557: BYD’s 5-Minute Megawatt Flash-Charging, China's Wild EVs, CEO Incentives, TSMC U.S. Fabs Economics, Cloudflare's AI Labyrinth, Google Gemini 2.5, and Reinventing Hollywood

"like a shiny new kitchen without a chef"

One idea I taught was the importance of regaining presence and clarity of mind after making a serious error.

This is a hard lesson for all competitors and performers.

The first mistake rarely proves disastrous, but the downward spiral of the second, third, and fourth error creates a devastating chain reaction.

—Josh Waitzkin

🗣️💬🧠💭➡️👤❓🤔💡❌✅🤷‍♂️ The illusion of transparency is a cognitive bias that causes us to overestimate how well others understand our thoughts, feelings, and intentions.

When you explain something that seems obvious to you, the other person typically understands far less than you assume. This happens because you have context in your head that they lack, making you believe you’re being much clearer than you actually are.

This has practical implications: Be more explicit in your communication than feels necessary, proactively ask for feedback to confirm understanding, and keep in mind that what's clear to you might be completely opaque to others.

This awareness can prevent countless misunderstandings. It’s especially valuable when it comes to invisible misunderstandings, which are the most insidious kind because both sides don’t realize that there’s confusion, so the error may not get corrected for a long time, if ever.

Wait! Did I explain all this clearly? 😅

💧🚰 In Edition #552, I wrote about buying a reverse osmosis countertop filter.

Having lived with it for a month, I can say a few things about it:

• It’s easy to set up and pretty well designed.

• My main UX complaint is that I'd prefer if it automatically reverted to default settings after a while, rather than staying at the last-used volume and temperature until manually changed.

• The biggest benefit — other than cleaner water — is the instant hot water. It’s very convenient for quickly making green tea or Aeropress coffee (my wife’s preferred brewing method). That feature alone makes it worth the counter space.

• It’s a bit annoying to have to refill the reservoir. We drink a lot of water, so we often have to refill it 2-3x per day. A system that is directly connected to plumbing with a faucet at the sink would be much more convenient. We may upgrade to that eventually.

• When I first set it up, I noticed that the water tasted different from what we got out of the Brita we had, likely because of the removal of mineral content. I worried I wouldn't adjust, but within a few days, it tasted completely normal. Now, I don't even notice it.

💚 🥃 🙏☺️ If you’re a free sub, I hope you’ll decide to become a paid supporter in 2025:

---

🏦 💰 Business & Investing 💳 💴

⏳ Median S&P 500 CEO Tenure: Less Than 5 Years ⌛️

One of the recurring messages from business greats is the importance of long-term thinking, building your life’s work, the magic of compounding over long periods, the value of experience and domain knowledge, how long it takes to build a high-performance company culture, etc.

Is it any wonder most CEOs don't practice this when their median tenure at large public companies is now less than 5 years?

Their incentive structure basically demands quick wins, and if they need to mortgage the company’s future to do that, well, it’ll be someone else’s problem. This is similar to politicians kicking the can down the road because they know they won’t be around to face the consequences of their choices.

the median tenure among the S&P 500 companies has decreased 20% from six years in 2013 to 4.8 years in 2022. [...] Between 2013 and 2022, the 20% decrease in median tenure compared to the approximate 5% decrease in average tenure indicates a considerable amount of chief executives are experiencing notably shorter terms in their leadership roles.

Warren Buffett is one of those people skewing the average upward: He has been CEO of Berkshire Hathaway for 55 years as of 2025! 🗓️🗓️🗓️🗓️🗓️🗓️🗓️

The disparity between median and average tenure figures signifies there is a subset of chief executives who have held their positions for relatively longer periods, skewing the average upward. In the aforementioned Barron’s article, Andy Serwer coined the term “forever CEOs” to describe CEOs with lengthy tenures. One example is Jensen Huang, the CEO of Nvidia, who has maintained his position for over 30 years.

Chief executives with over 20 years of tenure represented 8% of the S&P 500 in 2022 (when this data was compiled). Over 50% had been in their role for 10 years or less.

These short tenures are somewhat of a proxy for the agent-principal issue that plagues so many companies. Those who stay around longer are more likely to be founders, or at least to think more like founders.

Notable exceptions to the revolving door for non-founders include Satya Nadella at Microsoft (who became CEO in 2014), Lisa Su at AMD (also became CEO in 2014), and, despite the recent headlines, Tim Cook also fits the bill, having been CEO since 2011 (14 years ago) and having taken Apple from $300 billion in market cap to $3.3 trillion.

I’m not sure how markets could encourage more long-term thinking and long-tenured CEOs. In some cases, it’s very good that CEOs don’t stick around for long, because they are terrible and destroy value. But I feel like there’s probably a crop of CEOs who are quite competent and would make better decisions *if* they had a longer time horizon 🤔

Source.

h/t friend-of-the-show Nicoper (💚 🥃)

⚡️🔌🚘🪫→🔋⏱️ BYD’s 5-Minute Megawatt Flash-Charging & the Wild World of Chinese EV Innovation 🤔

The EV stuff coming out of China these days is pretty wild.

In the same way that Tesla made the old ICE carmakers seem slow and afraid to experiment and innovate, companies like BYD are making Tesla seem like a “legacy” EV maker (though possibly in good part because the CEO has been distracted with other things).

Check this out:

BYD launched the Super e-Platform, featuring flash-charging batteries, a 30,000 RPM motor, and new silicon carbide (SiC) power chips. The platform upgrades the core electric components, achieving a charging power of 1 megawatt (1000 kW) and a peak charging speed of 2 kilometers per second, making it the fastest for mass-produced vehicles - 5 minutes of charging for 400 kilometers of range. The Super e-Platform also delivers a single-module single-motor power of 580 kW and a top speed over 300km/h.

They published this video showing a 5-minute charge at close to ONE MEGAWATT 🤯

So in America terms, that’s roughly 250 miles of range in five minutes. Pretty competitive with stopping at the gas station.

Of course, announcing things is a lot easier than shipping them and having them work reliably in the real world. I can’t help but wonder how dumping this much energy into a battery over such a short period of time won’t damage it.

The physics of “flash charging” (because “fast” isn’t fast enough anymore) implies pretty gnarly thermals, and batteries are generally sensitive to that.

From the positive to the negative electrode, it has built ultra-fast ion channels, reducing the battery's internal resistance by 50% [...]

With the support of ultra-high voltage (1000V) and ultra-high current (1000A), BYD "Flash Charging Battery" can achieve the world's largest mass-produced charging power of 1 megawatt (1000kW).

BYD also unveiled the other half of this technology, what they call Megawatt Flash Charging stations:

Lian Yubo, Executive Vice President of BYD and President of the Automotive Engineering Research Institute, stated: "BYD has self-developed the world's first all-liquid-cooled Megawatt Flash Charging terminal system, with a maximum output capacity of up to 1360kW. In the future, we plan to build over 4,000 'Megawatt Flash Charging stations' in China (Source)

Good thing China is also good at building new power stations, because 4,000 of these megawatt stations is a big draw on the grid, even assuming they won’t all be utilized simultaneously all the time. But it’s still a theoretical 4GW load (more if you include losses).

On the financial front, BYD’s latest annual report shows 777 billion yuan ($107 billion) of revenue for 2024 (with a 19% gross margin and 5.2% net margin, though I’m not familiar enough with Chinese accounting rules to know how comparable these figures are to US GAAP).

That’s a 29% year-on-year increase, beating Tesla’s $97.69 billion for 2024. But BYD doesn’t just make EVs, so it’s not entirely 🍎 to 🍎 (though to be fair, Tesla also doesn’t just make EVs and has an energy division that sells stationary storage).

Berkshire purchased 10% of BYD in 2008. It was a Munger investment. Over time they’ve sold down the stake, but at last disclosure, it was still close to 5%.

The stock is up about 3,000% since that original purchase (it depends on which currency), representing about a 23.5% CAGR over 16+ years.

If BYD delivers on this stuff, the EV race just got a lot more exciting.

🏗️🐜🇺🇸 Producing wafers at TSMC Arizona is only 10% more expensive than in Taiwan 👀🇹🇼

Good news for the U.S., though perhaps not-so-good news for Taiwan:

The dominant factor of semiconductor production cost is the cost of equipment, which contributes well over two-thirds of overall wafer expenses. Tools made by leading companies like ASML, Applied Materials, KLA, Lam Research, or Tokyo Electron cost the same amount of money in Taiwan and the U.S.; they effectively neutralize location-based cost differences.

A major source of confusion about wafer prices comes from labor costs. Wages in the U.S. are roughly triple those in Taiwan, which many mistakenly take as a significant factor in chip production. However, with the advanced automation of today's wafer fabrication facilities, labor accounts for less than 2% of the total cost

This doesn’t mean that expertise isn’t scarce in the U.S. and heavily concentrated in Taiwan, though.

A fab without highly skilled workers is like a shiny new kitchen without a chef. 👩🏻‍🍳

Let’s also not forget that fabbing the wafers is just one part of the process of going from a pile of sand to a working computing device in someone’s hands or in a data-center, though:

It should be noted that wafers that TSMC currently produces at Fab 21 travel back to Taiwan to get diced, tested, and packaged. Some of them then go to China or elsewhere to be put into actual devices; some will travel back to the U.S. (Source)

TSMC also plans to build packaging capacity in the U.S., but this will also take some time.

---

🧪🔬 Science & Technology 🧬 🔭

🤖🚪🌀 Cloudflare AI Labyrinth: Purgatory for Bad Bots 🔥🤖🔥

This is clever:

AI Crawlers generate more than 50 billion requests to the Cloudflare network every day, or just under 1% of all web requests we see. While Cloudflare has several tools for identifying and blocking unauthorized AI crawling, we have found that blocking malicious bots can alert the attacker that you are on to them, leading to a shift in approach, and a never-ending arms race. So, we wanted to create a new way to thwart these unwanted bots, without letting them know they’ve been thwarted.

To do this, we decided to use a new offensive tool in the bot creator’s toolset that we haven’t really seen used defensively: AI-generated content. When we detect unauthorized crawling, rather than blocking the request, we will link to a series of AI-generated pages that are convincing enough to entice a crawler to traverse them. But while real looking, this content is not actually the content of the site we are protecting, so the crawler wastes time and resources.

As an added benefit, AI Labyrinth also acts as a next-generation honeypot. No real human would go four links deep into a maze of AI-generated nonsense. Any visitor that does is very likely to be a bot, so this gives us a brand-new tool to identify and fingerprint bad bots, which we add to our list of known bad actors.

The Sisyphean crawler is trapped in an infinite maze!

This raises an interesting question: Will the dumber bots that don’t realize the content is fake will scrape it and include it in the training data for some future model?

Could we end up in a scenario where a notable percentage of AI training data is AI Labyrinth nonsense? 🤔

At least they try to create “real” content, so maybe it’s not worse than a lot of the stuff that is being scraped online anyway. In fact, could Cloudflare accidentally become a synthetic data provider for AI training..?

To generate convincing human-like content, we used Workers AI with an open source model to create unique HTML pages on diverse topics. Rather than creating this content on-demand (which could impact performance), we implemented a pre-generation pipeline

To keep their costs low, they’re probably using a tiny version of Llama or Mistral, so I’m guessing the quality of this text won’t be high.

Just good enough to fool a bot.

We found that generating a diverse set of topics first, then creating content for each topic, produced more varied and convincing results. It is important to us that we don’t generate inaccurate content that contributes to the spread of misinformation on the Internet, so the content we generate is real and related to scientific facts, just not relevant or proprietary to the site being crawled.

This pre-generated content is seamlessly integrated as hidden links on existing pages via our custom HTML transformation process, without disrupting the original structure or content of the page… We also ensured that these links remain invisible to human visitors through carefully implemented attributes and styling.

More details here.

🤖 Google Gemini 2.5 — New LLM Arena Champion 🏆

Impressive pace of improvement from Google:

Our first 2.5 release is an experimental version of 2.5 Pro, which is state-of-the-art on a wide range of benchmarks and debuts at #1 on LMArena by a significant margin.

Gemini 2.5 models are thinking models, capable of reasoning through their thoughts before responding, resulting in enhanced performance and improved accuracy. [...]

We’ve been focused on coding performance, and with Gemini 2.5 we’ve achieved a big leap over 2.0 — with more improvements to come. 2.5 Pro excels at creating visually compelling web apps and agentic code applications, along with code transformation and editing.

On SWE-Bench Verified, the industry standard for agentic code evals, Gemini 2.5 Pro scores 63.8% with a custom agent setup.

Much like what OpenAI said about GPT-5, all Gemini models will incorporate reasoning going forward:

with Gemini 2.5, we've achieved a new level of performance by combining a significantly enhanced base model with improved post-training. Going forward, we’re building these thinking capabilities directly into all of our models, so they can handle more complex problems and support even more capable, context-aware agents.

The benchmarks are impressive, though as usual, I think it’s real-world use that will reveal the true “flavor” of the model, and its strengths and weaknesses:

I can’t wait to get my hands on it and try it. I’ve been very impressed by Gemini 2.0 Pro and Gemini 2.0 Flash Thinking.

---

🎨 🎭 The Arts & History 👩‍🎨 🎥

🎥 Jason Carman on Reinventing Hollywood, Sci-Fi’s Role in Inspiring Science, Embracing Agency, and Re-Igniting Optimism 🚀🛰️🛰️🛰️

If this conversation between Jim and Jason (J&J) doesn’t fire you up to become better at whatever you’re doing and building cool things, I don’t know what will! 🏋️‍♂️

They also had a great discussion about Bob Dylan and the film ‘A Complete Unknown’ (2024), which I wrote about in Edition #544.

Make sure to check out ‘New Space’, S3’s new film about the modern space industry. It’s excellent and it’s free. 🆓