If the AI bubble like the Internet, in what year are we now?

[WayWardCloud]

Why on Earth are we forbidding China to buy chips yet allowing Meta to release their models open source?? If this is such strategic tech protect it all the way!

 

(real question, no expert)

[Spekulatius]

13 hours ago, winjitsu said:

 

 

And this article is great too: https://www.interconnects.ai/p/deepseek-v3-and-the-actual-cost-of
showing that they were training on-top of ChatGPT (and I suspect open-source Llamma like a few other Chinese firms) which used billions to setup. So I think there's alot more to this than $5mm. Likely a marketing stunt that I'm sure they'll want to use to raise billions off of.

 

By of course happy to eat humble pie if I'm wrong

 

Could well be, but I am asking myself that even if true, does it matter? Deepseek shows that you can create their own LLM that runs efficiently build upon another LLM that is freely available. So it seems that you need less NVDA chips and datacenters either way and also that any nation even target of export controls can do so.

 

Also, the allegation that they trained on ChatGPT isn’t proven. And yes, they can likely raise billions.

[brianr27]

4 hours ago, WayWardCloud said:

Why on Earth are we forbidding China to buy chips yet allowing Meta to release their models open source?? If this is such strategic tech protect it all the way!

 

(real question, no expert)

Open source is not clearly defined for AI yet. A lot of companies are open washing their AI technology and not really open sourcing all elements of their AI. There are various efforts such as this one to define open source AI: https://opensource.org/ai/open-source-ai-definition. 

 

The main arguments for open source AI are:

• Transparency and collaboration allows for greater scrutiny, enabling researchers and developers to identify and address biases, vulnerabilities, or errors in models. It fosters innovation by enabling collaboration across organizations and reducing duplication.

• Decentralization of power reduces the dominance of a few large organizations, encouraging healthy competition.

• Access to models and training data speeds up scientific progress, allowing smaller entities to contribute.

The main risks stem from poor governance. If the governance doesn't promote a vibrant community of contributors from a broad array of sources in industry and academia to contribute, the projects tend to lose momentum. The most successful open source communities (e.g. Linux) are not dominated by one company or organization and are well governed by non for profit foundations.

National security risks were raised as a concern in the early days of open source software too. The same arguments were made around giving adversaries unrestricted access to powerful software. My view is that closed-source systems pose greater risks due to their opacity, making it harder to detect or prevent misuse. Insider threats are a far greater risk from closed source. 

 

This nature article has a bit more background:

https://openuk.uk/wp-content/uploads/2024/06/Not-all-%E2%80%98open-source-AI-models-are-actually-open-heres-a-ranking-.pdf

 

https://www.reddit.com/r/OpenAI/comments/1dm2odz/not_all_open_source_ai_models_are_actually_open/

 

[MungerWunger]

interview with the DeepSeek founder from mid 2024: 

https://drive.google.com/file/d/1DW5ohZWxoCEOdrUQjokKreuArHqJdtKb/view

 

"Because we believe the most important thing right now is to participate in the global
wave of innovation. For many years, Chinese companies have been accustomed to leveraging
others’ technological innovations for application and monetization. But this isn’t something that
should be taken for granted. In this wave, our starting point isn’t to capitalize on an opportunity
but to push to the forefront of technology and contribute to the development of the broader
ecosystem."

[winjitsu]

7 hours ago, Spekulatius said:

Could well be, but I am asking myself that even if true, does it matter? Deepseek shows that you can create their own LLM that runs efficiently build upon another LLM that is freely available. So it seems that you need less NVDA chips and datacenters either way and also that any nation even target of export controls can do so.

 

Also, the allegation that they trained on ChatGPT isn’t proven. And yes, they can likely raise billions.

Yeah I've come around to this same thought over past few hours.

The fact its open source, and a few people have started to replicate results on a smaller scale means its legit advancement in the state of the art. Big win for the open source model and building together.

[mattee2264]

If there is any truth to this then can't be good for NVIDIA as indicates chips may no longer be a limiting factor and a second hand market will soon develop as early start-ups with hugely expensive NVDIA GPU inventories may collapse unable to generate a sufficient return on their early investments, and latecomers have the advantage because they can buy cheaper and fewer chips and target niches. 

 

Monopolistic competition seems to be emerging and there will be a multitude of differentiated LLMs that make money in the short run but don't make any money in the long run. Just like autos and airlines. 

[Luke]

 

Interesting take and worth thinking about folks.

 

Not sure about 50x better efficiency but its surely ruining perplexity, gemini, and chatgpt if the model is as good for free...

Edited 18 hours ago by Luke

[dealraker]

17 hours ago, Spekulatius said:

Could well be, but I am asking myself that even if true, does it matter? Deepseek shows that you can create their own LLM that runs efficiently build upon another LLM that is freely available. So it seems that you need less NVDA chips and datacenters either way and also that any nation even target of export controls can do so.

 

Also, the allegation that they trained on ChatGPT isn’t proven. And yes, they can likely raise billions.

To the degree stock price number go up the resulting optimism turbo charges irrational expectations and wild-ass capital allocation.   And this theme always registers the hardest within those we consider the brilliant minds of the era.  It is nothing new.

 

The circles of technology expertise and superb capital allocation only overlap to a small degree, either by skill or luck.   But you can't grasp any of this until the stock prices quit going up.  Nothing can be fully rationally discussed or analyzed until prices stop escalating and that particularly includes our tech and crypto worlds of today.   

 

 

 

 

[73 Reds]

16 minutes ago, dealraker said:

To the degree stock price number go up the resulting optimism turbo charges irrational expectations and wild-ass capital allocation.   And this theme always registers the hardest within those we consider the brilliant minds of the era.  It is nothing new.

 

The circles of technology expertise and superb capital allocation only overlap to a small degree, either by skill or luck.   But you can't grasp any of this until the stock prices quit going up.  Nothing can be fully rationally discussed or analyzed until prices stop escalating and that particularly includes our tech and crypto worlds of today.   

 

 

 

 

I dunno, every speculative bubble that I can recall has been fully recognizable by those who really care to well before the bubble bursts.  Many folks are willing participants simply because it is often easier to profit from temporary human behavior than it is to make money any other way.  The fact that the last remaining greater fools lose most or everything does not in any way diminish the profits from a nice momentum ride by those who are less foolish.        

[Blake Hampton]

1 hour ago, 73 Reds said:

I dunno, every speculative bubble that I can recall has been fully recognizable by those who really care to well before the bubble bursts.  Many folks are willing participants simply because it is often easier to profit from temporary human behavior than it is to make money any other way.  The fact that the last remaining greater fools lose most or everything does not in any way diminish the profits from a nice momentum ride by those who are less foolish.        

But when do you decide to get out?

[james22]

1 hour ago, Blake Hampton said:

But when do you decide to get out?

 

It's much easier deciding to get out (risking further upside) when you've gains in hand than deciding to get in (risking further downside) - especially if you've catching up to do.

[james22]

 

[73 Reds]

2 hours ago, Blake Hampton said:

But when do you decide to get out?

That is up to everyone who plays the game.  Personally, I don't play because I'm better at making reliable money in other ways.