Ah, i see the confusion. It seems that I lost focus when writing the first part. I was referring to that an AI model is able to "understand" a topic (simplification due to the format of short text messages) and give you a better explanation that will fit your understanding level

Similarly to how you used toy have a textbook as your only source of information, but then you got new sources of information on the internet. However, with LLM, now you can actually get a description or explanation in any form you like, and with the benefit of asking questions

I should add more context: by "better" explanation, I do not mean a revolutionary news of describing things, but better, as in better tailored to the student. It's easier to ask questions and domb down description to make it easier to understand

If you actually read what I said, it will be easier to understand: bsky.app/profile/slap...

Btw in the format of short messages, it is common to make simplifications or broader claims

When it comes to AI, LLMs are giving good explanations on topics it has trained on and new advancements in preventing hallucinations, both imported quality and reliability of the models

That is strawman of my argument, and you are ignoring most of what I said. Have you ever tried LLM it for anything advanced but niche? Like certain configurations in Azure that are poorly documented? That was the first time I realized how valuable LLMs can be

To begin with yes. It happens to understand topics and concepts by providing much better explanations or giving much more extensive descriptions. Ofcourse, in addition to other advantages, but simply the power of being able to ask the stupid questions and request simplified explanations is important

It was referring to AI being good for explaining certain concepts that can be hard to learn on your own, while also being able to some some assignments for you which makes life easier in the short term, but robes you of learning how to solve those tasks independently. I do not see platitudes here

People at Nvidia and OpenAi probably see DeepSeek in their nightmares

All the stuff that DeepSeek used was publicly available (in papars and libraries), so I do not understand why they did not crate a version with the alternative approach as well

I guess deepseek turned of dlss

Most of the modern economy is built on hype. Everything from nft or meme coins to stock values is based on human perception of their value rather than objective factors. That is the result of our financial system

AGI can also be a developer in a county with unrestricted capitalism and lobbyism. In a county, were people with power think increases in profits are worth human life. In that case, AGI will also be a tool of oppression and wealth extractions rather than a loot for the betterment of humanity

The Ford Motor Company CEO taunts Reuther by asking, "Walter, how are you going to get these robots to pay union dues?" Reuther comes right back at Ford, asking, "Henry, how are you going to get them to buy your cars?"

Lastly, DeepSeek models are big (671 billion params), so it is heavy and not magically lighter than other models, but you can use distillations of models on smaller devices, which was used a lot in academia and, in my opinion, will become a common approach in the short term for smaller devices

DeepSeek-R1 is the reasoning model (similar to OpenAI's o-series) that generates the problem-solving steps before solving a given problem. I have read a bunch of articles that praise R1 for using... reinforcement learning... approach during training...

example: www.livescience.com/technology/a...