Max Kannen
@maxkannen.bsky.social
Machine learning research assistant and currently doing my masters in CS. Futurist and interested in all STEM fields. I train neural nets since 2016.
I want to recommend some books that I read this year, like I did last year. Without particular order:
December 20, 2025 at 12:04 PM
I want to recommend some books that I read this year, like I did last year. Without particular order:
The same way people go to the Gym despite there not being much physical work anymore to look good and be healthy, in the future people will learn or go to schools to appear smarter and stay mentally fit even though there will be less intellectual work.
October 17, 2025 at 9:12 PM
The same way people go to the Gym despite there not being much physical work anymore to look good and be healthy, in the future people will learn or go to schools to appear smarter and stay mentally fit even though there will be less intellectual work.
I wrote my longest blog post yet. It is about how AI became a religion and the resulting consequences.
My next blog post will be my yearly prediction post.
mkannen.tech/ai-and-the-n...
My next blog post will be my yearly prediction post.
mkannen.tech/ai-and-the-n...
AI and the New Faith: How the Singularity Became a Modern Religion - Maximilian Kannen
Artificial intelligence is advancing at an ever-growing pace, reshaping industries and capturing the world’s attention. While often framed as an economic and technological revolution, there’s a deeper...
mkannen.tech
September 28, 2025 at 3:16 PM
I wrote my longest blog post yet. It is about how AI became a religion and the resulting consequences.
My next blog post will be my yearly prediction post.
mkannen.tech/ai-and-the-n...
My next blog post will be my yearly prediction post.
mkannen.tech/ai-and-the-n...
So why would anyone use OpenAIs API?
Especially now that you can use their models for much cheaper elsewhere. Look at this contrast. Both OpenAI models. One is a new reasoning model at over 2000t/s for pennies and the other one is their mini model for twice the price.
Especially now that you can use their models for much cheaper elsewhere. Look at this contrast. Both OpenAI models. One is a new reasoning model at over 2000t/s for pennies and the other one is their mini model for twice the price.
August 5, 2025 at 6:12 PM
So why would anyone use OpenAIs API?
Especially now that you can use their models for much cheaper elsewhere. Look at this contrast. Both OpenAI models. One is a new reasoning model at over 2000t/s for pennies and the other one is their mini model for twice the price.
Especially now that you can use their models for much cheaper elsewhere. Look at this contrast. Both OpenAI models. One is a new reasoning model at over 2000t/s for pennies and the other one is their mini model for twice the price.
Lets be honest. At this point it does not really matter what architecture you use. As long as it is big enough and you have enough data it will produce a decent LLM.
Someone training a pure MLP model with xxB parameters?
arxiv.org/abs/2506.14761
Someone training a pure MLP model with xxB parameters?
arxiv.org/abs/2506.14761
From Bytes to Ideas: Language Modeling with Autoregressive U-Nets
Tokenization imposes a fixed granularity on the input text, freezing how a language model operates on data and how far in the future it predicts. Byte Pair Encoding (BPE) and similar schemes split tex...
arxiv.org
June 25, 2025 at 7:04 PM
Lets be honest. At this point it does not really matter what architecture you use. As long as it is big enough and you have enough data it will produce a decent LLM.
Someone training a pure MLP model with xxB parameters?
arxiv.org/abs/2506.14761
Someone training a pure MLP model with xxB parameters?
arxiv.org/abs/2506.14761
It is surreal to see my name on Arxiv. I hope I can do it again soon.
arxiv.org/abs/2506.09804
arxiv.org/abs/2506.09804
Regularizing Learnable Feature Extraction for Automatic Speech Recognition
Neural front-ends are an appealing alternative to traditional, fixed feature extraction pipelines for automatic speech recognition (ASR) systems since they can be directly trained to fit the acoustic ...
arxiv.org
June 12, 2025 at 6:53 PM
It is surreal to see my name on Arxiv. I hope I can do it again soon.
arxiv.org/abs/2506.09804
arxiv.org/abs/2506.09804
Today @giffmana.ai gave a talk at my University. Really nice to attend some events like this since I will not be able to go to Interspeech this year.
June 4, 2025 at 10:15 AM
Today @giffmana.ai gave a talk at my University. Really nice to attend some events like this since I will not be able to go to Interspeech this year.
Two years ago I gave a presentation about the future of AI and technology where I talked about drone warfare and asymmetry. Now we see how this looks. Cheap drones taking out billions in material.
⚡️34% of Russia's strategic aviation (cruise missile carriers) worth about $7 billion was hit by today's Spiderweb attack - Security Service of #Ukraine.
June 2, 2025 at 6:24 PM
Two years ago I gave a presentation about the future of AI and technology where I talked about drone warfare and asymmetry. Now we see how this looks. Cheap drones taking out billions in material.
Did you ever see the comparisons of AI models by asking them to generate a Hexagon rotating with balls inside?
Let me explain why this is a terrible way to compare models:
Let me explain why this is a terrible way to compare models:
May 29, 2025 at 8:55 AM
Did you ever see the comparisons of AI models by asking them to generate a Hexagon rotating with balls inside?
Let me explain why this is a terrible way to compare models:
Let me explain why this is a terrible way to compare models:
LLMs can do stuff:
deepmind.google/discover/blo...
deepmind.google/discover/blo...
AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms
New AI agent evolves algorithms for math and practical applications in computing by combining the creativity of large language models with automated evaluators
deepmind.google
May 14, 2025 at 4:42 PM
LLMs can do stuff:
deepmind.google/discover/blo...
deepmind.google/discover/blo...
I have a new job and I feel so lucky to be able to join the research community in this decade. Feels like the most interesting time to be in computer science.
May 8, 2025 at 8:20 PM
I have a new job and I feel so lucky to be able to join the research community in this decade. Feels like the most interesting time to be in computer science.
I like the feeling when I get a new PC. Choosing OS, Programms, Folder structure, moving over files, testing the hardware. I used to have the same feeling with a new phone, but Apple cloud sync killed it. After everything is synced I forget that it is a new phone in minutes.
April 28, 2025 at 1:27 PM
I like the feeling when I get a new PC. Choosing OS, Programms, Folder structure, moving over files, testing the hardware. I used to have the same feeling with a new phone, but Apple cloud sync killed it. After everything is synced I forget that it is a new phone in minutes.
GPT 4.1 family is announced and is API exclusive.
All support 1 Million context
Big focus on improved coding capabilities.
4.1 Pricing:
Input/Output per million 2/8
Mini: 0.4/1.6
Nano: 0.1/0.4
openai.com/index/gpt-4-1/
All support 1 Million context
Big focus on improved coding capabilities.
4.1 Pricing:
Input/Output per million 2/8
Mini: 0.4/1.6
Nano: 0.1/0.4
openai.com/index/gpt-4-1/
Introducing GPT-4.1 in the API
Introducing GPT-4.1 in the API—a new family of models with across-the-board improvements, including major gains in coding, instruction following, and long-context understanding. We’re also releasing o...
openai.com
April 14, 2025 at 5:12 PM
GPT 4.1 family is announced and is API exclusive.
All support 1 Million context
Big focus on improved coding capabilities.
4.1 Pricing:
Input/Output per million 2/8
Mini: 0.4/1.6
Nano: 0.1/0.4
openai.com/index/gpt-4-1/
All support 1 Million context
Big focus on improved coding capabilities.
4.1 Pricing:
Input/Output per million 2/8
Mini: 0.4/1.6
Nano: 0.1/0.4
openai.com/index/gpt-4-1/
Google is not winning because they outpace OpenAI. They are basically equal in quality. Google is winning because they can offer anything for much cheaper. OpenAI has to rush their own chips if they want to have a chance.
April 9, 2025 at 11:40 AM
Google is not winning because they outpace OpenAI. They are basically equal in quality. Google is winning because they can offer anything for much cheaper. OpenAI has to rush their own chips if they want to have a chance.
Social media is more and more used to direct the radicalization of the population towards other countries, groups or problems.
April 8, 2025 at 7:46 AM
Social media is more and more used to direct the radicalization of the population towards other countries, groups or problems.
So Llama 4 is out and they are all MoE models. Which is great for API costs, but is clearly not build anymore for consumer PCs. So while open source is making great progress, we are moving away from having our personal local models.
April 5, 2025 at 7:58 PM
So Llama 4 is out and they are all MoE models. Which is great for API costs, but is clearly not build anymore for consumer PCs. So while open source is making great progress, we are moving away from having our personal local models.
I am convinced that the current hardware costs will force AI companies to offload some of the compute onto the users device. I can imagine a few scenarios:
April 2, 2025 at 8:43 AM
I am convinced that the current hardware costs will force AI companies to offload some of the compute onto the users device. I can imagine a few scenarios:
GPT-4o image generation feels like a better version of Gemini flash image generation. I wonder if the new Gemini 2.5 Pro will get image output enabled and will be able to beat 4o
March 25, 2025 at 8:25 PM
GPT-4o image generation feels like a better version of Gemini flash image generation. I wonder if the new Gemini 2.5 Pro will get image output enabled and will be able to beat 4o
So exited to see if we can make progress on this, this year.
March 24, 2025 at 10:00 PM
So exited to see if we can make progress on this, this year.
If the US leaves NATO, will they also remove all their military bases around the globe? And if not, are technically invading half the world?
March 2, 2025 at 5:07 PM
If the US leaves NATO, will they also remove all their military bases around the globe? And if not, are technically invading half the world?
If we ignore cost for a second, GPT-4.5 is a great model. Reasoning models trained on this base could get scary good. Cost will come down when hardware and software gets better over time, so we just have to treat it as a look at the future.
March 1, 2025 at 5:00 PM
If we ignore cost for a second, GPT-4.5 is a great model. Reasoning models trained on this base could get scary good. Cost will come down when hardware and software gets better over time, so we just have to treat it as a look at the future.
Watching american politics makes me sick. Satire is so dead.
February 28, 2025 at 7:50 PM
Watching american politics makes me sick. Satire is so dead.
Really exited for ARC-AGI-2.
The first one was the most interesting benchmark to follow.
The first one was the most interesting benchmark to follow.
February 26, 2025 at 10:47 PM
Really exited for ARC-AGI-2.
The first one was the most interesting benchmark to follow.
The first one was the most interesting benchmark to follow.
It is so funny. Kimi and Deepseek seem to release papers at the same time and each time with similar content. Are they working together or are just on the same development roadmap?
February 18, 2025 at 5:10 PM
It is so funny. Kimi and Deepseek seem to release papers at the same time and each time with similar content. Are they working together or are just on the same development roadmap?