Category: AI

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 …

from Bartender 5 to ice

AI economy internet OSX

I look at my macOS Sonoma 14.5 screen for a couple of hours a day. I wasn’t a fan of how Google and Adobe placed their icons in the menu bar, as I never need them. When GPT-4 added its icon—which is actually helpful for me—I felt that cleaning up that part of my surroundings would be nice.

So, I learned about Bartender 5. I’m not sure if I will end up spending money (22EUR) to have a tidy upper right corner of the screen. Installation was simple, albeit a bit surprising since the app needs permission to screen record and access another setting, which I later disabled. So far, the menu bar has not reverted.

If you’re befuddled like me by the Byzantine mess that Mac preferences have become, check Privacy & Security > Accessibility and Screen Recording if you’d like to revert the settings after the menu has been set to your liking.

Some Concerns About Bartender 5

Seemingly—and I learned this only after my installation—the single developer, Ben Surtees, sold his wonderful and intricate app to Applause.dev earlier in 2024. It wasn’t smooth sailing. It’s not reassuring that Applause.dev seems to have no list of applications they have purchased so far. Their website only caters to developers who are looking for a cash exit.

I feel that Ben is totally fine with selling what he built. The app looks and works amazingly well. Single developer apps sometimes have a five-star restaurant touch to them. Only a person who cares intensely can get all those details right. More often than not, it never happens.

The world, however, does not automatically and magically reward individuals for their efforts. Regardless of how awesome somebody codes and thinks, linear scaling the revenue based on the efforts is—sadly—not a reality at all. It’s insane that this myth lingers, much like the lingering scent of perfume in an empty subway tunnel—ephemeral. It might even just have been a shattered perfume bottle a perfect while ago.

Challenges of Being a Single Developer

Applause.dev probably has no shortage of developers trying to sell what they’ve made. Life as a single developer is hard as it is: Customers expect support from everywhere in a non-linear fashion. Three percent of the prospects create 90% of the troubles and work, and of those, most will not even buy.

To make it worth your while, you want to have around 200,000 in revenue a year. At that number, you would still be better off financially in some corporation. Let’s assume that’s not your thing, and that you’re not in the daily-ramen phase of your life either. For an app, that means 10,000 people a year need to go through the motions. Let’s assume an amazing conversion rate of 20%. You have to deal with roughly four sad/bad/stupid/horrible people a day. All the while, you must be nice, since you can’t afford the potential backlash of losing your temper in the wrong way. That alone takes at least one hour out of each day, every day. If you take a weekend off, there are eight more of those waiting on top of the four that come each day.

Of course, you need to funnel them to you, which is even harder and becomes increasingly difficult each month. The internet is full of content—so full that you have basically zero organic reach now. You need to pay Google or Meta for eyeballs, and the cost of that increases continuously.

So, wanting to exit makes sense. I get it.

The Bright Future of Single Developers

On the other hand, the future for single developers—and, in turn, for us—is incredibly bright. A person who knows what they are doing can implement many things between 4 and 20 times faster than just a year ago. So, the scale of what that one-person bottleneck can churn out is absolutely amazing. I predict that we will see a rise in single-person solutions that will make all our lives so much better. I feel that this will be the first real, positive impact of transformers and the abundance of CUDA compute: single people doing things where you once needed 20 developers. And if you have 20 good developers, you usually need between 100 and—how many people do Meta, Google, and Apple have again?—around them. OK, maybe a bit over the top with that one. But six weeks ago, we all learned the hard way that Crowdstrike only had 19 and not 20 good developers among the 8,000 people they employ.

What Will I Do with My Menu Bar?

So, what will I do with my menu bar? For now, nothing. In four weeks or the next time I log out, I will have to consider whether I should turn my network off, give Bartender the rights it needs, change the settings again, and move on. Or should I look at a free alternative like Ice?

Or just go back to the old ways. Not sure.

Bartender is tricky, though. I don’t think Applause is particularly evil—they could be, they could not be—but having an app install base of Mac users who most likely let the config options for controlling your computer and making screen recordings on makes them a very juicy target for malicious third parties. Or is that a fourth party in this constellation? Not something I’m eager to partake in.

September 18, 2024: With the macOS 15 update Bartender vanished, since I didn’t give it any permissions.

Turns out installing Ice is super simple, and using it equally so. It also wants screen recording and accessibility permissions. Other as with Bartender, it wants to keep those permissions, otherwise the view reverts to its default.

Summary

To sum it up, Bartender 5 offers a sleek solution to tidy up your macOS menu bar, but the transition of ownership from a single developer to a larger company raises questions. While the future of single developers looks promising, the challenges they face are significant. As for Bartender, whether I stick with it or explore alternatives like Ice, I’ll keep a close eye on how things evolve—both in the software itself and in the security concerns that come with it.

When the Internet Leads You Astray – Trust the Source

AI internet linux unix

Putting a USB SSD on My Ubuntu Machine: A Journey Through Confusion

Recently, I decided to add a USB SSD to my Ubuntu machine. Pretty straightforward task, right? So, I set it up, partitioned it, and formatted it with an ext4 filesystem. Then came the question: What happens if the drive isn’t connected when I boot my system?

Since it’s an external drive and not critical to my system’s operation, I didn’t want my machine to throw a fit if the SSD wasn’t present at boot time. So naturally, I turned to GPT-4 for advice.

GPT-4’s Advice: The “nofail” Option

GPT-4 responded quickly and gave me a clear suggestion: use the nofail option in the /etc/fstab file. This would ensure that the system attempts to mount the USB drive if present but continues to boot even if the drive is not connected.

That made sense, and GPT-4 also reassured me that using this option for non-critical filesystems is common practice. But something bugged me. “nofail” sounded counterintuitive—shouldn’t this be used for important, “must mount” filesystems? So I turned to my trusty search engine to verify what exactly “defaults,nofail” meant.

The Internet Confusion Begins

I did a quick search using Kagi (or Google—take your pick, I got the same results). On the first page, I came across this page from Rackspace’s docs: Rackspace Docs on Linux – Nobootwait and Nofail.

It flat-out stated the opposite of what GPT-4 had told me! It described nofail as an option for critical filesystems, implying that the system would wait until the drive was mounted. This seemed strange since I wanted the system to boot even when the drive wasn’t there. This increased my doubts, so I dug deeper.

Deeper Dive – The Web Only Adds to the Confusion

I kept browsing, checking multiple sources. Each page seemed to explain nofail slightly differently. Some agreed with Rackspace, others said the opposite. At this point, I was more confused than when I started. How could such a basic option be so misunderstood on the web?

The Answer Was in the Man Pages All Along

Frustrated, I decided to check the man pages—the original source of truth for Linux users. Sure enough, GPT-4 was right. The nofail option is specifically there to ensure that the boot process does not stop or fail if the specified filesystem is not present. Perfect for my use case, as I wanted the system to keep booting even if the USB SSD was missing.

Conclusion: Crazy World We’re Living In

So here we are, in a world where search engines and online documentation can sometimes steer you in the wrong direction, while an AI (GPT-4) was spot on from the beginning. It’s crazy how something as fundamental as mounting a drive can be so muddled online.

I’ve learned two things from this: first, always double-check your sources—especially with something as complex as Linux configuration. And second, never underestimate the value of consulting the man pages. In this crazy, confused world of information overload, sometimes the simplest and most direct solution is the right one.

Oh, and yeah, nofail is exactly what you need for external drives that you don’t want to hold up your boot process. Crazy name? Maybe. But it does the job.

And, yes, this was written by gpt4o as well, based on the cryptic text

Intersting. 

putting USB SSD on ubuntu machine. 

Asking gpt4o what to do. Works. 

worried about this being external drive that boot would not stop when it is not there. So I ask, and it says should be fine, but to use “nofail” to be sure. 

I think: weird naming for option in open source. 

Use Kagi (or google, the same result) to look for “defaults,nofail meaning”

FIRST page 

https://docs.rackspace.com/docs/linux-nobootwait-nofail

says opposite of gpt4. 

I look further on the net. Confusion. Finally I look at the man page, 

gpt4o was right. The Internet as seen via search is plain wrong. 

Crazy world we are living in

random

AI

https://youtu.be/JfsnXXwu7-A

Hit the random wikipedia page 3 times.
Told gpt4o to make a story out of links I got and also a prompt that I can give to flux.
Took the resulting image to runway (weekest link here, sucks, but this is not about making something good)
Asked gpt4o to make a udio text for the image that I also gave
extended it in udio (twice, I guess) added intro and outro
Put it together in daVinci with obvious loops and fades.
Done.

Nothing got cherry picked here. I could basically automate this, since
I made NO CHOICES during the process. This is all the stuff how it fell out those various machines …

KI-Auswirkung

AI economy history

Bosch gab zum 125-jährigen Jubiläum 2011 ein Buch heraus. Auf Seite 54 findet sich diese Karte:

Ich fragte mich, wie das wohl als Animation aussehen würde. Früher wäre mir dieser Gedanke wahrscheinlich gar nicht gekommen. Das Gehirn wächst mit seinen Möglichkeiten. Heute ist es einfach möglich, aus dem Foto eine Animation zu erzeugen:

Das hätte ich zwar auch vor KI schon hinbekommen. Nur wäre es eben den Aufwand nicht wert gewesen. So wichtig ist es nicht für mich, die globale Entwicklung von Bosch zwischen 1897 und 1922 animiert dargestellt zu sehen. Wenn man KI benutzt, dann ist es nicht sonderlich aufwändig. Das Kosten-Nutzen-Verhältnis verschiebt sich. Nicht automatisch. Nicht magisch. Man muss ja immer noch wissen, was man tut, man muss wissen, was man will.

KI ist NICHT eine automatische Lösung. Es ist nicht so, dass alles heute mit einem einzigen Knopfdruck magisch entsteht. Amüsanterweise wird uns genau das versprochen, genau genommen seitdem es Computer gibt. Und trotzdem war es noch nie so.

Wie beim Eisberg gibt es in den aktuellen KI-Erwartungen aber auch Gruseliges unter der Oberfläche: Es war nie einfacher, Programmcode zu erzeugen, der scheinbar zu funktionieren scheint, es aber in der Realität dann nicht wirklich tut. Das war schon immer das Problem mit Programmierern, die ihre Arbeit nicht ausreichend beherrschen. Und dieser Personenkreis wurde plötzlich um das Hundertfache größer. Menschen, die in der Vergangenheit an Dingen wie Syntax, Dokumentation oder Lücken bei Stackoverflow scheiterten, können heute ihren Kunden und Arbeitgebern allerlei Unsinn unterjubeln. Und das passiert dann auch. Überall.

Blue Screen of Dichotomy

AI internet malware

Crowstrike, never heard of them. Lucky me. Sorry for all who got impacted.

Screenshot

While this could very well be the first case of a broader “the AI ate my homework”, reality is never that simple, easy or straightforward. No matter if the story works nicely.

Wikipedia featured the fix for a while. No idea if it is legit or will entirely burn your machine to the ground. Upon further reading it feels that the source for this “fix” is legit, but it might be that it is not a panacea. In my, very limited, understanding deleting those channel files might give the system a chance to reload valid ones on the next boot.

The actual root cause of this incident will be interesting though. Strange that it is possible to push something faulty to this many machines. One would think that avoiding this would be one of the core issues of an org like Crowdstrike.

Good luck to everybody affected. Directly or indirectly.

AI generated artists

AI art

I asked Claude to generate hypothetical artists, and then had gpt4o generate images based on them.

A Synthesis of Essence


Allegra di Firenze

Evelyn Zhao

Jasper Hawkins

“Unkempt Iris” listing @ CCFA

I gave it no guidance. Its choices are its own.

To me, these pages illustrate nicely the strengths and weaknesses of AI right now: The language is free from any obvious errors I would notice. The fabricated facts have some consistency to them.

And yet, it is all uninspired. One cliché follows the next. It is a dense condensation of all our prejudices and current assumptions. Utterly dull and uninspired. AI-generated.

This distinction between generating and creating is crucial. GenAI can generate content by synthesizing existing information and patterns. However, it often lacks the spark of true creativity. Generated things are rarely genuinely new; they are recombinations of what already exists. Creation, on the other hand, involves originality and innovation—elements that are currently more characteristic of human endeavor.

It remains open if quantitative progress, which can surely be expected from AI —after all, we keep pouring yottaflops, gigadollars, and terrawatts into the thing— will lead to a qualitative leap eventually. Then AI could actually be creative. We will see. If we are lucky. Again.

coding along LLMs, July

AI technology

July 2024 and it is still surprising how erratic LLMs can be when they get tasked to help with very small coding jobs. I work on many projects on my Mac. For a while, I have been using my own directory simple stack implementation. Remembering paths is what the computer can do for me.

It works well and has two parts: zsh functions loaded via .zshrc and a Python program doing the actual work, naturally except for the cd, change of the prompt.

I thought it would be nice if my ‘cd’ variation could check if a Python environment under */bin/activate would exist in the directory I changed into. If so, it can source it. If there is none, then it should not care, and if there are multiple, it should list them so that I could pick and choose.

Simple enough.

Parts would require zsh shell coding. Not something I tend to do a lot. Since Sonnet 3.5 has a limit even in the paid version, I tend to use my paid gpt4o first.

For this simple thing, I should not have. Today gpt4o was stunningly stupid. It managed to do zsh syntax well enough, but then completely failed. For a while, stuck in that dreadful loop where one hopes the next version would finally work. I still abort those loops of idiocy way too late.

Claude 3.5 got it right. In my frustration, I also had introduced a bug / typo on my end. Both gpt4o and Claude would have pointed it out easily if they had seen that part of the code. Claude stood out since its debug hints let me see what I had done wrong. That was beyond my current expectation horizon.

Speaking of: I am amazed how dumb LLMs still can be. Today gpt4o was utterly stupid. Not sure why. Is it zsh that it is not familiar with? Did the system prompt that got assigned to me, or my region, suddenly change? Who knows.

It must be hard to make a living based on some expectations from LLMs. They are really awesome, but can fall off a cliff at any point. Pretty much the opposite from computer work in general.

I expect that people develop all sorts of Cargo Cults in their work with these tools.

Meta Status

AI economy history internet technology

What does Meta do? It turns people into money. Those that are on the Internet, that is—not in a Soylent Green kind of way.

At least, that was the mantra up until 2018. Then Cambridge Analytica broke. And the Q2 2018 earnings gave an inkling of the possibility that not a fixed—and also rather large—ratio of people entering the Internet would become, just like magic, Facebook users.

Later, people seemed to forget about the fact that they get algorithmically nudged in Zuck’s wonderland every step of the way. Wall Street itself realized that revenues at 1 Hacker Way actually kept on rising—until they jumped in 2021. COVID, remember?

The Metaverse, however, wasn’t really that great of a hit, and after the virus bonus revenue fell back in line the following year, FB lost a staggering two-thirds of its value. A trillion-dollar meme stock.

An attribute that it then turned into current heights via hitching itself to the AI bandwagon.

Releasing the LLaMA weights is undoubtedly a commendable move. It sounds utterly impressive when you can claim, “While we’re working on today’s products and models, we’re also working on the research we need to advance for LLaMA 5, 6, and 7 in the coming years and beyond to develop full general intelligence,” in an earnings call. Pretty much like that strange man proclaimed five years ago: “I want 5G, and even 6G, technology in the United States as soon as possible.” Numbers: They go up, up, and up.

Hype aside, I am not really aware of any practical applications for LLaMA 3. Zuck bought lots of GPUs. Both Jensen and I are happy about that. Maybe they thought they had all this data that people have entered in their apps. Maybe they could train a LLM on it. With GPT-3, there was this notion that the size of the training corpus was all that mattered. After all, OpenAI’s chatbot was such a wonder, and it jumped into existence just via the increase of its training data. I speculate that a trillion training tokens derived from FB discussions yield surprisingly little meaningful reasoning power. Especially compared to actual content like, for instance, Wikipedia.

The pressure to come up with something must have weighed heavily on 1 Hacker Way. As those two transformer-based applications (LLMs and Image Diffusers) broke into public view and kicked the world into a frenzy that seemingly became the new normal, Meta itself had just spent around $50 billion on developing, well, the Metaverse. Which received rather little positive reaction, to put it mildly.

The total and utter failure of Zuck’s idea to come up with a whole new thing left Meta with no choice but to jump on the AI hype PR scheme. And up to this day, it has worked rather well. While revenue is ticking along as expected, the stock is kissing new heights. For now.

So, what’s next?
Nobody knows.

What will happen is that Internet population growth will end. There are simply no more people left that could join. Pretty much everybody who could go online already has done so. While 25% of the world’s population are younger than 15, many of them live in underdeveloped parts of Africa. Furthermore, young people hardly flock eagerly into the Meta family of products once they get their first Internet device.

Meta’s revenue growth would therefore stall together with the plateau in its user count. While they continue to make a lot of money, a PE ratio of currently around 30 is expecting something else: More money. You need to grow profits to justify such a valuation.
A quick way to bump revenues would be to reduce costs. Twitter is still up and running, despite Mr. Musk letting go of most of its workforce. A tempting move that could save the numbers for a quarter or two at Menlo Park as well. The problem is that this approach works only briefly: Costs go down to zero. But not more.

Which means that Meta needs to increase revenues while user numbers can no longer grow.

Can Zuck’s companies accomplish that? They might, but it would not be pretty: Billions of people have delegated a great part of their social existence into the “Meta Family of Products”. (What’s in a name?) A sticky situation in itself. Add to that the addictive aspects that rival nicotine, and you realize that half the planet as a user base won’t go anywhere fast.

Wealth as well as the inflexibility to change app use or social topology both tend to grow with increasing age. Meta owns people’s time and attention in staggering amounts.


Here comes the part that isn’t pretty: it is rather easy to manipulate people online. Tech is able to do it. And will increasingly be. There is a threshold after which you no longer realize that you got nudged.

When the magician manages to direct your attention successfully, all sorts of things are possible. With a serious difference: Magic lives from the effect, that the outcome shows you, that you must have missed something. You are supposed to notice that it is impossible what just happened.

Manipulation to gain, aka advertisement, has a different aim: You should be made to act in certain ways, all the while thinking that you want to do that.

The total spending of Meta family users is responsible for a mind-boggling share of GDP. And, as discussed, most of the users will not go anywhere. If Meta does not f*ck up royally, pretty much half of the global adults will continue to point their noses, eyes, minds, and wallets its way.

Turning on the manipulation engine will not be one deliberate conscious act or one magnificent large piece of software. Lots of little changes will yield lots of little benefits. With billions of people, you can do a whole lot of A/B testing. Nobody will notice. Everybody’s feed is different and the fact that you see wording that is ever so slightly different will not trigger any of the societal mechanisms that will raise a reaction.

Jacob Riis used flash photography at the end of the 19th century to show the world how poor people lived in NYC, and he changed the world for the better. I cannot imagine how we can illuminate the modern plight of getting nudged into an ultimately unhappy existence that looms on the horizon.

Navigating LLMs: Benefits and Drawbacks in June 2024

AI google history internet

LLMs have their limits, and where they excel makes a difference. As of June 2024, they continue to evolve. Anthrop\c Claude 3.5 works well for coding simple things with Python. It feels like the LLM has been heavily trained on existing code. Actually, it might be just as good in other applications as well. I wouldn’t know since I only use it for coding right now. Even on the paid plan, it has a message limit, which feels very 2023. So, I use the limited interaction where I get the highest benefit, which is coding. The artifact window is a great idea, and the speed of generation is appreciated. With gpt4o, I had to interleave work: make a request, switch to a different task while gpt4o sputtered out characters at Morse code speed. It probably runs on a colony of squids at the bottom of the Mariana Trench that OpenAI taught how to use Morse code with each arm.

And yes, an image like this I create with gpt4o. I don’t even know if Claude can do that. I don’t mind having multiple LLMs. I am gladly paying for both of them, as I do for search. Right now, I am very happy that there is more than one solution. I tried to use Google AI, but it was too complicated to figure out. To find the offering that fit mit my needs. And I am not aware of a key feature that I could only do with them. They already have all my email, read the entire Internet. If I can avoid it, I would not like to help them any further. Sure, if they were as good at coding as Claude, I would use them in a second. I have morals, but I cannot save the world single-handedly either.

One of the bigger fears I have is that LLMs might take the same turn that Google Search did. It was a great idea. It worked great, allowing for a phase of the Internet in the early 2000s that was very promising. Then it became what we suffer from today—a swamp. Barely functional. Generating around $150 profit for Google per user annually. Which means companies make more. Which means that I loose even more than that. The costs of using Google Search by being manipulated are much higher now than its benefits. The SEO world that Google Search presents is not a nice one. I happily give Kagi money to have some distance from that swamp.