Michiel Bontenbal
@mpbontenbal.bsky.social
Complex beings as us humans can not be summarised in a few lines but I am here for #AI #climate #EUtech #EU_politics. Lecturer in AI & IT. Some posts in Dutch. 🇪🇺
We need to talk about the 'sparkles button' that is becoming popular.
In my view, this is deeply problematic. Sparkles are associated with magic.
But AI is anything but magic. AI systems are highly sophisticated systems based on math.
The association with magic is simply misleading.
#learnAI
In my view, this is deeply problematic. Sparkles are associated with magic.
But AI is anything but magic. AI systems are highly sophisticated systems based on math.
The association with magic is simply misleading.
#learnAI
October 9, 2025 at 8:18 AM
We need to talk about the 'sparkles button' that is becoming popular.
In my view, this is deeply problematic. Sparkles are associated with magic.
But AI is anything but magic. AI systems are highly sophisticated systems based on math.
The association with magic is simply misleading.
#learnAI
In my view, this is deeply problematic. Sparkles are associated with magic.
But AI is anything but magic. AI systems are highly sophisticated systems based on math.
The association with magic is simply misleading.
#learnAI
Every week there are new models, but to get further in AI as a community we need open source models, that also disclose the data they are trained on.
#learnAI
#learnAI
October 2, 2025 at 7:42 AM
Every week there are new models, but to get further in AI as a community we need open source models, that also disclose the data they are trained on.
#learnAI
#learnAI
Thanks for the reply. I was referring to this well known diagram. I still think prompt engineering is more like classical programming than Machine Learning.
September 24, 2025 at 5:45 PM
Thanks for the reply. I was referring to this well known diagram. I still think prompt engineering is more like classical programming than Machine Learning.
I asked several AI's to calculate Fibonacci numbers in Python. Humans would do this in ± 10 lines of code.
A rule of thumb could be that the larger the model, the larger the code it produces. And 'agentic mode' is mostly bullshit.
In summary run local, small models to get high quality code!
A rule of thumb could be that the larger the model, the larger the code it produces. And 'agentic mode' is mostly bullshit.
In summary run local, small models to get high quality code!
September 16, 2025 at 2:05 PM
I asked several AI's to calculate Fibonacci numbers in Python. Humans would do this in ± 10 lines of code.
A rule of thumb could be that the larger the model, the larger the code it produces. And 'agentic mode' is mostly bullshit.
In summary run local, small models to get high quality code!
A rule of thumb could be that the larger the model, the larger the code it produces. And 'agentic mode' is mostly bullshit.
In summary run local, small models to get high quality code!
My institution recommends JetBrains PyCharm as the preferred IDE.
I have some interesting (but frustrating) chats with is AI code assistant...
#learnAI
I have some interesting (but frustrating) chats with is AI code assistant...
#learnAI
September 5, 2025 at 8:09 AM
My institution recommends JetBrains PyCharm as the preferred IDE.
I have some interesting (but frustrating) chats with is AI code assistant...
#learnAI
I have some interesting (but frustrating) chats with is AI code assistant...
#learnAI
Hey @ec.europa.eu you should not let your AI generate that alt-text as it is merely bullshit!
That's not Donald Tusk in the image.
That's not Donald Tusk in the image.
September 1, 2025 at 2:29 PM
Hey @ec.europa.eu you should not let your AI generate that alt-text as it is merely bullshit!
That's not Donald Tusk in the image.
That's not Donald Tusk in the image.
September 1, 2025 at 12:20 PM
I tried GPT-5 mini the other day in VS code and I broke it...
Now unable to use any code generator for that specific repo...
(but for some reason copilot still works in other repo's though...)
Now unable to use any code generator for that specific repo...
(but for some reason copilot still works in other repo's though...)
August 26, 2025 at 8:10 AM
I tried GPT-5 mini the other day in VS code and I broke it...
Now unable to use any code generator for that specific repo...
(but for some reason copilot still works in other repo's though...)
Now unable to use any code generator for that specific repo...
(but for some reason copilot still works in other repo's though...)
August 7, 2025 at 6:54 PM
Maybe the latest generation of models can come to the rescue. Here is a Vision Language Model (ChatGPT) that explains it rather well:
July 10, 2025 at 10:15 AM
Maybe the latest generation of models can come to the rescue. Here is a Vision Language Model (ChatGPT) that explains it rather well:
Typical mistake made by a supervised learning algorithm like CNN or YOLO trained on a limited number of images. The model learns the context -the position of the arm next to the head - rather than recognizing the object itself.
nos.nl/artikel/2574...
nos.nl/artikel/2574...
July 10, 2025 at 10:11 AM
Typical mistake made by a supervised learning algorithm like CNN or YOLO trained on a limited number of images. The model learns the context -the position of the arm next to the head - rather than recognizing the object itself.
nos.nl/artikel/2574...
nos.nl/artikel/2574...
I asked #Gemini to summarise a paper for me, but then it summarised the WRONG paper, not the one I asked for.
#weird #learnAI #hallucination
#weird #learnAI #hallucination
July 2, 2025 at 10:22 AM
I asked #Gemini to summarise a paper for me, but then it summarised the WRONG paper, not the one I asked for.
#weird #learnAI #hallucination
#weird #learnAI #hallucination
Claude gave it's verdict.
June 14, 2025 at 10:37 AM
Claude gave it's verdict.
Re: Identifying deepfakes with Vision Language Models.
I created some code to generate this gif. The code takes the image of Zelensky and gives arguments why it's a #deepfake. It uses Claude 4 to generate the 5 arguments. Claude correctly predicted this is a deep fake.
#AI-pocalypse #learnAI #VLM
I created some code to generate this gif. The code takes the image of Zelensky and gives arguments why it's a #deepfake. It uses Claude 4 to generate the 5 arguments. Claude correctly predicted this is a deep fake.
#AI-pocalypse #learnAI #VLM
June 8, 2025 at 11:43 AM
Re: Identifying deepfakes with Vision Language Models.
I created some code to generate this gif. The code takes the image of Zelensky and gives arguments why it's a #deepfake. It uses Claude 4 to generate the 5 arguments. Claude correctly predicted this is a deep fake.
#AI-pocalypse #learnAI #VLM
I created some code to generate this gif. The code takes the image of Zelensky and gives arguments why it's a #deepfake. It uses Claude 4 to generate the 5 arguments. Claude correctly predicted this is a deep fake.
#AI-pocalypse #learnAI #VLM
Re: Identifying deepfakes with Vision Language Models.
I created some code to generate this gif. The code takes the image of Zelensky and gives arguments why it's a #deepfake. It uses Claude 4 to generate the 5 arguments. Claude correctly predicted this is a deep fake.
#AI-pocalypse #learnAI #VLM
I created some code to generate this gif. The code takes the image of Zelensky and gives arguments why it's a #deepfake. It uses Claude 4 to generate the 5 arguments. Claude correctly predicted this is a deep fake.
#AI-pocalypse #learnAI #VLM
June 8, 2025 at 11:30 AM
Re: Identifying deepfakes with Vision Language Models.
I created some code to generate this gif. The code takes the image of Zelensky and gives arguments why it's a #deepfake. It uses Claude 4 to generate the 5 arguments. Claude correctly predicted this is a deep fake.
#AI-pocalypse #learnAI #VLM
I created some code to generate this gif. The code takes the image of Zelensky and gives arguments why it's a #deepfake. It uses Claude 4 to generate the 5 arguments. Claude correctly predicted this is a deep fake.
#AI-pocalypse #learnAI #VLM
See for yourself the quality of the reasoning here. This is with Claude. (Also note that the AI model does not describe that this is Zelensky and why he is holding this ace with a drone)
June 5, 2025 at 7:53 AM
See for yourself the quality of the reasoning here. This is with Claude. (Also note that the AI model does not describe that this is Zelensky and why he is holding this ace with a drone)
Europe gets it's gas from 3 main sources: Russia, USA and Qatar.
We don't want Putin's gas any longer, that leaves us with these two. We will be blackmailed with gas, sooner rather than later.
I am happy that I have installed a heat pump this week.
We don't want Putin's gas any longer, that leaves us with these two. We will be blackmailed with gas, sooner rather than later.
I am happy that I have installed a heat pump this week.
May 11, 2025 at 5:43 PM
Europe gets it's gas from 3 main sources: Russia, USA and Qatar.
We don't want Putin's gas any longer, that leaves us with these two. We will be blackmailed with gas, sooner rather than later.
I am happy that I have installed a heat pump this week.
We don't want Putin's gas any longer, that leaves us with these two. We will be blackmailed with gas, sooner rather than later.
I am happy that I have installed a heat pump this week.
Shared this slide with my students yesterday. If you are new to AI coding, you can work on both the prompt and the response to get better code from your AI code generator.
#learnAI #codingwithAI
#learnAI #codingwithAI
May 9, 2025 at 8:36 AM
Shared this slide with my students yesterday. If you are new to AI coding, you can work on both the prompt and the response to get better code from your AI code generator.
#learnAI #codingwithAI
#learnAI #codingwithAI
Yep. Amerikanen hebben door hoe prettig het leven is in Nederland. Amsterdam is al jaren populair bij Amerikanen, in 2023 waren zij de een na grootste groep immigranten.
Bron: onderzoek.amsterdam.nl/artikel/amst...
Bron: onderzoek.amsterdam.nl/artikel/amst...
May 5, 2025 at 5:03 PM
Yep. Amerikanen hebben door hoe prettig het leven is in Nederland. Amsterdam is al jaren populair bij Amerikanen, in 2023 waren zij de een na grootste groep immigranten.
Bron: onderzoek.amsterdam.nl/artikel/amst...
Bron: onderzoek.amsterdam.nl/artikel/amst...
One of the largest groups of migrants to Amsterdam are... Americans. In 2023 Americans were the second largest group of migrants to my city.
source (Dutch): onderzoek.amsterdam.nl/artikel/amst...
source (Dutch): onderzoek.amsterdam.nl/artikel/amst...
April 19, 2025 at 5:55 PM
One of the largest groups of migrants to Amsterdam are... Americans. In 2023 Americans were the second largest group of migrants to my city.
source (Dutch): onderzoek.amsterdam.nl/artikel/amst...
source (Dutch): onderzoek.amsterdam.nl/artikel/amst...
Can you ask questions about an image?
A 'Vision Language Model' links your question with the visual content of an image. It will generate a full response to your question.
#learnAI #VisionLanguageModel
A 'Vision Language Model' links your question with the visual content of an image. It will generate a full response to your question.
#learnAI #VisionLanguageModel
April 17, 2025 at 11:38 AM
Can you ask questions about an image?
A 'Vision Language Model' links your question with the visual content of an image. It will generate a full response to your question.
#learnAI #VisionLanguageModel
A 'Vision Language Model' links your question with the visual content of an image. It will generate a full response to your question.
#learnAI #VisionLanguageModel
For many years a lot has been said about a slow EU economy, outpaced by the US. But now it reminds me of the fable of the Tortoise and the Hare. Curious how you see that.
April 16, 2025 at 1:13 PM
For many years a lot has been said about a slow EU economy, outpaced by the US. But now it reminds me of the fable of the Tortoise and the Hare. Curious how you see that.
AI, and more specifically Large Language Models (LLMs), are becoming a key component in web applications, enabling new and advanced functionalities.
With the rise of local AI models, you do not run an API but it the app can run 100% in the browser or on device.
#webapplications #learnAI
With the rise of local AI models, you do not run an API but it the app can run 100% in the browser or on device.
#webapplications #learnAI
April 14, 2025 at 1:06 PM
AI, and more specifically Large Language Models (LLMs), are becoming a key component in web applications, enabling new and advanced functionalities.
With the rise of local AI models, you do not run an API but it the app can run 100% in the browser or on device.
#webapplications #learnAI
With the rise of local AI models, you do not run an API but it the app can run 100% in the browser or on device.
#webapplications #learnAI
