CO2 footprint of TikTok like that of Greece? Hell no.

internet

I’m not a fan of TikTok. I think it has an almost evil genius in its ability to condition minds into something akin to human “Legehennen” (battery hens). In general, I have no issue with left-leaning publications. But let’s call out nonsense where we see it.

The Guardian article from December 12, 2024, claims that TikTok has a carbon footprint on par with Greece. That’s a serious claim. But if we break down the math, it starts to look suspicious.

The Math Behind the Claim

Let’s take the numbers they’re using and do some calculations:

  • Start with a total of 50 million metric tonnes of CO₂.
  • Divide by 1 billion users: that’s about 50 kg of CO₂ per user per year.
  • Assume a carbon intensity of 0.475 kg CO₂ per kWh. At that rate, 50 kg CO₂ corresponds to roughly 105 kWh/user/year.
  • Spread over a full year (8,760 hours), that’s about 12 W continuously per user.
  • But users only spend about 276 hours/year on TikTok. If we concentrate that 12 W into just the active usage time, you get 380 W per active user session.

So the math itself is internally consistent: given the initial assumptions, you end up with 380 W per active user.

Does 380 W per User Even Make Sense?

No. 380 W per user just for streaming short-form video is wildly implausible. It’s off by orders of magnitude. To put this in perspective, 380 W is like running several light bulbs just to serve you a single video feed. That’s massive.

In reality, data centers serving millions of users simultaneously do so at a much lower per-user energy cost. The number from these calculations strongly suggests that the initial carbon footprint figure or some fundamental assumption in that claim is questionable.

If TikTok truly had the same carbon footprint as an entire country with millions of citizens and heavy industry, it would be a global scandal. But we need to consider whether that initial data point (50 million metric tonnes of CO₂) makes sense, and what it includes. Is it capturing the entire business operation, including servers, distribution, all upstream supply chains, and user device energy?

Conclusion

It’s critical to apply common sense and do the math before swallowing claims like this. Even if you dislike TikTok, as I do, you have to recognize that equating its energy usage to that of Greece might be an exaggeration—or based on questionable data or assumptions.

At the end of the day, 380 W per user to serve video streams is “No Fucking Way” territory.

[[ O1 Pro took its time to write this, and during that creation I might even have used 380W ]]

synthetic

AI

Obviously, I have no clue about how an LLM works. Having said that, I cannot fathom how synthetic data could improve things much. My suspicion is that it gets favored since it is so predictable to generate, while finding and filtering actual information is hard and messy.

I just tried to generate the most simple code for Apple Metal — it never worked. Looks like GPT could use a closer look at actual example code from that area. And, like that, there are probably countless other aspects of what humans have come up with so far that warrant a closer look.

“Organizing the world’s information and making it accessible”

That is still a good guideline, as is “Don’t be evil.”

Their mileage might vary.

photoshop spacebar not activating hand move?

cli photoshop unix

It has been like this forever for me. Most of the time. I realized that if I have a terminal open in MacOS then in PS the spacebar no longer can activate the hand move tool.


Funny that this bug can exist for so long. The number of people using that handy (!) shortcut AND using a shell if probably small.

ESP32-WROOM-DA internal LED, serial: pick one

technology

Starting out with a ESP32-WROOM-DA it might be tempting to make the internal LED blink. Turns out it is trickier than thought. Of course one needs to pick the correct board, otherwise you might see:

avrdude: stk500_recv(): programmer is not responding
avrdude: stk500_getsync() attempt 1 of 10: not in sync: resp=0x00

But even after that it might be that the LED does not blink if you use a different PIN ID than 1 – lots of examples are around with 2,4, 13 etc


A common way debug is to output something to the serial port. Which works, UNLESS you use PIN ID 1 to turn the LED on or OFF. In which case the serial output turns into gibberish.

No big deal, but can be confusing.


HP Color LaserJet Pro 4202dn with macOS Sonoma : don’t!

Apple history

Since more than 40 years there are HP printers that are connected to Apple computers. It is a simple thing: Some document is in the computer, it should get printed on paper. Happened in this specific Apple/HP combination many billions of times. HP’s market cap is around 35,470,000,000 dollars right now. That of Apple is 3,400,000,000,000.

Two devices connecting for this super simple thing. Should be easy in 2024.

IT IS NOT.

Fucking not.

The latest macOS Sonoma 15.1 will not work reliably with the HP Color LaserJet Pro 4202dn.

Printing a document multiple times will only work when you open the print dialog as many times and print one page.

The printer insists that it wants to double-sided print.

I spent hours with this. Tried everything. It did not work out of the box. Neither does it now.

Again: All I want is to print a simple text document.

Enshitification.

Here it is not intended. But infuriating. I love my MacBook Pro. The design for the HP printer I personally like. The hardware of it seems to be built right. But the software that moves the data to the printer does not work with the software of the printer. A task that should be solved since more than 40 years.

State of AI – October 2024 Edition

AI

Actual AI usage for me is a mixed bag.

Recently, Google offered a free trial of their Gemini AI within Gmail. Unfortunately, my first interaction couldn’t have gone worse:

I received an invoice for a lawnmower repair. Out of curiosity, I clicked on the ✨ icon to see if Gemini could generate a reply. To my surprise, it said it could not recommend a response because the email seemed suspicious. When I asked for more information, it claimed that the email was pretending to be from Capital One. I told it that this was not the case. Then it responded, “You’re right, this email is pretending to be from Wells Fargo,” which was equally false.

If an AI fails on the very first task—understanding the nature of a basic email—how could it possibly be useful? How can Google ship such subpar software? I can’t imagine the pressure they must be under! It’s still surprising that they would risk making such a flawed system the first AI experience for many users.

On the other hand, I had to cancel a DSL contract recently and found a more positive experience. I located the onboarding letter from the provider as a physical document, took a picture, and sent it to ChatGPT-4 with a request to generate a cancellation letter. Within two minutes, I had a completed draft—no thinking required, except for verifying the numbers were transcribed correctly. This saved me at least ten minutes of tedious work. Quite nice!

In more mundane AI experiments, I spent some time reading Stephen Wolfram’s latest essay on the nature of time. Using a notebook-based language model, I got a decent summary and a good introduction to the material.

Meanwhile, OpenAI’s O1 preview did a good job explaining a few concepts and was patient in addressing my requests to relate the theory to other topics like quantum wave collapse, Hubble Tension, and the need for dark matter/energy.

Another interesting angle into the state of AI is explored in the Lex Fridman podcast featuring the Cursor team. It dives into the challenges and future possibilities of integrating AI into coding environments, providing insights into the ongoing development process.

Sudo MacOS 15.0 Sequoia and timestamp_timeout

Apple OSX

It seems that MacOS 15.0 does not respect altered timestamp_timeout values.

Trying to find out why, and how it might be fixable I found nothing. Instead the ability to do sudo via touch ID came to light. Looking for

/etc/pam.d/sudo_local.template

will get you details. I tried for 5 minutes. And it is nonsense!

You work in a terminal, a sudo is needed. With this config enabled, a popup shows up, you have to reach for the mouse, give the window focus, and then do touch id. Typing a password every three days is SO much quicker.

Nexus

AI books

I recently finished reading Nexus and thought I’d share my thoughts. To start with, let me say that Sapiens was an incredible book. It was so good that, for a while, I wouldn’t hesitate to pick up any book written by Yuval Noah Harari. However, Homo Deus didn’t quite reach the same level for me, and neither does Nexus.

That being said, Nexus does have some interesting ideas and provides helpful historical context. The core notion that ‘more information’ doesn’t automatically translate into ‘truth and happiness’ is something I can definitely agree with. But overall, the book doesn’t deliver in a way that would make it stand out.

The author is undoubtedly a smart individual, and his intentions seem to be in the right place—trying to nudge readers in a positive direction. But the book itself feels like a wild mix: on the one hand, it contains interesting insights rooted in solid historical observations. On the other hand, these are paired with personal musings and extrapolations that, while thought-provoking, are not necessarily backed by extensive knowledge.

One of the central themes of Nexus is the state of the world in relation to AI, along with some views on how we got here. Interestingly, there’s no mention of transformers (the AI architecture, not the robots), which is quite telling about the author’s perspective.

In summary, while Nexus is an intellectually stimulating book in parts, it doesn’t quite live up to the standard set by Sapiens. It’s a mix of deep insights and personal thoughts, the latter sometimes feeling like they stem more from the author’s compulsion to think about things rather than from a solid foundation of knowledge.

Drum and Fiddle

politics

It was hers to lose. Her lead in the polls was melting away over the past few days. Strange, since neither candidate actually changed anything, nor was there any actual news that would tip the scale. Maybe a slight dip in the job numbers? “It’s the economy, stupid.”

The stakes were high. After all, the B/T debate had a dramatic outcome. Seen from the perspective of the self-proclaimed stable genius, the subsequent candidate swap of the Democratic ticket was clearly the result of his very own remarkable debate performance.

In light of the H/T event yesterday, it seems clear that it was the gerontological issue of 46 that the cameras captured, which made his dropping out of the race inevitable. Those snippets that the Internet repeated painted a drastic picture of Joe Biden’s issues.

After the Harris announcement, the Republicans honed in on a narrative that she would be a lightweight, somebody who could not even do an interview, which logically implied that her fate in a debate would be along the lines of her future predecessor. Trump would be victorious in two debates, clearing the entire Democratic bench in two TV innings lasting a mere 3 hours in total. The best work. He actually believed that. Republicans had 9 years now to stop worrying and love Trump, so they went blindly along with that narration.

During those 90 minutes in Philadelphia, it fell on their feet.

Countless times, Harris did not answer the questions that were put in front of her.

Donald repeated dutifully, and with all his zest, the tropes that get him cheers during his rallies. Frequently, since Kamala wanted him to do so.

She played him like a fiddle. When Biden wanted to get nominated, and still had a more than ephemeral pulse, he said he should be the candidate, since he would beat Trump like a drum. Which he did. And yesterday, the man from Queens became her instrument.

Kamala did the only thing you can do if you are in a debate with DJT in 2024. And she executed a good plan very well. Sure, it is strange that who fills the 4- or 8-year king role in the country with the most nukes gets decided in an event, almost WWF-like, with commercial breaks that has nothing to do with content and which gets decided by how people enter the room and what they do in the next 5 seconds.

What a fateful handshake that was. Good to touch a man that doesn’t want to get touched. Smart.

When topics came to questions she didn’t want to dwell on, she simply diverted to talk about his event crowd or inheritance sizes. And he took the bait, every time. Exploding in inner anger, he started to direct his rambles in a predictable manner.

A behavior that she eventually turned into her future presidency when she outlined that he would be equally easy to manipulate—something that became a stronger point by her letting people come to that conclusion by themselves instead of pointing it out explicitly. People like it if they are the smart ones in a story.

But their expectations were also that he would trounce her. So everybody would see what a lightweight she “is.” Hard to see how one could read that into yesterday’s event. I am sure Fox, etc., will find a way to edit things in this way, but they haven’t been given much to work with.

Again, her triggering his anger at will was case in point to the argument she made.

Our Fading Memory

AI history

I wondered what the percentage of people would be that could remember certain events. Based on some simplifying assumptions:

  • Only 1 in 10 Americans could have a memory of JFK.
  • Around 20% could remember the Moon landings.
  • For 40%, 9/11 is a historic fact, but not something they lived through.

These numbers might seem surprising. Many of us tend to assume that more people are aware of or experienced key historical events. Our societal memory is skewed towards older generations, dragging what should be considered history into the present.

To reach these approximations, I looked at U.S. demographic data and assumed people don’t retain significant memories from before the age of 10. While this isn’t entirely accurate, it works well enough for the purpose of this exercise. This also explains why the graph doesn’t reach 100% as we approach more recent events.

The insights gained here are fascinating—at least to me. It’s likely that this analysis has already been presented somewhere, and possibly in a more thorough and insightful way. However, the real problem lies in trying to find it.

Sure, I could try using Perplexity—and I did, just now. But no, it goes off in the wrong direction. From my experience, it often becomes tedious to track down something specific. Google? Pointless. I gave it a shot again, and the results were garbage.

I even tried @gemini in the search bar, and it hallucinated:

“A 2019 Pew Research Center survey found that 86% of Americans remember the day Neil Armstrong first walked on the moon.”

Kagi? I keep my subscription because I like the idea behind it, and it uses Google search without those pesky ads, but the results aren’t much better.

The truth is: Search seems broken. Fortunately, we are no longer entirely dependent on it. That hallucination from Gemini reminded me of something I’ve observed over the past few weeks: GPT-4o doesn’t seem to hallucinate for me anymore. I thought I caught it yesterday with this:

“C5b, C6, C7, C8, and C9 assemble together to form the MAC”
Are you sure? Sounds like you’re making it up. Please provide an external source.

It cited the right paper and textbook and was correct.

Getting the raw data and cooking up the graph was a good first project for Cursor. It worked. I might sound like an OAI fanboy here, but things got better when I switched the model to GPT-4o from its Claude Sonnet default. I really liked Claude for coding, but even in the paid version, it has a usage quota. So, I avoid jumping into the pool with walls when I can dive into the ocean instead. And considering the water is mostly the same for what I do, I stick with GPT-4o.

Without AI, I could have coded this myself, of course. It’s not rocket surgery. But it would have taken me more time and mental energy than I would have been willing to spend on it. It’s not that important to me. Which is the impact of AI: We can do things now that we didn’t do before.

And that, I feel, is very good.



What isn’t good is that 4 attempts to instruct gpt4o to NOT precede the spell checked text with “Here is the full blog post with the new section included, ready for you to copy and paste:” all failed. It even made a memory, but of what? Crazy how simple things still don’t work …