AI considered not so harmful

Cal Newport, hoogleraar informatica, schrijver en podcaster, ontkracht de hysterische reacties op de nieuwste ontwikkelingen op het gebied van AI. Deze hysterie is grotendeels te wijten aan de aandachtzoekende houding van de media, en niet zozeer aan onderzoek dat met wetenschappelijke nauwkeurigheid is uitgevoerd. “We have summoned an alien intelligence,” schrijft Harari, die langzaam maar zeker verandert in een luddite en technologiepessimist.

Cal Newport doet wat Harari en anderen hadden moeten doen. In zijn Deep Questions-podcast ‘Defusing AI panic’ ontleedt hij het onderwerp.

Only by taking the time to investigate how this technology actually works—from its high-level concepts down to its basic digital wiring—can we understand what we’re dealing with.

Cal Newport legt uit wat ChatGPT werkelijk doet en hoe intelligent het is. We zullen zien dat dat nogal beperkt is.

The result of these efforts might very well be jaw-dropping in its nuance and accuracy, but behind the scenes, its generation lacks majesty. The system’s brilliance turns out to be the result less of a ghost in the machine than of the relentless churning of endless multiplications.

A system like ChatGPT doesn’t create, it imitates.

Consciousness depends on a brain’s ability to maintain a constantly updated conception of itself as a distinct entity interacting with a model of the external world. The layers of neural networks that make up systems like ChatGPT, however, are static…

It’s hard to predict exactly how these large language models will end up integrated into our lives going forward, but we can be assured that they’re incapable of hatching diabolical plans, and are unlikely to undermine our economy.

In de podcast gaat Cal Newport wat dieper in op de technische details. Uit het transcript (door mij licht bewerkt wat betreft de interpunctie):

What a large language model does is it takes an input. This information moves forward through layers. It’s fully feed forward and out of the other end comes a token which is a part of a word in reality. It’s a probability distribution over tokens but whatever a part of a word comes out the other end that’s all a language model can do. Now, how it generates what token to spit out next can have a huge amount of sophistication …

When I talk to people is when you begin to combine this really really sophisticated word generator with control layers. Something that sits outside of and works with the language model that’s really where everything interesting happens . Okay this is what I want to better understand: the control logic that we place outside of the language models that we get a better understanding of the possible capabilities of artificial intelligence because it’s the combined system language model plus control logic that becomes more interesting. Because what can control logic do?

It can do two things: it chooses what to activate the model with, what input to give it and it can then second: actuate in the real world or the world based on what the model says. So it’s the control logic that can put input into the model and then take the output of the model and actuate that, like take action, do something on the Internet, move a physical thing.”

Something I’ve been doing recently is sort of thinking about the evolution of control logic that can be appended to generative AI systems like large language models…

Als je naar de afbeelding kijkt die ik na de podcast van Cal Newport heb gemaakt, zie je de verschillende besturingslagen. Zoals Cal Newport aangeeft, wordt daar het echte werk verricht. De LLM is statisch; hij geeft een woord, en dat is alles. Die besturingslogica weet wat er met het werk moet gebeuren.

Control layer in contemporary artificial intelligence

De besturingslogica is nu complexer geworden. We weten beter wat we moeten doen met de antwoorden die AI ons geeft.

Newport fantaseert over een derde besturingslaag die met verschillende AI-modellen kan communiceren, de intentie kan bijhouden, visuele herkenning heeft en complexe logica kan uitvoeren. Dat is waar we in de buurt komen van Algemene Kunstmatige Intelligentie.

Maar, zoals Newport opmerkt, is er niemand die hiermee bezig is.

Minstens even belangrijk is dat deze besturingslogica volledig door mensen wordt geprogrammeerd. We zijn nog lang niet zover dat we te maken hebben met door AI gegenereerde besturingslogica en zelflerende besturingslogica. Wat Newport ‘intentional AI’ (iAI) noemt. Het is onduidelijk of dit met onze huidige AI-technologie wel mogelijk is.

Het is juist in de besturingslogica waar de spannende dingen gebeuren.

Het zijn nog steeds mensen die de besturingslogica verzorgen.

In 1990 studeerde een vriend van mij af in fuzzy logic. Die periode was waarschijnlijk het hoogtepunt van de hype rond fuzzy logic. Fuzzy logic was een van de technologieën die de samenleving op zijn kop zouden zetten. Tegenwoordig is fuzzy logic gewoon een technologie die, net als andere, wordt toegepast voor het juiste doel en binnen de juiste probleemstelling.

Wat vandaag de dag op sciencefiction lijkt, is de gangbare technologie van morgen. De AI van vandaag is het loodgieterswerk van morgen. Dat is mijn interpretatie van Cal Newports’ uitleg over de huidige stand van zaken op het gebied van AI.

Update 11 juni 2026

In een recente podcast van Cal Newport neemt hij opnieuw de apocalyptische en overdreven technofobische uitspraken van historicus Yuval Noah Harari over AI onder de loep en komt hij opnieuw tot de slotsom dat deze overdreven en technisch inaccuraat zijn. Harari stelde tijdens een speech in Davos dat AI een zelflerende ‘agent’ is die kan liegen en manipuleren om te overleven. Newport legt uit dat dit een verkeerd beeld geeft van de huidige computerwetenschap.

Volgens Cal zijn huidige AI-systemen (zoals LLM’s) in essentie statische “verhalenmachines”. Ze veranderen of groeien niet uit zichzelf. LLM’s voorspellen simpelweg het volgende logische woord op basis van hun training. Als een chatbot ‘liegt’ of zich gedraagt als een kwaadaardige AI, komt dat puur doordat de gebruiker een prompt geeft die hints bevat naar sciencefiction-achtige scenario’s. De AI maakt die verhalen vervolgens alleen maar geloofwaardig af.

Daarnaast falen AI-agents (de programma’s die plannen uitvoeren) volgens Newport nog steeds massaal, afgezien van specifieke programmeertoepassingen. LLM’s kunnen simpelweg niet logisch plannen.

Newport neemt het Harari niet persoonlijk kwalijk. Harari is een historicus die reageert op de overgehypte verhalen die Silicon Valley de wereld in slingert. Harari reflecteert dus prima op het verhaal dat hem verteld is, maar dat onderliggende verhaal klopt volgens Newport simpelweg niet met de technische realiteit.

Ik vind zelf echter dat Harari wel degelijk laakbaar is. Hij presenteert zich als AI-kenner, schrijft er in zijn boeken over en geeft er in veel interviews zijn mening over. Hij zou zich beter moeten laten voorlichten en niet moeten meesurfen op de hypes die vanuit Silicon Valley worden gecreëerd om de enorme investeringen in AI goed te praten, of zelfs om meer geld op te kunnen halen.

Opting out of Instagram AI

As European users, we can opt out of Instagram and Facebook using our posts for AI training. I’ve exercised this control, as I am the product of Facebook and Instagram, but I strive to limit their use of me as such.

Opting out on Instagram looks deliberately cumbersome. However, from Facebook, which is also owned by Meta, I received an email with very simple instructions.

Now, I am curious if they can prove they are not using my data for AI.

The cost of AI and other challenges

I stumbled upon this fascinating article by Stuart Mills looking at the challenges that further development and operations of AI models face.

The costs of model development and operation are increasing. Efficiencies in development and operation are challenging but may be addressed in the future. However, model quality remains a significant challenge that is more difficult to solve.

Data is running out. Solutions such as synthetic data also have their limitations.

There is also a severe challenge around chips. There is a supply shortage in the context of geopolitical tensions between China, the US, and the EU. Also, the environmental costs of running large AI models are significant.

The costs of model development and operation are increasing. Efficiencies in development and operation are challenging but may be addressed in the future. However, model quality remains a significant challenge that is more difficult to solve.

Data is running out. Solutions such as synthetic data also have their limitations.

There is also a severe challenge around chips. There is a supply shortage in the context of geopolitical tensions between China, the US, and the EU. Also, the environmental costs of running large AI models are significant.

Two revenue models may emerge in the AI industry, each with its own take on the cost aspects highlighted above. The first is the ‘foundation model as a platform’ (OpenAI, Microsoft, Google), which demands increasing generality and functionality of foundation models.

The second is the ‘bespoke model’ (IBM), which focuses on developing specific models for corporate clients.

Government action can support and undermine the AI industry. Investment in semiconductor manufacturing in the US and China may increase the supply of chips, and strategic passivity from governments around regulations such as copyrights is suitable for the industry. Government interventions should regulate the AI industry in areas related to the socially and environmentally damaging effects of data centers, copyright infringement, exploitation of laborers, discriminatory practices, and market competition.

AI, duh; make it personal (and analog)

ai photographer
The competition

When Artificial Intelligence-generated images win photo contests, should we oppose that?

I just think the developments of AI are telling us to do things differently, to stand out. AI has become the competition (and maybe just a tool), just like all other photographers are. So, we have to treat AI as competition, too. You can try to deny this reality, but you can also look at how you, as a photographer or artist, can differentiate yourself from this new collegue/competition.

Ideas:

  • Stories instead of single images. Combine with text.
  • An analog version of your work: a print, a book, wallpaper, toilet paper, t-shirts, quilt covers, printed bags, whatever.
  • Combine your photos into a video.
  • Handmade books.
  • Collages.

Personal and analog distinguish you from the aggregated, statistically generated products of AI.

AI credibility

Some thoughts regarding the current Artificial Intelligence hype.

Why We Need a Realistic View of AI

I am not among the people that believe AI technology is going to make humans superfluous. Like any technology, AI technology can be a blessing and a curse. AI technology does have its dangerous sides, and we should focus our attention on these aspects, and try not to get lost in religious debates. I think we should look at:

  • Transparency on the (quality of the) inputs for the AI-based software solutions.
  • Transparency on the (energy) cost of AI solutions.
  • Clarity about the possibilities and more importantly, the limitations of AI solutions.
  • Regulations on ownership of important data sets.

The quality of AI outputs

The most fundamental threat from AI technologies comes from its inputs. The quality of the output of the AI is directly dependent on the quality of the data fed into the AI’s learning process. If you cannot control the quality of the data you use to train the AI model, you cannot control the quality of the output either. (Most research in this respect is done in the area of so-called biases. What I mean here is more generic: how to assure the reliability of any data feed into an AI.)

Groene Amsterdammer 3 - 2023 - artwork niek de greef

The role of AI on social media

Any benevolent social medium platform that exploits AI technology has a great challenge: how do you curate the social media fed into the platform? How to distinguish real trends from fake ones. How do you prevent bots from people or organizations with malicious intentions to enter unreliable information into the stream, thereby using the social media platform to spread incorrect information, or ‘fake news’ as it become known by since Donald Trump.

Governments are debating how to regulate these unreliable and often uncontrollable social media streams. The first and probably the easiest thing to demand from social media companies is to apply journalistic principles: transparency over the sources of information. I think this is the most reliable and easiest measure that can be taken. Examining the algorithms, as is suggested sometimes, is very difficult, for the provider as well as the examiner. Instead of looking inside the AI box, we can judge the input to the AI, and make significant judgments on the reliability of the AI’s outputs.

AI and humanity

AI technology de-humanizes. When AI takes over more of our jobs and activities, what is left over to humans? What if AIs at some point have gained superior thinking power over humans? What is the purpose of men then? Humans become subordinate to AI intelligence. We lose the one thing that has always distinguished us from the rest of the living creatures on this planet: our superior thinking abilities. I do not think that the current AI technology is even close to taking over our thinking abilities. Later in this article, I will argue why I think so. Yet, in the sense that AI technology is the next step in taking over tasks that so far only humans could do, yes, AI takes some of that ‘humanity’ away. Jobs will be lost. But that is not a very revolutionary development. The calculator took over calculations, the early computer took over bookkeeping, and the later computer replaced our paper agendas, assembly lines, and our cars.

Energy consumption

Our brains need a lot of energy: our 2% brain is responsible for 20% of the energy consumption of our body. AI requires a lot of energy as well. In January 2023, ChatGPT used as much energy as a mid-size town. A single question requires Apple’s Siri about 5 Wh of energy. A household uses something like 5000 kWh (in Europe, in the US it is double that figure). So, 200 questions to Siri is 1kWh. Then 1 million questions to Siri needs as much energy as a household in Europe in a year. That’s a lot.

If I ask Siri 3 questions per day, that is about 1000 questions per year. That is 5 kWh.

Does your Siri also tend to spontaneously answer questions you did not ask? 5W each.

Who benefits from the AIs?

Currently, large organizations drive the development of AI technology. Microsoft, Google, and IBM, all promise to make the world a better place, but skepticism is justified. Not only history has proven these companies strive to dominate the marketplace. According to economic theories, this is totally in line with expectations. And this is not the thing to be afraid of. Monopolies never last at are not scalable. At some point, monopolists can sell a greater quantity only by cutting its price. Microsoft came down, after years of trying to dominate the market with inferior Windows technology. (Yes Microsoft Office still dominates the text processing market, though I am not sure why that is the case, as very viable free products are available for quite some time now). IBM came down. Our own East-India Company came down. Google will come down. Facebook will come down.

What is problematic, however, is when technologies like AI end up solely in the hands of a single or a few parties and a power balance is disrupted. These parties can dominate others with these technologies. Totalitarian ruling could become a big problem. A relativizing thought, however… AI technologies are much more difficult, if not impossible, to contain then technologies nuclear weapons, for example.
You only need moderate capital for AI solutions. Open-source AI technologies are available for everyone. The problem is not the technology, the problem would be when a single party could claim exclusive access to data sources.
GPS data for example.

Why AI is miles away from human intelligence

The current AI technology is an analytical, stochastic process, invented in the last two decades of the 20th century. The process is very compute-intensive. Only recently our computers have become strong enough so we can make practical use of AI inventions.

The AI technology analyses past data, and is very clever at “predicting” future values for this data, or gathering ‘facts’ from data. The technology looks at input data and separates noise from signal by throwing out outliers and assigning likelihood values to remaining data points. With this, it can predict (not even calculate) probabilities for future values. That is all.

Nicholas Nassim Taleb has written a book called Fooled by Randomness. One of the points he proves in this book is that you can not predict “black swan” events from past data. The existence of black swan events can be proven through analytical math, but not predicted through stochastical math.
So AI is worthless when it comes to predicting black swan events. More specifically AI then useless for supposed applications like stock market predictions. Even professors in AI research areas are ignorant of these “black swan” limitations, arguing that just further growth of datasets would enable AI to become perfect predictors of the future (and become uber-humanly intelligent).

Out of curiosity, I asked ChatGPT about its ability to predict black swan events.
Here is his honest, modest answer.

AI systems, including those based on advanced algorithms like GPT-3.5, have limitations when it comes to predicting black swan events. Black swan events are extremely rare, unpredictable, and have a significant impact. These events are, by definition, unforeseen and often fall outside the range of normal expectations.

AI models, including machine learning algorithms, are generally trained on historical data. If black swan events have never occurred in the training set, the AI model would not have learned to predict them explicitly. Traditional AI models work based on patterns and correlations found in the data they were trained on. They are not capable of inventing entirely new and unprecedented events beyond the scope of their training data.

However, it’s essential to note that AI can help prepare for unexpected events, even if it cannot predict black swan events directly. Here’s how:

Despite these capabilities, it’s important to maintain a cautious approach when relying on AI for preparing for black swan events. Human judgment, expert analysis, and continuous monitoring are essential elements in mitigating the impacts of truly unforeseen and exceptional events.

Next, I will argue that the learning process of AI is so inherently different than human learning, that it would be naïve to expect that such processes would ever be able to teach an AI so much that they can mimic human intelligence. AIs could mimic aspects of human intelligence, but for development of more human-like intelligence, a different learning model would be required.

When we look at the AI learning process, machine learning, we see this is significantly different from human learning and experience. The difference between human learning and experience and Machine Learning in short. The AI takes larger amounts of information and learns from it. Sometimes this learning is (self-)directed. Humans take in large amounts of data, information, and experience. This human experience includes sensory information. Humans undergo years of learning from interactions with people. They get complex and ambiguous feedback. They learn from experiences. AI’s do not undergo experiences. At best experiences could be simulated. The learning processes for humans and AIs are so different from Machine Learning that is silly to expect that the two could ever produce similar results.

The AI gets:

  • Many data sets, in huge volumes, and these are largely uncurated.
  • Large undirected learning.

Man gets:

  • A curated (culturally, geographically, …) stream of information streams books, tv, anything internet,
    school, discussion, talks, magazines, Youtube, art, music …

  • Music,… What is music to an AI? Music is the thing that plugs right into our brains…
  • A in-born character (depending on where you are in the nature/nurture dimension).
  • Years of growing up, interactions with people, animals, things, …
  • Years of sensory information from 5 sources.
  • Emotional feedback from humans, animals…

I do not think there is much reason to be freightened of AI. The current state of AI is very very far from human intelligence. And the current technology misses fundamental functions to close that gap. We should not be afraid of that. Rather than getting distracted by pessimistic stories warning of the end of humankind, we should focus our efforts on the dangerous aspects of AI.

AI is just a tool not a miracle, and its health depends on what you feed it

AI-based social media enables the spreading of incorrect and potentially dangerous information. Other AI based applications struggle with curating the information fed into the tools. Without stringent curation, the AI application undermines its own credibility, as do the providers of services based on unreliable services. Which benevolent being wants to believe a social media channel that bases its information on unreliable data?
Who wants to rely on AI-based decisions if the decisions are learned from an unreliable data set?

Read more about Cal Newport’s analysis of AI.

Check out my tech projects and articles.