maxkannen.bsky.social - Profile | ThreadSky | a Reddit-style client for Bluesky

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?

submitted 4 days ago • 0 comments

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.

submitted 5 days ago • 0 comments

Watching american politics makes me sick. Satire is so dead.

submitted 5 days ago • 0 comments

Really exited for ARC-AGI-2. The first one was the most interesting benchmark to follow.

submitted 7 days ago • 0 comments

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?

submitted 16 days ago • 0 comments

In the future we will have custom kernels for new models hours after they come out. This will make inference so much better. developer.nvidia.com/blog/automat...

submitted 20 days ago • 0 comments

You can get small models to be really good at math benchmarks, but they are no longer language models in this case. They become math models. We had a point around a year ago where the models where very general and now we go back to specialized models. pretty-radio-b75.notion.site/DeepScaleR-S...

submitted 22 days ago • 2 comments

I might soon submit research to a conference for the first time. Not sure yet if I'll stay in research, but this feels good.

submitted 25 days ago • 0 comments

Mistral is single-handedly keeping Europe competitive. Their updates to le chat adds all the futures that became standard for other services. mistral.ai/en/news/all-...

submitted 27 days ago • 0 comments

Karpathy is back with a new Video. I am a huge fan of his educational content. I wish more experts in the field would make videos like this. youtu.be/7xTGNNLPyMI?...

submitted 29 days ago • 0 comments

I honestly do not understand how everyone is talking about R1 and o3-mini and how no one is talking about Gemini. Their reasoning model is somewhere between o1-mini and o3-mini, and available for free for weeks now.

submitted 32 days ago • 2 comments

New OpenAI strategy: Let o3-mini con investors out of their money.

submitted 34 days ago • 1 comment

I really want a model that can generate high quality Manim code to explain stuff. I tried building this a year ago, but the models weren't good enough back than. Are they now? A reasoning model with video understanding could maybe do it in multiple iterations.

submitted 37 days ago • 0 comments

You don’t need 500 billion to build AGI. But you need it to deploy at scale. openai.com/index/announ...

submitted 43 days ago • 0 comments

Deepseek R1 is Open Source now. It uses the same base model as Deepseek v3 and is probably the biggest reasoning model that exists at the moment. I am really exited for the benchmarks. huggingface.co/deepseek-ai/...

submitted 45 days ago • 1 comment

I talked about material science in my 2025 predictions. I think that outside of biology, material science is the best field for AI. I expect a lot more progress here in the next 12 months. www.microsoft.com/en-us/resear...

submitted 49 days ago • 0 comments

More experts for faster inference is a general trend that we see, because the generation speed becomes more important. And it becomes more important because they are no longer just Chatbots, and reading speed is not enough. Reasoning and agents need speed t.co/IswiryeO0z

submitted 50 days ago • 0 comments

So now that ChatGPT has tasks... What are you supposed to do with them? Set a timer for cooking? Wake me up? Give me the daily news? I have a phone that does all of this natively. What is the use case when I have to open the app anytime?

submitted 50 days ago • 0 comments

Mistral is finally back with a code model. Since copilot is free, it will be hard to convince me, but I will give it a try. mistral.ai/news/codestr...

submitted 52 days ago • 0 comments

So where is the difference between that and what Putin is saying?

submitted 57 days ago • 0 comments

My recap of 2024 and predictions for 2025. I do this every year and had a really hard time this year. I hope you still enjoy and I am open to feedback. mkannen.tech/looking-back...

submitted 65 days ago • 1 comment

And here we see the advantage: cerebras.ai/blog/cepo

submitted 66 days ago • 0 comments

The Deepseek v3 paper is out and the training is very interesting. 1. They use Multi-token prediction during training which Meta released a paper about a few months ago. 2. They used their r1 reasoning models to distill reasoning into v3. github.com/deepseek-ai/...

submitted 69 days ago • 1 comment

Deepseek released a giant MoE model. 685B is massive, but it has 256 experts and uses top 8 as default. As far as I know this is the first time such a huge number of experts was used in a OS production model. huggingface.co/deepseek-ai/...

submitted 71 days ago • 0 comments

When I was in middle school, I was obsessed with Conway's game of life and the idea of self-emerging live in a simulation. This feels like a Christmas present for my younger self. pub.sakana.ai/asal/

submitted 72 days ago • 0 comments

Nvidia needs to dominate inference otherwise they will lose their valuation soon. Nothing is more important in the next two years than getting faster and cheaper inference. We are going to need trillions of tokens per hour.

submitted 73 days ago • 0 comments

I do not see myself providing anything of value to an economy when I leave university which an AI will not be able to do better at this point. I am just glad that we have such a stable global political system that will manage that. Oh wait ...

submitted 76 days ago • 0 comments

It is very expensive and not available yet, but this is at least some form of proto-AGI. arcprize.org/blog/oai-o3-...

submitted 76 days ago • 1 comment

How cool is that: genesis-embodied-ai.github.io

submitted 77 days ago • 0 comments

For the first time I found something useful. Doing research can actually be fun.

submitted 79 days ago • 0 comments

This was my first programming experience.

submitted 79 days ago • 0 comments

This is absolutely insane for this size

submitted 85 days ago • 0 comments

Whatever OpenAI announces today, DeepMind won. Gemini 2.0 has everything that OpenAi is promising since the 4o release

submitted 85 days ago • 0 comments

Gemini 2.0 looks very promising. The advantage of TPUs becomes very apparent. deepmind.google/technologies...

submitted 85 days ago • 1 comment