardopoulos

And my own personal honorable mention!

Il Pagliaccio by Marco Cavazzana

Parallax by Ariel Kotzer

This film has outstanding cinematography, just look at those two shots!

Dominant by Nik Kleverov

Newton’s Cradle by Jeff Synthesized

Probably the best sound design in the festival.

The Philadelphia Project by Jesse Torres for its social commentary

You can learn more about the context from this documentary on Channel 5 with Andrew Callaghan.

Fin

That’s a wrap! I hope you enjoyed at least some of the films here, and I hope you were able to look past the flaws and spot something you resonate with.

Lastly, I want to thank the organizer for making this festival possible, the attendees for all the fun conversation and the participants for their courage to put their work out in public.

And, hopefully, I’ll be on the roster next year!

To which I say — first of all, you are five days late to my birthday, but I’ll take that. Second of all, a 32B reasoning model that rivals R1 (a 600B+ model)?!

What kind of sorcery is this?

Well, it’s the open kind.

You get it? Open Sorcery, like Open Source? I’m not proud of this pun so it’s ok if you don’t like it. But as a quick anecdote — the Chinese AI community has long referred to training LLM as 炼丹 (alchemy), so I feel justified calling the open release of LLM as open sorcery.

In this blog post, I aim to demystify another part of the sorcery, namely

How do open source AI make money?

But first, let me briefly go over the benefits of open source AI. So without further ado…

Why open source

What a crazy quarter for open source AI! We started the year with a wake up call from DeepSeek, then we had Alibaba open sourcing the Wan 2.1 video model, and finally, the Alibaba’s QwQ reasoning model release yesterday.

Not only big corporations are pouring resources into open source AI, startups are emerging quickly as well. Recently, I came across a really cool startup called Oumi (Open Universal Machine Intelligence), who just raised $10 mm and is committed in building an unconditionally open AI platform. (Note that open AI is not to be confused with OpenAI.) By unconditionally open, Oumi strives to build a platform that “offers state-of-the-art foundation models with open code, open data, open weights—and open collaboration,” as opposed to just releasing the model weights like Llama.

In preparation for this blog post, I had a chat with Oumi’s CEO about their recent launch, and one thing he said sticked with me

“Open source AI is the future, the same way Linux dominated.”

To folks like Manos, open source AI is their conviction, which explains why Oumi’s founding team left their previous roles building closed source models, like Gemini, for this new mission. However, for other folks who haven’t taken the pill, here’s the tl;dr on the benefits of open source AI

1. Flexibility — finetune on proprietary data, deploy anywhere, avoid vendor lock in.

2. Security — code, behavior and performance are constantly under scrutiny.

3. Innovation — openness encourages innovation and collective advancement.

In addition, open source adoption can also bootstrap go-to-market motions. In a recent podcast, Ron Gabrisko, Chief Revenue Officer of Databricks, fondly recalled that rooms were packed on roadshows because everybody wanted to meet the star creators of Apache Spark. I bet the same would happen to DeepSeek as well.

For a more extensive reading list, I recommend the following

• Open Source AI is the Path Forward by Mark Zuckerberg (CEO @ Meta)

• If AI isn't truly open, it will fail us by Emmanouil (Manos) Koukoumidis (CEO @ Oumi)

• Keep the code behind AI open, say two Entrepreneurs by The Economists, Martin Casado (GP @ a16z) and Ion Stoica (cofounder & chairman @ Databricks)

Now you might say — well, if open source AI is truly as amazing as you say, then why is OpenAI still closed source? (even though the wind is changing?)

What is closed source AI holding onto? Expelliarmus!

Well, it’s pretty simple — they will tell you that closed source AI makes more money is safer to humanity. The argument is that bad actors could wreak havoc with AI and there isn’t a way to turn off the facet if the model is open source. This sounds more like a protectionist propaganda to me, plus, I believe there are way more good actors who will learn how to counter bad actors equipped with AI.

So besides the safety reason, what else? Well, the biggest revenue stream for closed source AI is usage fee. If OpenAI open sourced their model, two things can happen

1. someone will self host instead of paying OpenAI for usage

2. someone will start hosting OpenAI model as a service and erode its marketshare

In other words, you would be leaving money on the table, that is — if your key revenue stream is from hosting model as a service.

This naturally brings us to the next section — how do we make money with open source AI, and are there additional revenue streams we can explore to make the business case more attractive for open source AI?

Making money with open source AI

Making money as a big corporate

Big corporates like Alibaba, Meta, and Tencent have been making a lot of moves in open source AI lately, so there must be some sort of commercial interest behind it. While Alibaba makes money from hosting its models on Alibaba Cloud, it’s not immediately obvious what Meta is after. Meta doesn’t host Llama as a service and inference providers like Together AI don’t even need to pay Meta for commercial licenses because Meta only requires it “if your services exceed 700 million monthly users.”

At first sight, it seems like Meta is doing all this because open sourcing AI encourages community innovation, which will improve Llama’s performance. A superior model will then create a better user experience through product integration, which will eventually trickle down to Meta’s earning because user stickiness is money.

However, this doesn’t feel like the complete picture to me. Luckily, I was able to find an answer in an illuminating blog post titled Laws of Tech: Commoditize Your Complement. In short, the Laws of Tech suggests the following strategy — say you are a quasi-monopoly (like Meta/Instagram) and wants to leverage AI in your product. How do you make sure that you don’t give your AI vendor too much leverage? One obvious choice is to develop your own AI. However, the real genius move is to open source it, as it levels the playing field in the AI vendor space. In other words, by open sourcing Llama, Meta has fostered competition in its upstream supply chain, reducing vendors’ leverage and hence solidifying its own position in the value chain.

Making money as a startup

Figuring out how to self sustain is crucial to open source AI startups, not only from a commercial lens, but also from an independence standpoint. If open source AI companies can’t prosper on their own, then only large corporates like Meta or Alibaba can afford to open source models, which will be inevitably biased by their own priorities.

Let’s survey the top open source AI startups and see how they are making money

• Mistral

  • usage fee from hosted model (1P and 3P like AWS Bedrock)

  • subscription fee for applications like Le Chat Pro

  • professional service like model customization

  • commercial license for premier models like Codestral

• DeepSeek

  • usage fee from hosted model

• Genmo (Open Source Video Model)

  • subscription fee for hosted model access

• OUMI

  • professional service like model customization, training and deployment

• Unsloth

  • paid tier for premium features like memory and speed optimization

Charging for hosted model access isn’t the greatest business strategy because that’s not a model provider’s core competency. There will be other companies, like AWS or Together AI, who have a deeper expertise as an inference provider. Moreover, users might also find those inference providers more flexible because they tend to offer many models at the same.

The better way to make money on open source projects is to build premium offerings on top of the open core, such as

• Service: RedHat makes a huge chunk of its money from its enterprise support

• Premium Feature: MongoDB is built on the MongoDB NoSQL database but offers advanced features like enterprise security and query optimization. It brought in $1.68 billion last year and has a market cap of $21 billion.

• Certification: Linux Professional Institute offers vendor-neutral Linux certification, which in estimate brings in $17mm per year

Similarly, we may have the following revenue streams for open source AI startups

• Service: help with customization, deployment, etc

• Premium Features: finetuning, RAG, agents, memory or performance optimization all benefit from a stronger base model.

• Certification: models still have their own quirks and mastering them is a skill that could be certified

Another interesting revenue stream is licensing open source models for commercial deployments. However, if anybody can just grab your model from Hugging Face, how do you enforce the license terms?

After all, you don’t want people to pirate your models, right?

How to enforce model license

In order to enforce model license, you must first be able to detect who’s using or modifying your model. Without this capability, it really undermines the defensibility of the open source licensing strategy.

After all, how can IP owners possibly enforce copyright if YouTube doesn’t know how to detect infringement?

There are 2 main ways to detect model infringement

1. Output watermark: detect whether someone is using your model as is

2. Model watermark: detect whether someone finetuned/distilled your model

Output Watermark

Google’s SynthID “watermarks and identifies AI-generated content by embedding digital watermarks directly into AI-generated images, audio, text or video.” On a broad stroke, SynthID works in the following ways

• For text outputs, it tweaks the probability of each token it generates without affecting the output quality, creating a statistical fingerprint that can be detected

• For audiovisual outputs, it does clever signal processing tricks that adds in signatures imperceptible to human.

Other ways to water mark text output includes

• Green red watermark: like how SynthID operates but it suffers from the tradeoff between detectability & quality if the probability disturbance is too large

• Gumbel watermark: it does not perturb the output distribution

With output watermark, open source AI companies can routinely scan the Internet for watermarked output and take actions accordingly. However, what happens if the model pirates are not using your original model as is, but a slightly modified version through finetuning or distillation? Model watermark comes to rescue!

Model Watermark

This technique is becoming increasingly relevant as model distillation gets easier. While DeepSeek shocked the world with its $6mm training cost, another team soon distilled the DeepSeek model with less than $50 and got extremely close to DeepSeek R1 in benchmarks like MATH500. While DeepSeek might be fine with this, other AI companies like OpenAI and Anthropic both prohibit others to distill their models.

Keen readers might question why we need model watermark when we already have output watermark. Since distilled model should more or less inherit the behavior of the original model, wouldn’t the output watermark be preserved?

To some degree it’s true, but the output watermark can be easily removed by a counter technique called ensemble distillation, where averaging the outputs of multiple models can significantly reduce or even erase the watermarks, which necessitates model watermarks.

One of the key paper in this field, Protecting Language Generation Models via Invisible Watermark, introduced a model watermark mechanism called GINSEW. The basic idea of GINSEW is similar to output watermark — However, instead of injecting a generic probability perturbation, GINSEW uses a sinusoidal signal perturbation, which is difficult to erase by the averaging process during ensemble distillation.

I reached out to Xuandong Zhao, first author of GINSEW, for additional comment. Zhao mentioned that while the GINSEW results are promising, it has only been tested on smaller models (<1B parameters). As a summary, Zhao said

“Watermarking could help detect fine-tuning, but it remains an open problem that still needs to be solved.

Fine-tuning is relatively easier to detect; for instance, we can introduce backdoors or modify specific parameters in ways that remain stable when fine-tuned, enabling detection.

For distillation, detection is much more challenging. If someone distills just 10 samples, given the billions of parameters in a model, it is nearly impossible to detect. However, some approaches might work if the distilled dataset is large enough, making detection more feasible.”

Conclusion

As you can see, there isn’t a perfect solution yet. If you ask me to rank the open source commercialization strategy for startups, I’d say

licensing to 3P inference provider > service / become 1P inference provider / premium feature > certification

I know that VCs hate to hear the word “service”, but each model has its own quirk and who has a better insight on how to tame of fix those quirks than the team that built this model? This is especially true for image or video models, which is why I think the Moonvalley x Asteria partnership is so interesting.

Normally, premium features would be ranked as a very effective commercialization strategy for open source teams. Imagine if Llama were the only LLM, then building advanced feature like finetuning or RAG on top of Llama makes sense. However, given that there are so many model providers out there, the opportunity to monetize through advanced features is slim as there are plenty horizontal players.

While the commercialization strategy for open source AI evolves, I really do hope we can unite behind the open source mission and stop spending millions of dollars, energy and intellect recreating the same wheel. In the end, the real sorcery of open-source AI isn’t just in the code—it’s in the collective drive to push the boundaries of innovation together.

Somebody please kill the idealism in me, and yea, subscribe to make me feel needed.

Subscribe now

here’s the tale of its SUCKcess.

Back in 2023, the Chairman of Beijing Enlight Media, the company behind Ne Zha 2, said that “My greater hope is to see if it’s possible to drive China's box office into the $1 billion era, and I hope the film that hits the milestone is Ne Zha 2.”

Well, he’s way off — at the time of writing, the film has raked in close to $2 billions, surpassing Inside Out 2 as the highest grossing animated film in history and the 7th highest grossing film worldwide. It currently boasts a 8.2 rating on iMDb, 99% rotten tomatoes, and the stock price of the studio tripled in 7 days.

No doubt, this is a resounding, massive commercial success, but as you can guess from the title, this film sucks. In addition to that, if we zoom out a bit, there’s a much larger dynamic at play — the battle for cultural supremacy. So allow me to first take you to the 10,000 foot view, and then I’ll explain what are the implications of Ne Zha 2 within this broader context.

Hollywood, China, and Cultural Supremacy

Towards the end of Red Carpet: Hollywood, China, and the Global Battle for Cultural Supremacy, Schwartzel wrote

We are about to watch the final act in a classic tale. China started this century as the savvy ingénue, eager to learn. Countless Hollywood movies have taught us what happens next.

The ingénue grasps for the leading role.

Since early 2000s, in search for more growth, Hollywood has increasingly turned towards China for its massive population and insatiable need for entertainment. As a result, we saw movies like Transformers: the Age of Extinction that scored a franchise low in the U.S. but an all time high in China or Warcraft, which made $47 millions in the U.S. but $221 millions in China.

However, what stood out the most was the Kung Fu Panda series. It was an instant hit in China upon its release in 2008, but more importantly it was a wake up call to the Chinese filmmaking industry. The then President of the China National Peking Opera Company questioned

The film (Kung Fu Panda)’s protagonist is China’s national treasure and all the elements are Chinese. Why didn’t we make such a film?

The box office is not the only motivating factor here, as cinema is also an incredibly effective channel to broadcast a country’s soft power. For decades, the Chinese audience was enamored with the world portrayed by Hollywood films, and China has long wanted to do the same with its own treasure trove of stories.

Hence commenced the final act that Schwartzel predicted — the rise of Chinese films. The act was first kickstarted by a series of films that highlighted Chinese heroism in action/military/adventure films like

In the meantime, another movement started highlighting the traditional Chinese culture through adaptation of classic literature like

Fast track to today, the final act has hit new heights with the release of Ne Zha 2. As an avid fan of Chinese cinema, there would’ve nothing I’d like to see more than a Chinese film win big. However, despite Ne Zha 2’s earth-shattering success, despite the countless media coverage that brands this success as a win for China, I simply cannot be happy for it. In fact, I’m worried that this is setting the wrong kind of example, where the box office becomes the only North Star.

Since 2013, the Chinese government has published various memos on this topic, and the slogan is “让中国电影走出去”, meaning “let Chinese film walk onto the world stage.” Ne Zha 2 is not the answer despite having the illusion of being one. In order to really propel Chinese films onto the world stage, two criteria must be met

1. It needs to be an outstanding film

2. It needs to get at least 60% of its ticket sales from foreign market

Keen readers might have already noticed that the two timeline graphics in this section revealed something very interesting —

While Chinese films have scored major box office wins, the vast majority of the revenue came from the local China audience alone. 96% is the lowest number, that’s crazy localized!

Similarly, for Ne Zha 2, the ~$2 billions box office was hit with little help from the rest of the world, with only $30 millions from foreign markets. In other words, 98.4% of the ticket sales came from China alone, and this is exactly why Ne Zha 2 is not the right answer because a great film is a pan-human experience that could not possibly just be a localized hype.

While the hard numbers of foreign market share is hard to argue with, many will disagree with me calling Ne Zha 2 a SUCKcess, so allow me to elaborate.

Why it sucks

I’m not going to talk about super contextual elements because these are hard to explain in a blog, instead I will focus on aesthetics, which is the visual foundation of a movie. Like it or not, if the aesthetic sucks, you can’t possibly have a good movie.

The aesthetics of this film is as if Marvel and TikTok conceived a baby together.

This is not particularly surprising as ByteDance is indeed Ne Zha 2’s strategic content partner. It’s pretty much 2 and half hours of sensory overload on top of a predictable plot and cringey humors.

As a reminder, Ne Zha 2’s plot loosely follows Investiture of the Gods, a classic Chinese mythology novel. I’m not saying that humor is not allowed in this kind of work, but bad humor is unacceptable. For instance, at one point in the film, Ne Zha is sent to arrest a bunch of goblin groundhogs, and for whatever reason the director thinks a cross over with the popular Internet meme of screaming groundhog would be a good idea. Come on! Even this meme itself has ceased to be funny! Not to mention that given the deep mythology origin of Ne Zha, this would be the equivalent of having Poseidon fight SpongeBob in Nolan’s upcoming Odyssey. Absurdism can work when you lean into it, but not when the rest of the film is severed from that style.

In terms of aesthetics, you might think that it looks fine initially. However, if we compare Ne Zha 2 with NeZha Conquers the Dragon King (1979), the regression of artistic quality is obvious. I’m just gonna leave this here, in comparison Ne Zha 2 doesn’t look that much better than a League of Legend trailer.

Here are a couple more shots from the 1979 version just so you get the idea.

Not only the aesthetics is atrociously cheap, some of them look…well, let’s just say

extremely unoriginal

Sidebar — a funny conspiracy theory

How else might the battle for cultural supremacy manifest? Perhaps by throwing some shade! I saw a couple articles floating around the Chinese Internet about the symbolisms in Ne Zha 2. The conspiracy theory postulates that

• Ne Zha, the protagonist = China,

• The Immortals, the villain = the United States.

To back it up, the engraving on the jade badge that signifies the immortal status looks extremely similar to the Great Seal, so it represents the green card/permanent residency.

The poison that the Immortals tried to feed to Ne Zha represents the COVID virus that some people believe was incubated at U.S. labs in order to sabotage China.

The cauldron that the Immortals used for alchemy represents the Federal Reserve because of the dollar sign on on the cauldron.

The Final Final Act

Going back to the main course of discussion on why Ne Zha 2 is simply not it

1. This film sucks

2. This film has negligible international traction.

Where do we go from here?

If I have any say, it’s time that China learns from its earlier films such as Crouching Tiger, Hidden Dragon that won the hearts, mind and box office of the world. Ironically, Crouching Tiger, Hidden Dragon has not only held the record for the highest grossing Chinese films in the U.S. for 20+ years, but also only $1 million out of the $213 millions world wide ticket sales came from China, which is the exact opposite of what’s happening now.

So perhaps the better example would be Kung Fu Panda 2, which grossed $665 millions world wide, with $93 millions coming from China.

Now how does China get there?

The same way how Kung Fu Panda was made, with care and freedom.

These two things do not magically grow out of nowhere over night. If bad films are rewarded with recognition and success, then commercial filmmakers will be guided with that perverse incentive. Freedom takes even longer to cultivate, and probably the hardest challenge for Chinese films given the increasing tightening of lips in China. Till then, it will take a couple more SUCKcess for China to pop its own bubble, and hopefully the ensuing vacuum will recalibrate the Chinese film industry and finally “让中国电影走出去.”

Subscribe now

Lastly, please subscribe to make me feel needed, thanks.

This is a bit of an unusual post, I originally wrote this in response to ’s policy competition.

While I was sadly not one of the winners, this was a fun thought experiment.

Lastly, a thank you to the team at the Federation of American Scientist for helping me polish this memo!

Summary

Chips are power, and, for much of the past century, semiconductor technology and production has been largely dominated by the United States and its allies. However, in recent years, China has been rapidly developing its semiconductor industry, thereby increasing its leverage in the global market. Since October of 2022, U.S. policies sought to curb China's advancement in the semiconductor industry, but most of them targeted advanced node production, defined by the CHIPS and Science Act as chips smaller than 28 nm. While recent export controls and other measures have likely slowed China's progress in developing cutting-edge technologies, they have not curbed China's growth in the mature-node sector of chips larger than 28nm.

By the end of 2023, China accounted for approximately 31% of global mature node production, a significant increase from 17% in 2015. Furthermore, China is expanding its manufacturing capacity, with 44 operational fabs and 22 additional ones under construction. In comparison, the U.S. has 24 fabs and 6 under construction, with key allies like South Korea, Japan and Taiwan trailing behind.

Policymakers should recognize that mature nodes are of strategic value as well, as they can be found in both civilian and military hardware such as vehicles, electronics, factories, medical devices. The inability to produce enough mature nodes to meet domestic demand would be devastating, as the 2020-2023 chip shortage previewed.

There are various theories of harm in circulation regarding the motive behind China’s increasing dominance in mature node production. A smart U.S. strategy should soberly assess the potential risks and hedge against potentially dangerous future scenarios.In response, it will require a two pronged approach – a foreign policy that fosters new partnership, specifically India, and domestic policies around stricter export control and defense procurement.

Challenge and Opportunity

There are four primary theories of harm to explain why China’s growing market share in mature node semiconductors could potentially challenge US and global interests:

1. Dumping

2. On-shoring

3. Buying leverage with market share

4. Cybersecurity warfare

Dumping

The first theory posits that China has been increasing its mature node production through substantial state subsidies to dump overcapacity into the international market, thereby gaining an unfair economic advantage. This strategy is often used to destabilize competing markets and industries by flooding them with cheap products, undermining local production and leading to potential market failures.

The overcapacity in China's semiconductor production is likely due to mistakes in demand forecasting following the COVID-19 supply chain disruptions rather than a deliberate economic strategy. According to the report "Legacy Chip Overcapacity in China: Myth and Reality" by the Center for Strategic and International Studies (CSIS), it is challenging for foundries to swiftly adjust production lines in response to demand fluctuations, particularly for mature-node productions operating on very thin margins. As a consequence, foundries and customers prefer long-term contracts to secure a specific type of semiconductor, particularly in industries with long product life cycles and stringent quality requirements such as medical devices and automotive applications. Li Ke, vice president of China Center for Information Industry Development’s consulting arm, pointed out that in the past few decades the average company's chip inventory was only 7 days, but now companies generally stock months worth of inventory. Moreover, fabs take 3 - 5 years to build, so the delayed overstocking effect after the initial supply shortage could be misinterpreted as intentional overcapacity.

Moreover, since the first quarter of 2024, the utilization rates of most Chinese fabs have normalized. SMIC's utilization rate increased to 80.8%, HuaHong to 91.7%, and Nexchip has been operating at 110% since March 2024. Inventory levels in major Chinese semiconductor companies have significantly declined since the third quarter of 2023, indicating that the perceived overcapacity is being effectively managed.. According to industry reports, the median inventory of 33 representative semiconductor companies listed on the Chinese A-share stock exchange fell from a peak of 9.5 months in the first quarter of 2023 to 6.7 months in the first quarter of 2024.

In other words, it is unlikely that China is intentionally dumping chips as its utilization rate and inventory level are both returning to industry average.

On-shoring

The second theory suggests that China is increasing its mature node production to onshore more elements of the semiconductor lifecycle, from design to full-scale production.

This strategy is clearly articulated in the “Made in China 2025” initiative, which aims to reduce China's reliance on imported chips and on shore production. Data from the Guidelines to Promote National Integrated Circuit Industry Development also indicated the same strategy by setting an aggressive target of 70% localization rate by 2025. From 2018 to 2022, China's chip self-sufficiency rate increased from 5% to 17%, with an estimated 25% self-sufficiency in 2023.

According to the International Semiconductor Industry Association (SEMI), Chinese chip manufacturers' capacity is expected to expand by 12% this year. If the expanding capacity is mostly serving foreign demands, then the on-shoring claim would be weak. However, SMIC has shifted from 60% foreign customer production five years ago to nearly 80% domestic production. Similarly, 80% of Huahong's production is for domestic customers, which indicates a strategic pivot towards meeting internal demand. In other words, China is expanding capacity and redirecting existing capacity in order to serve its domestic demand.

Between 2021 and 2023, China's import volume of semiconductors decreased from 635.5 billion units to 479.6 billion units, while the export volume decreased from 310.7 billion units to 267.8 billion units. This data indicates a strategic emphasis on on-shoring rather than dumping excess production.

Buying Leverage with Market Share

The third theory posits that China is increasing its mature node production to gain market share and subsequently use this dominance as geopolitical leverage. Dependence on Chinese mature node exports could create opportunities for China to coerce technology transfers, retaliate against sanctions, or influence political matters.

China has demonstrated its willingness to use market share as a foreign policy tool through its export control policies on critical raw materials. For instance, in July 2023, China imposed export controls on gallium and germanium, essential for various high-tech applications. In October 2023 it imposed temporary controls on graphite items, and most recently in December 2024 they announced further critical mineral export controls. These actions highlight China's strategic use of its market dominance to exert pressure on other nations.

While this theory of harm may not be a standalone threat, it could be a side effect of on-shoring. China's decreasing dependence on the global semiconductor industry increases its leverage in international negotiations, potentially destabilizing existing power dynamics.

Cybersecurity Warfare

The final theory involves cybersecurity warfare, which is a threat that cannot be ignored despite the lack of public evidence. There have been allegations about Chinese hackers interfering with the semiconductor supply chain, yet there is little concrete proof. For instance, a Bloomberg report highlighted a security vulnerability due to an unidentified chip on motherboards that originated from China, though the company disputed the report's accuracy. The lack of evidence could stem from the fact that the hardware community is naturally reluctant to discuss these vulnerabilities in public, since many of these backdoors could be unpatchable as they are on a hardware level. However, the ramification of hardware vulnerabilities could be dire and deserve consideration.

State capacity

It is important to realize that the theories of harm only speculate on the intention of the Chinese government. The actual impact depends on its ability to execute. Therefore, to properly right size the challenges one must consider how effective the Chinese government is in directing the semiconductor industry.

One of the most touted examples of the Chinese government’s determination in semiconductors is the launch of the China Integrated Circuit Industry Investment Fund, with almost $100 billion raised.

However, despite substantial investments in the semiconductor industry, China's capital allocation has been marred by inefficiencies and corruption. Reports indicate a disproportionate focus on short-term profiteering in secondary markets rather than nurturing long-term industry research and development.

According to 21st Century Business Herald, the first phase of the fund has entered its investment return stage, resulting in major share reductions for many A-share companies, including Sanan Optoelectronics (down 70% since 2021), Goke Microelectronics (down 80% since 2021), China Resources Microelectronics (down 60% since 2021), etc. This kind of short term profiting is highly contradictory to the claim that the Chinese government is throwing money at the problem and heavily subsidizing the industry without a concern for quick returns, it would appear that the China Chip Fund has been a state-sanctioned pump and dump scheme.

There’s also corruption, as high-level executives of the China Chip Fund, Wenwu Ding and Jun Lu are under investigation of corruption charges. Although there’s only circumstantial evidence, these executives likely were trying to profit from the IPO activities from its portfolio. For example, UNISOC planned for an IPO back in 2019 until it hit a snag when its major shareholder went bankrupt. Since then, the Chip Fund has been aggressively investing into UNISOC, owning 17.7% of UNISOC’s share from the Fund I and II activities. Following the announcement of the third fund, UNISOC immediately received a $4 billion investment, the only significant investment of that quarter. This move would appear to be a last-ditch effort to prepare the company for IPO.

Plan of Action

While there is no substantial backing to the dumping claim, China’s on-shoring strategy and the accompanying leverage it creates for China on the international stage warrant policy responses that hedge against these challenges. In particular, should China continue to decrease its manufacturing capacity available for foreign chip manufacturing demand, new trade partnerships must be forged. In addition, the cybersecurity threat, although currently theoretical, also requires proactive measures.

In summary, the U.S. should

1. Deepen partnership with India by continuing the CHIPS act in support of existing commercial partnerships as well as workforce education.

2. Tighten EDA export control either through total ban or mandating EDA cloud migration, as well as perform routine auditing of corporate structure.

3. Expand defense procurement vendors by making it more attractive to conduct business with the government through demand aggregation with defense startups, as well as continuing the conventional procurement restrictions such as ICT controls

Deepen Partnership with India

In the event of increased Chinese aggression or overcapacity dumping, it is essential to have a reliable partner. India, with its vast population and industrial capacity, could be a decisive factor. The U.S. should foster stronger partnerships with India rather than solely relying on existing allies or on-shoring production.

Onshoring semiconductor production to the United States presents certain challenges, including community opposition, stringent regulatory environments, and cultural clashes. Despite the recent good news of TSMC Arizona’s superior yield rate, the conflicts among workforce will continue to be a destabilizing factor that jeopardize the smooth operation and growth of the on-shoring initiative, as exemplified by the recent class action lawsuit filed by current and former employees of TSMC.

In addition, deepening the relationship with India also helps smooth over the collateral damage caused by the exports control policy. For instance, under U.S. guidance ASML has instructed American employees to cease collaboration with Chinese customers in response to U.S. export restrictions. However, Dutch and Japanese engineers continue to service many machines in China, much to the frustration of their American counterparts. ASML reports that China accounts for 49% of its Q1 total sales and 20% of its backlog orders, underscoring the complexity of decoupling from the Chinese market. In other words, policy actions such as export control erodes the interest and goodwill of U.S. allies. Therefore, rather than jeopardizing the U.S. 's relationship with existing alliances, it is prudent to expand and nurture new partnerships.

U.S. companies are generally reluctant to invest heavily in mature nodes, given that advanced nodes such as 2nm and 3nm processes can command prices ranging from $20,000 to $50,000 per wafer, compared to just a few hundred dollars per wafer for legacy chip nodes. Collaborating with countries like India allows the U.S. to concentrate on developing cutting-edge semiconductor technologies while leveraging India's growing industrial capabilities to meet global demand for legacy chips.

Current U.S.-India partnerships include

1. Synopsys' MOU with Tata

2. Micron Tech's $825 million plan for a chip packaging and testing facility

3. Lam Research's plan to train 60,000 Indian engineers

4. Applied Materials' $400 million investment in semiconductor manufacturing equipment 5. AMD’s plan to build its largest design center in India.

Beyond these ongoing partnerships, the $200 million CHIPS for America Workforce Education fund could create joint curriculum development, student exchanges, and workforce development initiatives, including skilled trades essential for building and operating semiconductor fabs.

At the time of writing (June 2024), this proposal initially recommended that the U.S. Chips International Technology Security and Innovation Fund could direct more investment towards India to build semiconductor capacity and workforce, which has now been officially announced. The announcement is a vote of confidence for the idea of deeper partnership with India, and this direction should be carried through by the next administration.

Learning from the previous chip war with Japan, preferential treatment to allies like India, such as beneficial tax treatment for future chip exports, could be highly effective. This approach not only strengthens bilateral relations but also builds a robust semiconductor manufacturing ecosystem in India, reducing global dependence on China.

Export Control

Strengthening export controls, particularly around EDA tools where the U.S. controls 90% of the global market, is crucial. The existing EDA export control policy has several loopholes. Companies like Synopsys, Cadence, and Siemens have reported ways to tweak their EDA software to comply with restrictions and continue selling to Chinese engineers. For instance, Cadence is selling an EDA version that does not contain the GAAFET feature that’s required for cutting edge chips, which means Chinese companies still have access to EDA tools for the mature node designs.

To address these issues, the U.S. government can start by patching the loophole by banning all EDA software export to China and moving EDA software to cloud service providers based in the U.S. to facilitate more effective enforcement and revocation of access. In addition, the Bureau of Industry and Security (BIS) can mandate that the licenses for EDA export are only granted if the software is provisioned through cloud services.

Beyond export control policy, routine auditing of joint ventures for any ties to Chinese companies is recommended as well since given precedents of Chinese companies trying to circumvent the license control by partnering up with a non-listed entity.

Defense Procurement

According to the "Assessment of the Status of the Microelectronics Industrial Base in the United States," the U.S. lacks visibility into China-based dependencies in the defense industrial base. Addressing this requires conventional ICTS controls, Treasury SDN sanctions, and procurement restrictions for organizations critical to national security through initiatives like the Trusted Foundry Program.

Additionally, increasing transparency and traceability in the defense supply chain is essential. Implementing stringent requirements for suppliers to disclose their sources of semiconductors and other critical components can help identify and mitigate dependencies on Chinese products. Investing in domestic semiconductor manufacturing capabilities for defense applications can further reduce reliance on foreign suppliers. For instance, the most trusted foundry in the Trusted Foundry Program is GlobalFoundry, a UAE corporation. While the UAE has been historically an ally for the U.S., it has also become close to China in terms of trading activities, with a 40% rise in trade over the last decade. This tension creates a vulnerability in the Trusted Foundry Program as the U.S. is heavily indexed on a company located in a region heavily courted by China.

The difficulty to expand the Trusted Foundry Program network is partially caused by the lack of commercial justification for companies to go through the exercise. Commercial entities like Apple order far more chips than the Department of Defense (DoD), so it is often unattractive for foundries to jump through the hoops for government orders. One way to incentivize more participation in the program is increasing the volume of the government order.Given the recent rise of defense tech startups like Anduril, it is possible to group orders together since those chips will all be going to defense products. DoD’s Office of Strategic Capital should spearhead this initiative by surveying opportunities for demand aggregation with the top defense startup executives.

Conclusion

The rise of China's semiconductor industry presents multifaceted challenges that require a strategic and realistic response. Addressing these challenges involves strengthening export controls, enhancing defense procurement practices, fostering international partnerships, particularly with India, and increasing investment in R&D and workforce development. By implementing these policies, the United States can mitigate the risks associated with China's increasing dominance in the semiconductor industry while ensuring a stable and secure global supply chain.

Furthermore, proactive measures to address cybersecurity threats and promote transparency in the defense supply chain are essential to safeguard national security interests. Through collaborative efforts and strategic investments, the U.S. can maintain its technological leadership and build a resilient semiconductor ecosystem that supports economic growth and innovation.

Subscribe now

The most powerful open source node-based application for creating images, videos, and audio with GenAI.

Fun fact: the name comfy came about because its creator was generating AI art with Stable Diffusion and people said his images had a comfy feeling. I don’t know exactly what that means but this comes to mind.

Still don’t know what it does? Fine, let’s see if this helps — this is what ComfyUI looks like for a Live Portrait workflow.

If you still don’t get it, don’t worry! You are not dumb! You most likely don’t understand ComfyUI because you don’t fully understand the problem it’s solving for, that’s what this blog is for.

What’s your problem?

Suppose you are an artist — and let’s just say you have impeccable taste and outstanding vision already — what stands between you and greatness? The answer is quite simple

The boundary of your creativity is defined by your tools.

So, how does ComfyUI help unleash our creativity? Let’s start from the beginning. What is the one thing that AI content creators or AI artists do the most? It’s clicking that generate button over and over again until you tweak your way to a clip you like. If you are using a closed source video generation model like Runway or Pika, your bill can stack up quickly. When your money runs out, you no longer have access to those tools, so now you turn your attention towards publicly available models like Genmo or Hunyuan.

While these models are free, often times you don’t want to just run it vanilla because then you hardly stand out from the crowd and that is unacceptable! Artists want to have their own magic sauce, so they might add additional preprocessing or post-processing steps like inpainting, outpainting, noise, denoise, LoRA, upscaling, area composition, controlnet, live portraits and the list goes on. What’s more, these steps all come with its own set of parameters that you can tweak. The point is, you start to realize that this is really a modular and parametric process.

At this point, there are two options to present this modular workflow

1. it could use a linear notebook style like Google Colab, or

2. it could use something like a flow chart.

Those two representations are formally equivalent, but ComfyUI’s team made a smart choice to adopt a node-based drag-and-drop UI as many content creators are used to node-based workflows with color grading tools like Davinci Resolve.

It’s quite mesmerizing to admire a complicated ComfyUI workflow, but ComfyUI is more than what meets the eye. Remember, the ComfyUI backend is the engine that actually keeps everything running. Content creators don’t often have access to large clusters of GPU, so ComfyUI team spent a lot of time optimizing the backend (e.g. smart memory management during model loading) so that many workflows can run smoothly even without a RTX4090.

Lastly, the open source nature of ComfyUI means you can push the boundary as you like! If you don’t like something or need something, go build it!

Is Comfy in a Comfy spot?

Yea, duh. Just look at their stars!!! ComfyUI is currently on the GitHub Top 150 list, and it has 10,000+ custom nodes written by the community members, which really goes to show how much ecosystem adoption it had gathered almost organically.

But what really is the best indicator of success for an open source project? Is it really the number of GitHub stars? Is it really the number of active contributors? If it were up to me, I’d say those are all just scratching the surface. What you can’t fake is how drunk people are at its community meetup. I, for one, was pretty drunk. Incredibly unprofessional, I know, but I was just really thirsty from all that talking. The event enjoyed a wide spectrum of attendees — filmmakers like Paul Trillo, brand experience agency like Appmana, AI motion graphics startup like glyf, AI film competition like Project Odyssey, and boring BD people like me.

Here’s a recap of all the interesting things from the event, from my best recollection, which is not very good.

Met one of my favorite YouTubers, Theoretically Media! This man probably spends 25 hours per day on experimenting with AI video workflows.

Third time running into Teslanaut! Give him a follow! Here’s my favorite shot from his latest work Daydreaming.

Sean, Head of Engineering at BentoML, demoed a really nifty tool called Comfypack, which makes it super easy to share and run ComfyUI workflows by packaging everything you need to run a workflow such as model versioning, custom node, python dependencies, etc. In short, it’s like a docker-compose file for ComfyUI and it allows artists to spend more time creating and less time wrestling with the dependencies.

Next Up

So what’s next for ComfyUI?

• Hiring! They are hiring across the board in design, engineering and ops. This is a great team and they know how to have fun.

• New functionality like better Mac compatibility, training capability, notebook mode, subgraphs and easier collaboration.

That’s it for now. As always, please subscribe and make me feel needed, thanks.

Subscribe now

Despite splitting the tech stack into two parts, the visual and audio elements are not mutually independent from each other. The score must work with the scene, the sound effect must match the action, and the lip movement must match the speech. In short, audio and visual have to exist in harmony, as demonstrated in the image below.

Music

My favorite part of editing is finding that perfect piece of music. If you got the right music, you double the magic. Music has the power to render additional meaning or influence the emotion of a scene. Watch the following video for a fun demo.

Some people believe that the end game for AI music generation is about talking to a computer and expect Bach to reincarnate, but I firmly believe that’s not the case. Instead, AI music generation is more about generating inspirations. Granted, if what you are after is generic elevator music, then by all means just type in some words and hit ‘generate’. Sure the infinite monkeys may actually have a stroke of genius once in a long while, but good music simply cannot be generated just by prompting.

Keen readers may ask — why didn’t you take this stand towards AI video generation? Why do you believe prompting can yield good video but not good music?

Good question and let me explain. If you paint a mental picture of a scene, you can usually describe what’s in this footage, how they are interacting with each other and the overall visual style of the scene. In other words, you visualize then verbalize. However, for music, the equivalent of “visualization” is humming a melody. In that case, just grab a keyboard and start composing. Modern softwares like Logic or Ableton allow you to be your one person band already. Text prompts that describe music genre, mood, pace, etc simply do not have the intentionality that is required for good music.

As I result, I never intended for AI music generation tools to be the actual place for production, and the following table does not account for editing features like stemming or in-painting. There are also features like mood, genre, bpm, instrument dials that I don’t care about because I believe the model will just get better at complying with the prompt so those hard coded dials will be obsolete.

Most notably, I find the combination of Upload Audio + Extend feature surprisingly useful in my own creative process. In my 2025 New Year piece, I was stuck for a really long time on one section. Since I was writing this blog, I fed Suno my composition up til that point and asked Suno to extend it. I generated about 20 minutes worth of melodies and in the end found 15 seconds that had the spark. What’s more interesting is that I would never come up with those melodies myself, and deviating from my own creative inertia and toying with the anticipation made this piece much more interesting. Can you guess which section is AI generated melody?

Sound Effect

Moving on, we are now in the sound effect department. Here’s a demonstration on how important sound effects are.

Sorry, wrong video. Here you go again.

Most of the tools only have the ability to generate sound effect through text prompting, but Lightricks, MovieGen and FineVoice have this neat feature that can generate more fitting or tailored sound effects based on the actual video footage.

Voice

The three main voice capabilities are

1. Text to speech: this comes in handy when you need a narrator that talks over the footage such as a documentary host

2. Voice changer: you might decide that you don’t like the actor/actress’s voice and want something more cinematic

3. Voice clone: what if you screwed up an actual voice recording and cannot get the actor/actress to re-record? Try voice cloning, with consent of course.

Next Up

Next, I will be making an AI short film using the tech stack I just went over, and god knows how long that’s gonna take or if I will make anything that I find worth sharing. So I will be publishing on some other fun topics. Stay tuned!

Subscribe now

You can’t afford much so you will probably end up writing, directing and editing yourself. However, at the very least you need to figure out

• What the audience will see: Cinematography, acting, lighting, color grading…

• What the audience will hear: Sound design, music…

That’s the bare bone of course. Depending on your film, you might need other cool people like intimacy coordinator, weapon master, make-up artists, etc. I won’t go into details as I have literally zero experience in that department, the closest thing I had to a make-up artist is my mom applying sunscreen to my dad.

Sidebar — at the very least, you should really be figuring out how to make the audience feel something, but let’s stay pragmatic here.

The Visual Stack

AI tools can help create cinematic footage in two ways

1. video generation (from scratch, i.e. vision to visual)

2. video modification (restyle, object removal, etc)

Generating a video is easy if you know how to communicate your creative direction effectively. Simply feed the AI with a text description of your vision with an optional reference image, and voila! You now have a footage! Some advanced capabilities include

• end frame control: this allows you to dictate the shot the footage should end on, which is helpful if you have a sequence of key frames that needs animation.

• camera movement control: this allows for more precise controls of your camera work such as pan, tilt, zoom, etc.

• motion brush: this allows you to specify motion for entities in the reference image

Video modification can come in handy when you have a video already but

• want to transfer it into a completely different style

• want to add, replace or remove something from the footage

The following tools are not publicly available yet. I was going to call out Sora specifically for the insufferable wait but they did launch today!

Video Upscaling

Keen readers might notice that none of the tools generates 4K footages. In fact, you often have to pay extra just to get 1080p, not to mention that some tools just straight up don’t have anything more than 720p. Are we doomed?! How can I watch anything less than 4K?! Luckily video upscalers can save the day here.

There are a couple players in the vertical as well, and you can clearly see how much of an improvement you might get. This image came from a demo of one such tool called Topaz — original image on the left and upscaled version on the right. Pretty amazing, right?

Fun Effects

Lastly, there are other models out there with some pretty fun video effects, which is particularly interesting as they really lean into the AI-ness. For instance, Midjourney’s image generation video is pretty cool, it’s like an MRI of the AI brain.

Replicate also hosts some models that lean into their own styles such as this Cog Stable Diffusion Model that creates cool interpolation among key frames.

Pika also released a suite of special video effects called Pikaffects. My favorite is this effect that can pretty much inflate anything (but your ego).

Next in line, both Luma and Sora have the ability to create infinitely looping video by morphing the video back to a previous frame. This clip below came from an OpenAI demo, where we see a mountain biker realizing that it’s really all downhill from here.

Lastly, Runway greens screen could save the day when you want to isolate certain elements of the video and overlay it on other visual elements.

Next Up

Alright, now you have everything you need…

to make a silent film, so in the next post we will cover the audio stack for AI film making, stay tuned!

In the meantime, please subscribe and make me feel needed, thanks.

Subscribe now

Incidentally, getting rejected by Sundance and working at AWS has nothing to do with each other, I just wanted both pursuits to work out I guess. Ideally, the pursuit for productivity yields enough money to fund my pursuit for art. No strings attached, full creative liberty.

Did I think my film was actually good enough for Sundance? To be honest, I thought the opening scene was pretty damn good. Upon reflection, it is becoming absolutely clear to me that I didn’t get in because of my camera, so I am now a proud owner of Sony FX30 and 2 lens after a hazy Black Friday afternoon.

Anyways, I am currently starting my new job at AWS's AI/ML team doing BD (Business Development), and I’ll have future essays covering BD stuff as I grow in this capacity.

I’m still onboarding at AWS right now and I have been trying to figure out this question — what do I want people to think of when they hear my name?

• One of my favorite bloggers is , so naturally I thought about AI in defense. I even went to a couple hackathons, but eventually I realized that I would be going nowhere in this field without the right pedigree, a.k.a at least a U.S. citizenship.

• Then I thought about AI in cybersecurity, but upon more introspection it seems like I just like the sound of it. It sounds serious and important, like Logan Roy. Am I Logan? Kendall? Roman? Shiv? Tom? or GREG? I don’t know, but I definitely was not crazy about CS161.

• Still learning about AI in fundamental science, and slowly recovering from the shame of dropping that Quantum Field Theory class in my senior year, sorry Prof. Murayama.

Then I thought, hey what about AI in film? This feels right — here’s my decision making process, I often suffer from analysis paralysis so I’m trying to keep this simple

1. I like film

2. Film is big

3. AI in film is early

See my demo below if you don’t think so, a reference image is supplied and the prompt was “A green alligator leaps out of the water and bites the giraffe's neck. Photorealistic, 4K.”

Clearly, we have some work to do, but I do believe that AI in film will trigger another Cambrian explosion in creativity. It will never replace taste, but I believe it will make it easier for everybody to create with less reliance on experience and capital.

So what will this blog be about? The blog name () should be revealing, it’s going to be about many things, but we shall start with a couple essays on AI in film.

The next essay will be about the tech stack for AI film making, please subscribe and make me feel needed, thanks.

Subscribe now

Share

Subscribe now