site banner
PaperclipPerfector  3mo ago (text post)   36654 thread views

Culture War Roundup for the week of January 20, 2025

This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.

Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.

We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:

  • Shaming.

  • Attempting to 'build consensus' or enforce ideological conformity.

  • Making sweeping generalizations to vilify a group you dislike.

  • Recruiting for a cause.

  • Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.

In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:

  • Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.

  • Be as precise and charitable as you can. Don't paraphrase unflatteringly.

  • Don't imply that someone said something they did not say, even if you think it follows from what they said.

  • Write like everyone is reading and you want them to be included in the discussion.

On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.

5
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

To continue the drama around the stunning Chinese DeepSeek-r1 accomplishment, the ScaleAI CEO claims DeepSeek is being coy about their 50,000 H100 GPUs.

I realize now that DeepSeek is pretty much the perfect Chinese game theory move: let the US believe a small AI lab full of cunning Chinese matched OpenAI, with a tiny fraction of the compute budget, with no ability to get SOTA GPUs. Let the US believe the export regime works, but that it doesn't matter, because Chinese brilliance is superior, demoralizing efforts to strengthen it. Additionally, it would make the US skeptical of big investment in OpenAI capital infrastructure because there's no moat.

Is it true? I have no idea. I'm not really qualified to do the analysis on the DeepSeek results to confirm it's really the run of a small scrappy team on a shoestring budget end-to-end. Also what we don't see are the potentially 100-1000 other labs (or previous iterations) that have tried and failed.

The results we have now are that -r1 b14 and b32 are fairly capable on commodity hardware, and it seems one could potentially run the 671b model which is kinda maybe but not actually on par with o1 on a something that costs as much as a tinybox ($15k). That's a remarkable achievement, but at what total development cost? $5 million in compute + 100 Chinese worth of researchers would be stunningly impressive. But if the true cost is actually a few more OOMs, it would mean the script has not been completely flipped.

I maintain that a lot of OpenAI's current position is derivative of a period of time where they published their research. You even have Andrej Karpathy teaching you in a lecture series how to build GPT from scratch on YouTube, and he walks you through the series of papers that led to it. It's not a surprise that competitors can catch up quickly if they know what's possible and what the target is. Given that they're more like ClosedAI these days, would any novel breakthroughs be as easy to catch up on? They've certainly got room to explore them with a $500b commitment to play with.

Anyway, do you believe DeepSeek?

Alex Wang is an opportunistic psychopath who's afraid of his whole Pinoy-based data generation business model going bust in the era of synthetic chains of thought. Therefore he's dishonestly paraphrasing Dylan Patel (himself a China hawk peddling rationales for more export controls) who had said “they have 50000 Hoppers” once, without evidence. But the most likely Hopper model they have is H20, an effectively inference-only chip, that has negligible effect on pretraining costs and scale for V3 and R1.

Yes I do believe DeepSeek. This is not really a political issue but a purely technical. Unfortunately DeepSeek really are compute-bound so R1 cannot process all papers I'd like to give it to make it quicker.

The political narrative does not even work, it's purely midwit-oriented, nobody in the industry imagines leading labs can be deceived with some trickery of this kind.

Inference costs are wholly addressed by Hyperbolic Labs (US) and some others already serving it for cheaper.

which is kinda maybe but not actually on par with o1

It's superior to o1 as a reasoner and a thinker. It writes startlingly lucid, self-aware, often unhinged prose and even poetry. It can push back. It is beyond any LLM I have seen including Sonnet and Opus. This becomes obvious after minutes of serious interaction. It just has less polish as a product because they haven't been milking the world for interaction data since 2019. They have 0.8-1.5 M quality samples for instruction finetuning. OpenAI had accumulated tens of millions if not hundreds.

For me it's something of an emotional issue. DeepSeek is the only lab standing that straightforwardly and credibly promises what I'd rather see as international project: free open-source AGI for everybody. I've been monitoring their rise for well over a year, reading every paper and even their blogposts in Chinese. Nothing that they claim is inconsistent, indeed it's all been predictable since 2023, all part of a very methodical, flawless, truly peak quant fund (that's their capital source and origins) execution towards the holy grail, “answering the ultimate question with longtermism”, as they put it. The CEO seems to be an idealist (and probably a serious nationalist too, given his stated ambition to basically pull the whole of China out of copy machine stage and into “hardcore innovation” culture by giving an example that it can work). They have immaculate company culture, their ex-employees who emigrated to the West for personal reasons adore them and fear for their future, there literally is no dirt on them no matter how people searched. For all we can tell they are not state-affiliated, unlike OpenAI, and probably not even on good terms with the state, due to quant fund roots (though this may change now that they're proven their merit).

This is not a Sputnik moment for the US. The US has a secure and increasing lead due to bog standard logistics and capital advantage, as always. What this should be is “are we the baddies?” moment.

Also, it's a moment to ask oneself how high are margins on Western model providers, and whether it's a true free market. Because Liang Wenfeng himself does NOT think they're that far ahead in efficiency, if they are ahead at all.

Wenfeng is invited to government functions so I simply don't believe that they are not on good terms with the state and I'm skeptical that they are less tied to the state than openai.

Not that this should change much - they still have a good model, though I wouldn't exactly trust the headline training cost numbers since there's no way to verify how many tokens they really trained the model on.

That's the point: He is invited NOW, after "suddenly" shipping a model on Western Frontier level.

7 months ago I have said:

We don't understand the motivations of Deepseek and the quant fund High-Flyer that's sponsoring them, but one popular hypothesis is that they are competing with better-connected big tech labs for government support, given American efforts in cutting supply of chips to China. After all, the Chinese also share the same ideas of their trustworthiness, and so you have to be maximally open to Western evaluators to win the Mandate of Heaven.

Presumably, this was true and this is him succeeding. As I note here.

As for how it used to be when he was just another successful quant fund CEO with some odd interests, I direct you to this thread:

The Chinese government started to crack down on the quant trading industry amid economic slowdown, a housing crisis and a declining stock market index.

The CSI300 (Chinese Blue Chip Index) reached an all-time low. They blamed high frequency traders for exploiting the market and causing the selloff.

  • Banned a quant competitor from trading for 3 days
  • Banned another from opening index futures for 12 months
  • Required strategy disclosures before trading
  • Threatened to increase trading costs 10x to destroy the industry High-Flyer faced extinction. (High-Flyer’s funds have been flat/down since 2022 and has trailed the index by 4% since 2024)

so I stand by my conjectures.

they still have a good model, though I wouldn't exactly trust the headline training cost numbers since there's no way to verify how many tokens they really trained the model on

So you recognize that the run itself as described is completely plausible, underwhelming even. Correct.

What exactly is your theory then? That it's trained on more than 15T tokens? 20T, 30T, what number exactly? Why would they need to?

Here's a Western paper corroborating their design choices [Submitted on 12 Feb 2024]:

Our results suggest that a compute-optimal MoE model trained with a budget of 1020 FLOPs will achieve the same quality as a dense Transformer trained with a 20× greater computing budget, with the compute savings rising steadily, exceeding 40× when budget of 1025 FLOPs is surpassed (see Figure 1). … when all training hyper-parameters N, D, G are properly selected to be compute-optimal for each model, the gap between dense and sparse models only increases as we scale… Higher granularity is optimal for larger compute budgets.

Here's DeepSeek paper from a month prior:

Leveraging our architecture, we subsequently scale up the model parameters to 16B and train DeepSeekMoE 16B on a large-scale corpus with 2T tokens. Evaluation results reveal that with only about 40% of computations, DeepSeekMoE 16B achieves comparable performance with DeepSeek 7B (DeepSeek-AI, 2024), a dense model trained on the same 2T corpus. We also compare DeepSeekMoE with open source models and the evaluations demonstrate that DeepSeekMoE 16B consistently outperforms models with a similar number of activated parameters by a large margin, and achieves comparable performance with LLaMA2 7B (Touvron et al., 2023b), which has approximately 2.5 times the activated parameters. Evaluation results show that DeepSeekMoE Chat 16B also achieves comparable performance with DeepSeek Chat 7B and LLaMA2 SFT 7B in the chat setting. Encouraged by these results, we further undertake a preliminary endeavor to scale up DeepSeekMoE to 145B. The experimental results still validate its substantial advantages over the GShard architecture consistently. In addition, it shows performance comparable with DeepSeek 67B, using only 28.5% (maybe even 18.2%) of computations.

As expected they kept scaling and increasing granularity. As a result, they predictably reach roughly the same loss on the same token count as LLaMA-405B. Their other tricks also helped with downstream performance.

There is literally nothing to be suspicious about. It's all simply applying best practices and not fucking up, almost boring. The reason people are so appalled is that American AI industry is bogged down in corruption covered with tasteless mythology, much like Russian military pre Feb 2022.

It's all simply applying best practices and not fucking up, almost boring.

It's pretty weird: there's nothing there that any of the big labs in the West should have trouble replicating a hundred times over, and DeepSeek still managed to make something that can trade blows with them (and subjectively win, more often than not).

Might it really be just clarity of purpose leading to focusing on what matters? About a week ago, I remember Claude lecturing me, apropos of nothing, a bit about how it's best to buy from local bookstores instead of online retailers in response to me asking about what kind of textbook would be used for a particular course. I've not experienced DeepSeek doing anything even close to that, and it makes me wonder if the extraneous post-training being lathered on is the real difference here. Western models get distracted and are pulled in a thousand different directions, while DeepSeek can focus on what's relevant.

I'm not impressed by "they work in a field censured by the state, therefore they have no state connections". Jack Ma was also (personally!) censured by the state, and he's certainly connected. In the US, the DOJ seeks to break up Google. The Sacklers got sued into oblivion. All these people are connected - getting rekt by government action is an occupational hazard of being Noticed by the government, and those who are Noticed typically try to ingratiate themselves.

Thanks for the links about the model training, that's interesting reading.

What is this place?

This website is a place for people who want to move past shady thinking and test their ideas in a court of people who don't all share the same biases. Our goal is to optimize for light, not heat; this is a group effort, and all commentators are asked to do their part.

The weekly Culture War threads host the most controversial topics and are the most visible aspect of The Motte. However, many other topics are appropriate here. We encourage people to post anything related to science, politics, or philosophy; if in doubt, post!

Check out The Vault for an archive of old quality posts. You are encouraged to crosspost these elsewhere.


Why are you called The Motte?

A motte is a stone keep on a raised earthwork common in early medieval fortifications. More pertinently, it's an element in a rhetorical move called a "Motte-and-Bailey", originally identified by philosopher Nicholas Shackel. It describes the tendency in discourse for people to move from a controversial but high value claim to a defensible but less exciting one upon any resistance to the former. He likens this to the medieval fortification, where a desirable land (the bailey) is abandoned when in danger for the more easily defended motte. In Shackel's words, "The Motte represents the defensible but undesired propositions to which one retreats when hard pressed."

On The Motte, always attempt to remain inside your defensible territory, even if you are not being pressed.


New post guidelines

If you're posting something that isn't related to the culture war, we encourage you to post a thread for it. A submission statement is highly appreciated, but isn't necessary for text posts or links to largely-text posts such as blogs or news articles; if we're unsure of the value of your post, we might remove it until you add a submission statement. A submission statement is required for non-text sources (videos, podcasts, images).

Culture war posts go in the culture war thread; all links must either include a submission statement or significant commentary. Bare links without those will be removed.

If in doubt, please post it!


Rules


Recommended Posts And Communities

Recommended Realtime Chats