What is the impact of DeepSeek on the AI sector? 🔥

[qrdl]

n2u casually dropping alpha, lol.

Regardless of what happens, imho, I think this premarket behavior confirms the analysis above. It was not an ‘irrational’ drop as folks are saying, one that is purely based on some notion of reduced computer usage.

Factors such as reduction in NVDA moat and potential future export restrictions also were an issue, though I believe the latter one was very BLIP, while the former might be sticky given Friday’s AMDs announcements and efforts alongside deepseek coupled with r1

[qrdl]

OpenAI says DeepSeek used its models illegally, and it has evidence to prove...

Lol. I mean, this was obv from day one. But deepseek has published so much as opensource and revealed so much, they have given back, where close sourced companies are just leaching.

There is quite an irony when companies who there entire product is based on the IP of others argue that others are stealing their IP.

The problem is that unless OAI says that they retain copyright and patents to all of their content they give users, how do they stop takeoff from happening? (see my other thread on this)

It’s very very possible, imho, that DeepSeek did not use OAI directly but instead just trained on their outputs provided by others.

[sashirestela]

Markets need more time to see the real impact

[qrdl]

Agreed, but I think we can disentangle some of the underlying impacts by observing realtime some of the reactions that are occurring and how this evolves.

While I still believe the watershed moment happened over Christmas, and that was when I wanted to talk about it but nobody else did, it’s here now in everyone’s consciousness.

And let me tell you something, as I said then, this is a BFD. Don’t ignore it. A singularity is rapidly approaching.

However, given the crying above from OpenAI, they might be setting the stage for some false flag legal attack on DeepSeek. They may have been told / encouraged to do this by government on their board.

This might result in what @N2U is alluding to in terms of export restrictions.

And, I will grudgingly admit, China is an adversary, so perhaps they have a point. I certainly do not agree with their policies on free speech.

As long as it’s China they are dealing with and not OpenSource. If it’s the latter, well, that is upsetting.

@dannetstudio

And really interesting things have quickly emerged, such as Aider’s architect mode that allows combining r1 with sonnet 3.5, achieving an efficiency never seen before: R1+Sonnet set SOTA on aider’s polyglot benchmark | aider

Love the aider stuff! So much fun to use. Any programmers here should take a look at it.

Leaders: I’ll try to stop with the double posting. Very exciting times and folks are actually paying attention - even though it took portfolio volatility for make them do it, lol.

[razvan.i.savin]

@qrdl

OpenAI says DeepSeek used its models illegally, and it has evidence to prove it, new report claims | TechRadar

Well, isn’t this the same as what OpenAI did in the past? → “Trained a few iterations”

[qrdl]

Oh course. OpenAI’s entire product is the distillation of IP generated by everyone else. They know this, everyone does.

However, DeepSeek is seriously disrupting things and likely panic has set in so folks are flailing and looking for a way to throw sand in the gears here.

The questions are:

1. How much of this legal attack is this OpenAI and how much is it govt on OpenAI’s board?
2. How much of it is Anti Chinese versus Anti OpenSource? (If the former, fine, if the latter, not so good)
3. How far are they going to take this?
4. Will govt get involved?

[anon10827405]

And how does OpenAI prevent “Write training data for $20/hr” websites from laundering ChatGPT output.

It’s an impossible battle that, as always, will hurt the common person.

OpenAI, Microsoft, and it looks like the US government are fighting a battle against something inevitable. The country of origin is irrelevant IMO and is just causing pain out of some grotesque bullying tactic

[qrdl]

Yes, you have succinctly summarized the entire drama.

Singularity will occur very soon if we collectively allow all our models to distill and improve each other.

At first I thought Trump was OK with this, but now I’m not so sure.

https://www.reuters.com/technology/artificial-intelligence/alibaba-releases-ai-model-it-claims-surpasses-deepseek-v3-2025-01-29/

I already pointed out Qwen2.5-Max above, ofc. Likely distilled as well. It does emphasize the global takeoff here, but imho, it’s only 50% of the problem, the other is the alg improvements like RL which are very very likely being at least hinted at by the models themselves.

Chinese firms ‘distilling’ US AI models to create rival products, warns OpenAI

The OpenAI spokesperson added that it was now “critically important” that the company worked with the US government to “best protect the most capable models from efforts by adversaries and competitors to take US technology”.

Sand in the gears! I suppose, as much as I want to see AGI, there are some risks to massive and sudden social upheaval.

The White House press secretary, Karoline Leavitt, said the US national security council was looking into the potential implications the AI app posed.

[razvan.i.savin]

If OpenAI does not take significant countermeasures, Chinese could repeat this process, continuously improving their models by distilling knowledge from OpenAI’s responses.

I’ve heard multiple versions of how this was done:

• A few months ago, there were discussions about a potential GPT-4o leak.
• Even without direct access to weights, massive prompt-response collection (Q&A data) could be enough to reconstruct model behavior.
• If I were in their place, I would:
  • Use embeddings to capture semantic relationships.
  • Combine embeddings with a vocabulary extracted from OpenAI responses.
  • Apply a parser to analyze response structures, like this one I made.
  • Leverage my custom attention mechanism for refining token dependencies, similar to this implementation.
  • Utilize Fill-in-the-Middle (FIM) to reconstruct missing sequences, just as DeepSeek likely does with FIM completion.

[qrdl]

It’s this sort of post which is a distraction from the reality of what is happening. It’s the same reason people missed what happened over Christmas.

You’re not seeing the super obvious forest because of trees. Distillation of intelligence is less about mass collection of prompt/responses and more about just the generation of raw intelligence.

In fact, arguably, the fewer prompt/responses used probably the better.

Read deepseeks paper, read kaparthy. It’s about reinforcement learning - not about imitation.

These things will learn by reasoning and training themselves. The validation tests of raw intelligence will be generated from other models, but that is very likely not going to involve some sort of mass collection as you imagine it.

GitHub - openai/evals: Evals is a framework for evaluating LLMs and LLM...

Andrej Karpathy (@karpathy) on X

There are two major types of learning, in both children and in deep learning. There is 1) imitation learning (watch and repeat, i.e. pretraining, supervised finetuning), and 2) trial-and-error learning (reinforcement learning). My favorite simple example is AlphaGo - 1) is learning by imitating expert players, 2) is reinforcement learning to win the game. Almost every single shocking result of deep learning, and the source of all magic is always 2

That all said, you need need an advanced base in order to do 2. Can’t start from scratch. x.com proved that.

But that’s OK, this cambrian explosion will lead to diverse models which means collectively we will have more effective intelligence.

Another misunderstanding here I think is that these models are going to hyper focus on math and code, with the other stuff coming later. This is different than the more fuzzier stuff being done in the west.

There are a lot of good reasons to focus on math and code. First, it’s what we need to self improve these models, but also it’s cheap to validate/train and requires little data.

[razvan.i.savin]

I agree that reinforcement learning plays a critical role in model advancement, but RL alone doesn’t explain DeepSeek’s OpenAI-like behavior. Models require a strong pretrained base before RL refinement.
If OpenAI responses were used in pretraining → whether through distillation, embeddings, or fill-in-the-middle techniques → then RL would only be fine-tuning an already similar system. The real question is: how did DeepSeek get an OpenAI-like base to begin with? If it wasn’t mass Q&A collection, was it some other indirect distillation method?

[qrdl]

Yeah, I mean, technically and maybe legally you might be right, but morally you’re so far off base I don’t know where to start. DeepSeek has given back, whereas OpenAI gives very little back in return for the mass amount of IP it has taken from humanity.

If you want to join in OpenAIs false flag attack, ok, the Chinese and any anti free speech culture are adversaries for sure. I just wish there was a way to leave open source out of it.

[razvan.i.savin]

I’m just focusing on the technical aspects here, not on politics or morality. The question is about how DeepSeek might have been trained → if OpenAI responses were used in pretraining, that could explain the similarities.
Let’s leave the broader ethical debates for another time; I’m done with this thread.

[qrdl]

Technically and legally you are very possibly right. I don’t know.

I don’t wish to offend, just to speak as honestly and sincerely and clearly as possible. I believe this story and many of its root causes is about the politics and morality of rushing towards AGI, and ignoring them is really missing the underlying activities that are occurring.

Also, I think some accusations are being thrown around here without proof. So far OpenAI has wisely avoided directing their accusations at DeepSeek specifically.

Update:

Microsoft’s security researchers in the fall observed individuals they believe may be linked to DeepSeek exfiltrating a large amount of data using the OpenAI application programming interface, or API, said the people, who asked not to be identified because the matter is confidential. Software developers can pay for a license to use the API to integrate OpenAI’s proprietary artificial intelligence models into their own applications

https://www.bloomberg.com/news/articles/2025-01-29/microsoft-probing-if-deepseek-linked-group-improperly-obtained-openai-data?embedded-checkout=true

We may never know the truth, but so far this has the momentum to the eventual removal of these open source models.

I’m not surprised at all and certainly was an outcome I have been expecting since Christmas. Whether the barn doors can be closed though remains to be seen, I think.

Going back to the original point of export restrictions, I can’t imagine Chinese govt will be happy to hear these accusations thrown at domestic companies (DeepSeek, Alibaba) which are pretty innovative and accomplished.

Trump, I thought, was wise to calm the waters here. For all the disagreements I’ve had with him, I think he’s pretty good at chilling things out.

Update: Yikes, this is escalating quickly. Oh well. Fun while it lasted. Pretty funny when the only reasonable person you think is Trump in the entire affair.

Probably not much to say at this point, will loop back in a few days. Folks have gone into full blown panic mode. Clearly a full court press to suppress DeepSeek and its advancements. Could see further declines in Nvidia as they get caught up in the geopolitics. Very likely to see this blowback against open source itself as well, especially in terms of its potential use of deepseek weights. Would not be surprised to see Alibaba accused as well.

[N2U]

@qrdl

n2u casually dropping alpha, lol.

I should probably disclose that Monday’s Nvidia drop constitutes the largest one-day loss I’ve ever taken, and I’m still hodl’ing.

Was the market overreacting? Absolutely, but a reaction was warranted. The question is just how much? Every single piece of analysis in this topic is actually valid, even if conflicting, because the market is made up of independent actors who trade based on their own ideas.

The release of DeepSeek doesn’t touch Nvidia’s moat; if anything, it cements it because DeepSeek is also built on Nvidia’s hardware. OpenAI’s moat is now reduced, and the competition may force a price reduction that could impact their revenue. Nvidia is selling shovels in a gold rush, and the commoditization of AI will only increase the need for GPUs.

@qrdl

And, I will grudgingly admit, China is an adversary, so perhaps they have a point.

The US foreign policy has consistently been to treat China as competitors instead of adversaries, but will Nvidia be allowed to continue selling GPUs to China if the development of AI suddenly becomes a major national security concern?.. There are enough Americans trying to hand over their personal data to the Chinese government that it actually DDoS’ed DeepSeek’s servers, which seems like a win for the current export restrictions.

Having the newest hardware will help stay ahead in AI development because the limiting factor at this scale is the total available compute and its efficiency.

@qrdl

How much of it is Anti Chinese versus Anti OpenSource? (If the former, fine, if the latter, not so good)

OpenAI’s lawsuit is likely driven by politics and bureaucracy. We all know that suing a Chinese company is an uphill battle, but they still have to do it to protect intellectual rights if they believe there’s some violation. But I think it’s worth noting that they haven’t sued all the other companies that make open-source models.

More Alpha

• The reaction: A price drop on Nvidia stocks seems warranted due to an increase in risk.
• The overreaction: Institutional investors, such as index and pension funds, rebalance their holdings on a daily basis. A large drop in any stock will cause them to dump stocks as they adjust for the loss in market cap, driving prices down even further. People selling their index funds also cause the drop to spread across the entire chip sector in ways that don’t correlate with reality.

If you’re okay with accepting the risk, I’d wager that this is a strong buying opportunity on Nvidia and other chip stocks.

[qrdl]

As always n2u with the highest signal to noise ratio posts. My only addendum is that things are moving very very fast right now. Everyone in power is going on the attack against DeepSeek. They see all the threats, obviously. China, AGI, OpenSource.

Only question mark now is how Trump plays this. He is the wild card to beat all wild cards. My guess is he will have to get on board and help throw sand in the gears.

Where does that leave OpenSource? Probably toast, but I’m still 50 50 on that.

Where does this leave NVDA? Hard to say. It could become like a strategic defense company, but with much higher multiples.

Biggest risk, imho, is the taiwan question. Yes, it’s very possible going there already. Welcome to the singularity, or at least its approach.

Added: One hope I have, and again I am not a fan of anti-free speech cultures (so this is kinda ironic), is that China will embrace Open Source successfully. If US goes close source, they could fall behind.

[kamil00]

If we both read the same books, will we have a similar understanding?
On the other hand, does openai pay book authors? Because one can steal and the other cannot.

" and any anti free speech culture are adversaries for sure."
that’s cool, can we learn about genders, blm and lot other things from openai? what about all the covid bans on fb? if thats not censorship then I dont know what is. what about tiktok ban, it’s ok for ig/meta/yt to hots same videos as tiktok, but its tiktok the bad one (disclaimer I hate tiktok)

[N2U]

@qrdl

As always n2u with the highest signal to noise ratio posts.

Thank you, I try my best, but much of this is just me trying to rationalize my financial losses

The question of Taiwan is definitely a precarious one, but it has been like that for some time. It would be great if this would just stay as tension between AI companies, but if the US and Chinese governments get involved, it will most likely result in mutual sanctions.

Chip fabrication is really hard; it typically takes a few years to build a factory and another few years before it can run at full capacity. Imagine trying to explain the chip manufacturing process to someone from the Middle Ages; it would sound like complete magic:

“So we dip the crystals in the special water, then bend light to etch tiny symbols into them. We can then force lightning through those crystals to make them do our work for us.”

The US does not currently possess the manufacturing capabilities to actually make their own chips. Attempts are being made to alleviate that issue, but it will still take years before they will be able to meet the demand for chips. A total collapse in chip trade between the US and Taiwan would leave Nvidia without products to sell.

@qrdl

Where does that leave OpenSource? Probably toast, but I’m still 50/50 on that.

Open source will never be completely toast; it’s just a completely different business model, one that can be entirely sustainable if done correctly. But a for-profit company will always have an edge simply because they have more money to blow on development.

What should OpenAI’s response to DeepSeek be? They’ve already responded legally, but I don’t care about that. I believe it would be good for OpenAI to spend 6 million of Altman’s pocket change on developing their own open-source model to deflate the DeepSeek hype balloon.

[merefield]

Yes: merefield (Robert) · GitHub

[dannetstudio]

Even the web interface of the new model launched by Alibaba Qwen 2.5 Max and its chat is vastly superior to what we were seeing with ChatGPT, as it has integration with Artifacts, web, image generator, video, and others.

• Qwen2.5-Max: Exploring the Intelligence of Large-scale MoE Model | Qwen
• https://chat.qwenlm.ai/

And behind this is no longer DeepSeek and its mediocre infrastructure, there is the Asian giant AliBaba.

This is also very interesting:

• Open-R1: a fully open reproduction of DeepSeek-R1
• GitHub - huggingface/open-r1: Fully open reproduction of DeepSeek-R1

A project that fully and transparently reproduces the r model, so that each piece is accessible and reproducible, not only the final model but also the data used, training processes or other tools for its development.

From what I am seeing, the superiority of these models compared to OpenAi is no longer limited only to what has to do with programming.

I think this is just beginning, and it is unstoppable, thanks to Open Source.