DeepSeek touts new training method as China pushes AI efficiency

[Jan]

DeepSeek touts new training method as China pushes AI efficiency

DeepSeek published a paper outlining a more efficient approach to developing AI, illustrating the Chinese artificial intelligence industry’s effort to compete with the likes of OpenAI despite a lack of free access to Nvidia Corp. chips.

The document, co-authored by founder Liang Wenfeng, introduces a framework it called Manifold-Constrained Hyper-Connections.

[Bloomberg]

 

[Brian_G]

DeepSeek, SkyNet, same, same...

 

[Totempole]

Hands up who prefers deepseek

View attachment 1874949

 

[Herr der Verboten]

DeepSeek touts new training method as China pushes AI efficiency

DeepSeek published a paper outlining a more efficient approach to developing AI, illustrating the Chinese artificial intelligence industry’s effort to compete with the likes of OpenAI despite a lack of free access to Nvidia Corp. chips.

The document, co-authored by founder Liang Wenfeng, introduces a framework it called Manifold-Constrained Hyper-Connections.

[Bloomberg]

They should have gone with deep-throat.

 

[garyc]

Yesterday tried a comparison between ChatGPT and Deepseek for generating Linux bash scripts. I had previously abandoned ChatGPT for this since it always screws up. It did so again, but Deepseek got it right.

 

[Sapphiron]

Hands up who prefers deepseek

View attachment 1874949

I spent today giving Deepseek another go with some of the problematic items that ChatGPT keeps stumbling on.


Most of my questions pertain to configuration of ERP, CRM and inventory systems. Things with definitive answers in documentation (although the documentation is usually very fragmented)


1. ChatGPT will regularly hallucinate answers referring to features that don't exist and when asked to cite a source, it will generate an fake URL and insist that its a real one. Deepseek does not do this. It seems to rely more on the contents of realtime internet searches than it does its training data.

2. Deepseek does not try and butter me up and stroke my ego.

3. ChatGPT is about 2x faster

4. Deepseek is far more likely to site and quote sources, rather than just abstracting it.

5. Deepseek does not seem to do as well in summarizing Legal documents

 

[neoprema]

No thanks, i am not interested in any commie-trained AI that may or may not be stealing data for Chinese intelligence.. ChatGPT is fine for me.

 

[Sapphiron]

No thanks, i am not interested in any commie-trained AI that may or may not be stealing data for Chinese intelligence.. ChatGPT is fine for me.

Make no mistake, your data is also being stolen, just by oligarch robber barons.

Privacy and confidentiality has become an illusion.

 

[neoprema]

Make no mistake, your data is also being stolen, just by oligarch robber barons.

Privacy and confidentiality has become an illusion.

I mean I won’t argue. But it feels less threatening not being the Chinese lol

 

[kolaval]

Wanting to play around I asked Gemini, chatgpt and Claude about creating a Docker image with nodejs, vite and persistent volume for creating a ShadCN app.
They all gave answers assuming I already had a ShadCN project wich I just wanted to containerize.

Sometimes they're like a family member that tries to convince you the earth is flat.

 

[w1z4rd]

Wanting to play around I asked Gemini, chatgpt and Claude about creating a Docker image with nodejs, vite and persistent volume for creating a ShadCN app.
They all gave answers assuming I already had a ShadCN project wich I just wanted to containerize.

Sometimes they're like a family member that tries to convince you the earth is flat.

I have issues as well. Finding ChatGPT great for problem solving, but you often have to spend a bit of time working with the AI to get it going in the right direction. It can be stupid AF sometimes.

Im kinda at the point I assume anyone who thinks AI will take all our jobs, hasnt used AI. It might in the future, but its nowhere near ready to replace the human (at least in my field of work). If someone who didnt know the work I did, tried to do my job with AI they would destroy the servers we work on. They wouldnt know which outputs were legit commands to run and which are hallucinations.

I do not work in the content creation field (at least with consent) so perhaps its different in different industries.

 

[semaphore]

Please share your prompt and the resulting output?

There’s nothing wrong with DeepSeek, it generates about the same slop as Claude or ChatGPT.

 

[Sapphiron]