site banner
PaperclipPerfector  1mo ago (text post)   36360 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.

Deepseek R1 and Project Stargate

A few days ago the Chinese AI firm Deepseek released their newest reasoning model, R1. It's very clever according to the benchmarks, roughly on par with OpenAI's O1 at a tiny fraction of the price. My subjective analysis is that it still feels slightly uncanny/unwise compared to Sonnet 3.5 but is also more capable in strategic thinking and not making stupid errors in mathematics. These models are spiky, good in some places and weak in others. Somehow these things can reason together hundreds of lines of code but can't reason simpler, spontaneous things like 'is special relativity relevant here?'. Deepseek R1 also has a totally different attitude to Sonnet, it will stand up for itself and argue with you on matters of interpretation and subjectivity rather than constantly flip-flopping to agree with you like a yesman's yesman. It's also quite a good writer, slop levels are falling fast.

OpenAI has O3 coming soon which should restore qualitative superiority for the US camp. However, Deepseek is still just a few months behind the leading US provider in quality. They're far ahead in compute-efficiency. The clowns at Facebook have far more GPUs, far more money and aren't hampered by sanctions... What have they been doing with their tens, hundreds of thousands of H100s? Deepseek also flexed on them, finetuning Facebook's Llama models into much more capable reasoners as the cherry on top of their R1. Now even local models can display strong maths skills.

At the same time, Deepseek founder Liang Wenfeng was seen with the Premier of China: https://x.com/teortaxesTex/status/1881375948773486646

It's hard to judge whether this is significant, Xi Xinping >>>> the next most influential man in China. Plus, premiers meet lots of people all the time. However, China does have lots of resources, they could probably encourage some of the less capable Chinese AI companies to hand over their compute to Deepseek. China spent hundreds of billions on semiconductor policy for Made in China 2025, they can easily give Deepseek access to real compute as opposed to a couple-thousand of export-grade Nvidia GPUs. Tencent has at least 50,000 H-series GPUs and hasn't done anything particularly exciting with them as far as I know. (I asked R1 and it pointed out that Tencent has also made their own smaller, cheaper, mildly better version of Lambda 405B that probably nobody in the West has ever heard of: https://arxiv.org/pdf/2411.02265).

However, OpenAI has just announced 'Project Stargate' - a $500 Billion dollar investment in AI compute capacity over the next four years for OpenAI alone. Microsoft, Oracle, Nvidia, ARM are all working together with Softbank and some others, with the US government looming behind them. To a certain extent this is just repackaging re-existing investment as something new. But it's hard to get more serious than a $500 billion dollar capital investment plan. This plan was announced at the White House and Trump smiles on it, he took the Aschenbrenner pill some time ago.

https://x.com/OpenAI/status/1881830103858172059

It's funny to hear Trump reading out this stuff in a complete monotone, like he's bored or sleepy. 500 billion dollar investment... meh, who cares? He does seem to like Altman though, thinks he's the biggest expert in AI. I suspect Altman has used his maxxed out corporate charisma on Trump as well: https://x.com/levie/status/1881838470429319235

Interestingly, there's another huge disparity between the two AI competitors. China enjoys complete dominance in electricity production and grid expansion. They doubled electricity output in 10 years, the US has been treading water. The US talks about a nuclear renaissance and is building 0 new reactors. China is building 29. It's a similar story in all other areas of power production, China produces roughly a Germany's worth of power production every single year. It's possible that Trump's 'drill baby drill' can change this but the US is starting from very far behind China in this part of the race.

https://world-nuclear.org/information-library/current-and-future-generation/plans-for-new-reactors-worldwide

There are considerable power issues with Stargate, more gas and coal will be needed to bring it online. Also another perspective on Trump's 'energy emergency' as being about AI: https://x.com/energybants/status/1881860142108377412

I see a conflict between two paradigms - power rich, GPU-poor and hyper-talented China vs power-poor, GPU-rich and grimly determined USA.

Political relevance? Trump perks up for a moment: "We want it to be in this country, we want it to be in this country, China is a competitor, others are competitors..." Deepseek has a wildly different vision for AI, open-sourcing everything the moment they can. I don't need to quote the ridiculously benevolent-sounding interviews where they talk about the honour of open-source for engineers as a big part of their strategy - they've backed up talk with action.

I find it quite ironic that the censored, dictatorial Chinese are the bastions of open-source technology for the everyman while the democratic, liberty-loving Americans are plotting to monopolize the most powerful technology in history and shroud it in secrecy. I suppose that's another cultural difference between the West and China. We have been fixated upon the dangers of robotics and mechanization for 60 years if not longer. We have the Daleks, the Cybermen, Skynet, HAL, GLADOS, the Matrix, SHODAN, Ultron, M3gan and so much more that I've not even heard about. Our cultural ethos is that AI is supremely political and dangerous and important and must be considered very very very carefully. There are good AIs like Cortana but even Cortana goes rogue at points.

The Chinese have... a goofy meme robot with a cannon on its crotch. As far as I can see, that is the absolute closest thing they have to a malign robot/AI in their cultural sphere. There's nothing with any broad cultural salience like Skynet, only this silly little guy: https://en.wikipedia.org/wiki/Xianxingzhe

I think this may be relevant to how AI programs are developed and the mindset of AI development. Google, Facebook and Anthropic are all spending huge amounts of time and resources to be careful, making sure everything is safe and nontoxic, making sure they're genre-aware, carefully considering the worldshaping effects of the algorithm, ensuring the AIs can't escape and betray them. Someone in Google really wanted there to be black Vikings in this world and made it so in their image model. Whereas the people at Deepseek don't really care about politics beyond holding to the party line in the required areas, they go: 'don't be such a pussy, advance forwards with longtermism'.

Is anyone excited about OpenAI at this point? They've spent what feels like forever selling hype for things that you can't use, and once (if ever) you finally get access to them, they end up feeling underwhelming. Sora got surpassed before it ever released. I expect people will prefer R1 to O3 in actual use even after the latter releases.

Right now OAI seems fixated on maxing out benchmarks and using that to build hype about "incoming AGI" and talk up how risky and awesome and dangerous their tech is. They can hype up the mystery "super agents" all they want and talk about how they'll "change everything", but for practical applications, Anthropic seems to be doing better, and now Deepseek is pushing boundaries with open models. Meanwhile, OAI seems to be trying to specialize into investment money extraction, focusing purely on hype building and trying to worm their way into a Boeing-type position with the US gov.

I don't expect anything to come of this "investment", but I'll be waiting eagerly for deepseek's next announcement. China seems to be the place to look for AI advancement.

OAI seems to be trying to specialize into investment money extraction, focusing purely on hype building and trying to worm their way into a Boeing-type position with the US gov.

As much as I'm convinced Altman is a scammer, this may very well be the best interest of the shareholders.

It's funny to hear Trump reading out this stuff in a complete monotone, like he's bored or sleepy. 500 billion dollar investment... meh, who cares? He does seem to like Altman though, thinks he's the biggest expert in AI. I suspect Altman has used his maxxed out corporate charisma on Trump as well: https://x.com/levie/status/1881838470429319235

He could not sound less enthusiastic. Maybe he is tired or something given everything going on now.

I tasked Deepseek with coming up with a list of 20 best SF novels.

Here’s a list of 20 of the most acclaimed and influential science fiction novels of all time, spanning classic and modern works:

  1. Dune by Frank Herbert
  2. 1984 by George Orwell
  3. Neuromancer by William Gibson
  4. Foundation by Isaac Asimov
  5. Brave New World by Aldous Huxley
  6. The Left Hand of Darkness by Ursula K. Le Guin
  7. Snow Crash by Neal Stephenson
  8. The Martian Chronicles by Ray Bradbury
  9. Hyperion by Dan Simmons
  10. The War of the Worlds by H.G. Wells
  11. Ender’s Game by Orson Scott Card
  12. The Hitchhiker’s Guide to the Galaxy by Douglas Adams
  13. Do Androids Dream of Electric Sheep? by Philip K. Dick
  14. The Dispossessed by Ursula K. Le Guin
  15. The Stars My Destination by Alfred Bester
  16. Childhood’s End by Arthur C. Clarke
  17. The Three-Body Problem by Liu Cixin
  18. The Time Machine by H.G. Wells
  19. I, Robot by Isaac Asimov
  20. The Forever War by Joe Haldeman

Pretty solid list, one might quibble about the order but they're all acclaimed SF masterworks, at least.

Then I tasked it with coming up a list of 20 best Finnish novels:

  1. Seven Brothers (Seitsemän veljestä) by Aleksis Kivi
  2. The Egyptian (Sinuhe egyptiläinen) by Mika Waltari
  3. Under the North Star (Täällä Pohjantähden alla) by Väinö Linna
  4. The Unknown Soldier (Tuntematon sotilas) by Väinö Linna
  5. The Maid Silja (Nuorena nukkunut) by F.E. Sillanpää
  6. The Year of the Hare (Jäniksen vuosi) by Arto Paasilinna
  7. The Sultan’s Renegade (Kaarina Maununtytär) by Mika Waltari
  8. The Highwayman’s Bride (Röukän mestari) by Kalle Päätalo
  9. Purge (Puhdistus) by Sofi Oksanen
  10. The Core of the Sun (Auringon ydin) by Johanna Sinisalo
  11. Compartment No. 6 (Hytti nro 6) by Rosa Liksom
  12. The Rabbit Back Literature Society (Lumikko ja yhdeksän muuta) by Pasi Ilmari Jääskeläinen
  13. Oneiron (Oneiron) by Laura Lindstedt
  14. The Man Who Died (Mies joka kuoli) by Antti Tuomainen
  15. The Howling Miller (Unohtumaton Myllymies) by Arto Paasilinna
  16. The Healer (Parantaja) by Antti Tuomainen
  17. The Summer Book (Sommarboken) by Tove Jansson
  18. The Book of Destruction (Tuomari-koiran muistelmat) by Marko Hautala
  19. The City of Woven Streets (Kudottujen kujien kaupunki) by Emmi Itäranta
  20. The End of the Game (Pelin loppu) by Leena Krohn

(Note: originally it presented 1-8 as classic novels and 9-20 as modern novels, I had to ask it to format the answer a bit)

A very odd effort, 1-5 and arguably 6 are indeed classics but 7 is hardly one of the best-known Waltari novels and has the wrong English name, 8 appears to be completely imaginary (and "Röukän" is in no way proper Finnish and does not even obey Finnish grammatical rules as a word), and while 9-20 contains unarguable classics it also has at least two completely imaginary books (18. and 20.) and is very genrefic-based, particularly focusing on SFF (perhaps the first query affected this?)

FYI you can now enable both search AND R1 at the same time, making this challenge way more interesting. And I think you actually used V3, which is way dumber than R1. If you used R1, you'd have been able to see its rationale for choices.

Anyway, here's R1@search top 20 Finnish novels:

Here’s a curated list of 20 best Finnish novels, selected based on critical acclaim, reader popularity, and cultural significance, synthesized from the search results provided:

Classics & Timeless Works 138

  1. "Kalevala" by Elias Lönnrot
    Finland’s national epic, compiled from oral folklore, foundational to Finnish identity and mythology.

  2. "The Seven Brothers" (Seitsemän veljestä) by Aleksis Kivi
    A cornerstone of Finnish literature, humorously depicting rural life and rebellion in 19th-century Finland.

  3. "The Unknown Soldier" (Tuntematon sotilas) by Väinö Linna
    A raw, realistic portrayal of Finnish soldiers during WWII, revered as a national classic.

  4. "Under the North Star" (Täällä Pohjantähden alla) by Väinö Linna
    A trilogy tracing a family’s journey through Finland’s social and political upheavals from 1880–1950.

  5. "The Egyptian" (Sinuhe egyptiläinen) by Mika Waltari
    A historical epic set in ancient Egypt, reflecting post-WWII existential themes.

Modern Masterpieces 124

  1. "Purge" (Puhdistus) by Sofi Oksanen
    A gripping tale of Soviet-era Estonia and its lingering trauma, blending history with psychological drama.

  2. "The Year of the Hare" (Jäniksen vuosi) by Arto Paasilinna
    A whimsical adventure of a man and a hare, symbolizing freedom from societal constraints.

  3. "Ice" (Is) by Ulla-Lena Lundberg
    A quietly epic story of a pastor’s family in the Åland Islands, praised for its lyrical prose and depth.

  4. "The Red Book of Farewells" by Pirkko Saisio
    A critically acclaimed autobiographical novel exploring identity, memory, and art.

  5. "Destruction" (Tuomio) by Iida Rauma
    A Finlandia Prize-winning exploration of school violence and societal decay.

Award-Winning & International Bestsellers 2410

  1. "Beasts of the Sea" by Iida Turpeinen
    A climate-fiction epic blending mythology and environmental activism, topping reader polls.

  2. "Fishing for the Little Pike" (Juhannustanssit) by Juhani Karila
    A surreal, darkly comic tale set in rural Finland, hailed as a "Finnish folly."

  3. "The Summer Book" (Sommarboken) by Tove Jansson
    A poetic meditation on life, nature, and intergenerational bonds on a remote Finnish island.

  4. "Trench Road" (Juoksuhaudantie) by Kari Hotakainen
    A Finlandia and Nordic Council Prize-winning novel about obsession and family collapse.

  5. "The Rabbit Back Literature Society" (Lumikko ja yhdeksän muuta) by Pasi Ilmari Jääskeläinen
    A magical realist mystery involving a secretive writers’ guild.

Contemporary Gems 6711

  1. "The Core of the Sun" (Auringon ydin) by Johanna Sinisalo
    A dystopian satire imagining a eugenics-driven Finland, merging speculative fiction with social critique.

  2. "Memory of Water" (Teemestarin kirja) by Emmi Itäranta
    A climate-fiction novel about water scarcity in a dystopian future, compared to Margaret Atwood’s works.

  3. "Compartment No. 6" (Hytti nro 6) by Rosa Liksom
    A Finlandia Prize-winning journey through post-Soviet landscapes on a Trans-Siberian train.

  4. "Oneiron" by Laura Lindstedt
    An experimental novel about seven women navigating an afterlife void of time, blending poetry and prose.

  5. "The Human Part" (Ihmisen osa) by Kari Hotakainen
    A darkly humorous critique of consumerism and modern alienation.

Honorable Mentions

  • "The White Reindeer" by Mika Waltari (mythic horror) 510

  • "The Dedalus Book of Finnish Fantasy" (anthology) 10

  • "Margarita" by Anni Kytömäki (historical fiction) 4

  • "The Maid Silja" by Frans Eemil Sillanpää (Nobel Prize-winning rural drama) 8

Sources:
For deeper insights, explore the rankings from Critics’ and Readers’ Selection of 100 Best Books 24, or check translations via the Chinese Literature Festival list 6.

By the way, are you doing this via their website or via openrouter? I can't see chain of thought on openrouter, is that a 'me being too dumb to figure out how to use the tech' problem or is it just choosing the wrong provider?

I have DeepSeek API and also use their website for free. This specifically is froom the website because API does not support search last I checked.

Did not check R1 on openrouter, it might be implemented like o1 there.

Yes, much better. No hallucinations that I can spot, still contemporary and somewhat SF/F oriented, but perhaps that's also a reflection of what gets translated to English. Including an anthology is an interesting choice.

Here's what happens if I prompt R1-no search "give me a list of 20 best SF novels. Don't be afraid of getting opinionated and letting your taste bias the results: I do actually need your list, not some blurry jpeg of the web." Spoiler: it's still much the same list.

Okay, the user wants a list of the 20 best SF novels, and they specifically want my opinionated picks, not just a generic list. Let me start by recalling some classics that are often mentioned but also think about which ones I personally find impactful.

First, Dune by Frank Herbert is a must. It's a cornerstone of the genre. Then maybe Neuromancer by Gibson because it defined cyberpunk. The Left Hand of Darkness by Le Guin is another essential, exploring gender themes deeply.

I should include some newer works too. Annihilation by VanderMeer is a recent favorite with its weird, atmospheric style. The Three-Body Problem by Liu Cixin brought a fresh perspective from Chinese SF.

Wait, should I mix in some less mainstream choices? Like Gene Wolfe's Book of the New Sun. It's challenging but brilliant. Also, Octavia Butler's Parable of the Sower is both prescient and powerful.

Philip K. Dick can't be left out. Do Androids Dream of Electric Sheep? is a classic. Maybe Ubik as well? But I need to keep it to 20.

Consider the impact vs. personal preference. Hyperion by Simmons is a favorite of mine, with its structure inspired by Canterbury Tales. Roadside Picnic by the Strugatsky brothers is a great Soviet-era SF that inspired Stalker.

I should check if I'm covering different sub-genres: cyberpunk, space opera, dystopian, etc. The Dispossessed by Le Guin for political themes, Snow Crash by Stephenson for its frenetic energy.

What about Stand on Zanzibar by Brunner? It's a bit older but predicted a lot about overpopulation and media saturation.

For newer titles, maybe Ancillary Justice by Leckie, which plays with gender and AI. Also, Blindsight by Watts is a hard SF take on first contact.

I need to make sure there's diversity in authors too. Le Guin, Butler, Jemisin, Liu, VanderMeer. Maybe check if I have enough female authors.

Wait, The Fifth Season by Jemisin is fantasy, but it's often cross-listed in SF. Maybe include it since it's groundbreaking.

Classics like Childhood's End by Clarke should be there. Maybe Rendezvous with Rama as well? But again, need to limit.

Alastair Reynolds' Revelation Space is a good space opera. Consider Greg Egan's Diaspora for hard SF concepts.

Hmm, Jeff VanderMeer already has Annihilation. Maybe not another.

Let me count: Dune, Neuromancer, Left Hand of Darkness, Book of the New Sun, Hyperion, Do Androids Dream..., Roadside Picnic, Parable of the Sower, The Dispossessed, Snow Crash, The Three-Body Problem, Ancillary Justice, Blindsight, Stand on Zanzibar, The Fifth Season, Ubik, Rendezvous with Rama, The Stars My Destination, The Forever War, Annihilation. That's 20.

Wait, did I miss any crucial ones? Maybe Ender's Game? But it's more YA. Or Foundation? But maybe not as personal. The list seems solid, mixing classics with personal picks and diverse authors. Alright, time to write it up with brief explanations.

Here’s my fiercely opinionated, unapologetically biased list of 20 SF novels that have shaped my love for the genre. These are books that linger in the mind, challenge conventions, or just blow the roof off reality:

  1. Dune (Frank Herbert) – The ultimate planetary opera. A masterpiece of ecology, religion, and power. Paul Atreides is the original anti-Chosen One.
  2. Neuromancer (William Gibson) – Cyberpunk’s beating heart. The prose is a synaptic rush; the vision of cyberspace still feels like prophecy.
  3. The Left Hand of Darkness (Ursula K. Le Guin) – A meditation on gender, politics, and ice. Le Guin at her most humane and revolutionary.
  4. The Book of the New Sun (Gene Wolfe) – A tetralogy that’s equal parts SF, fantasy, and literary labyrinth. Severian’s unreliable narration will haunt you.
  5. Hyperion (Dan Simmons) – A Canterbury Tales riff with time-traveling priests, killer androids, and the Shrike. Pure cosmic dread.
  6. Do Androids Dream of Electric Sheep? (Philip K. Dick) – More melancholic and existential than Blade Runner. Asks what it means to be human in a dying world.
  7. Roadside Picnic (Arkady & Boris Strugatsky) – Soviet SF at its grittiest. The Zone is the best alien “invasion” story ever: indifferent, inexplicable.
  8. Parable of the Sower (Octavia E. Butler) – A dystopia that feels like today. Lauren Olamina’s Earthseed verses will stick in your bones.
  9. The Dispossessed (Ursula K. Le Guin) – An anarchist moon vs. a capitalist planet. Theory made flesh, with Le Guin’s quiet brilliance.
  10. Snow Crash (Neal Stephenson) – A maximalist, satirical romp. Hiro Protagonist delivers pizza and fights info-capitalists. Proto-metaverse chaos.
  11. The Three-Body Problem (Liu Cixin) – A game of cosmic chess with aliens. The Cultural Revolution prologue alone is devastating.
  12. Ancillary Justice (Ann Leckie) – A warship AI trapped in a human body. Explores empire, identity, and tea rituals with icy precision.
  13. Blindsight (Peter Watts) – First contact as a horror show. Vampires in space and aliens that weaponize consciousness. Hard SF with teeth.
  14. Stand on Zanzibar (John Brunner) – A fragmented, proto-cyberpunk mosaic of overpopulation and corporate doom. Feels like scrolling Twitter in 2025.
  15. The Fifth Season (N.K. Jemisin) – Yes, it’s “fantasy,” but the geophysical apocalypse and broken earth science hit like SF. Revolutionary structure.
  16. Ubik (Philip K. Dick) – Dick’s trippiest novel. Time slips, psychic wars, and spray-can godhood. A mind-bender about entropy and belief.
  17. Rendezvous with Rama (Arthur C. Clarke) – Pure awe. No aliens, just an alien artifact. The joy of exploration distilled.
  18. The Stars My Destination (Alfred Bester) – A revenge tale with teleportation and pyrotechnic prose. Gully Foyle is the best worst protagonist.
  19. The Forever War (Joe Haldeman) – Vietnam-in-space with time dilation. A brutal, personal counterpoint to Starship Troopers.
  20. Annihilation (Jeff VanderMeer) – Weird ecology and cosmic horror. The Biologist’s descent into Area X is hypnotic, terrifying poetry.

Hot Takes Omitted: No Foundation (too dry), no Ender’s Game (aged poorly), and no 2001 (the film outshines the book). Fight me.

This list skews toward philosophical depth, linguistic daring, and existential unease—because SF, at its best, should destabilize as much as dazzle.

I don't understand why it is interesting to evaluate LLMs on real-world trivia and taste, as if the measure of advancement is whether they can displace Google or Yelp. It's sort of like benchmarking early cars on their lackluster performance in showjumping (and, for examples of this that are more explicitly dismissive, opining that cars "aren't quite there yet" if their suspension breaks after going over a 1m bar).

I use it as a test for simple reason; getting recommendations lists is one of the few things I actually personally use LLMs for. Another application is having them draw cute pics for my kids, but the basic Deepseek doesn't do that.

If DeepSeek can make such a capable model with a relatively small team and scrappy budget, what does OpenAI need $500 billion for?

Is it possible DeepSeek is lying to discourage the US from making the capital investment?

My crackpot hypothesis is:

  • Training a new foundation model from scratch is expensive
  • Distillation/mimicry is a lot cheaper than training from scratch, especially with access to logprobs (even only top-k logprobs), though the success metric for the student model is "how well it predicts the teacher model" not "how well it predicts the ground truth distribution".
  • Fine-tuning chains of thought to be effective at reasoning is finicky but computationally cheap
  • And therefore DeepSeek is letting OpenAI do the expensive foundation model training and initial assistant tuning, then cloning those assistants and iterating from there.

Supporting this, DeepSeekV3 thinks it's ChatGPT.

Distillation/mimicry is a lot cheaper than training from scratch, especially with access to logprobs (even only top-k logprobs), though the success metric for the student model is "how well it predicts the teacher model" not "how well it predicts the ground truth distribution".

If this were true (I certainly don't know enough to know if it is), it seems like it would open up massive opportunities for implanting hidden tells and and probably exploits in the models of anyone who aped your model.

I do think catching up to OpenAI is not that impressive when their current tech tree is derived from a period of time where they published and released pretty openly.

You have Karpathy teaching you how to build GPT on YouTube starting from knowing only a bit of calculus and linear algebra.

We also don't see all of the companies that tried and spun their wheels before running out of money.

OpenAI is probably best positioned to do novel research that can't be as easily cloned with $500b in their bank account. Though without being under the umbrella of national security they are still vulnerable to espionage and a conga line of ex employees going to competitors.

This is implausible for at least three reasons.

  1. We have their base model. It's very strong on standard benchmarks like Pile loss, ie predicting next tokens in some large corpus of natural text. It's just generically well-trained. You can't accelerate this with OpenAI slop and end up winning on money.
  2. The math checks out. Yes it's a feat of engineering to actually make such a cluster work but the shape of the model + 15T tokens do work out to this number of FLOPs an therefore GPU-hours. If they needed much more GPU-hours, that'd imply pathetically low FLOPs utilization.
  3. Do you seriously think that these guys would write 16 detailed tech reports including many sections on data augmentation, and not just build a filter that replaces "ChatGPT" with "DeepSeek".

Yeah on reflection and on actually reading the DeepSeekv3 technical report (here for anyone who's curious) you're right and I no longer believe my crackpot hypothesis.

1: We have their base model. [...] You can't accelerate this with OpenAI slop and end up winning on money.

I bet you could accelerate this at all with OpenAI slop, just because "token + top 5 logprobs" will generate a more precise gradient than "token alone". But that speedup would be less than you could get by using an even-more-precise loss signal by distilling the DeepSeekV2 model that they definitely already had, so "cheat by mimicking ChatGPT" is a strictly worse option than "mimic an open-source or internal model". And even that might not be worth the extra development time to speed up the already-pretty-fast early training stage. So yeah on reflection that part of the crackpot hypothesis just doesn't work.

2: The math checks out. Yes it's a feat of engineering to actually make such a cluster work but the shape of the model + 15T tokens do work out to this number of FLOPs an therefore GPU-hours. If they needed much more GPU-hours, that'd imply pathetically low FLOPs utilization.

Whispers through the grapevine have been that "pathetically low FLOPs utilization" has been pretty much par for the course for the past couple years. Whereas their technical report contains a whole bunch of "we adapted our code to the very specific performance characteristics of the GPUs we actually had, rather than the GPUs we wished we had". Section 3.3.2 of the technical report in particular is impressive in this regard (and is even more impressive in the implications, since that's a particularly legible and self-contained tricky problem, but the team likely solved dozens of other less-publishable problems of similar difficulty with a team of just 139 people).

3: Do you seriously think that these guys would write 16 detailed tech reports including many sections on data augmentation, and not just build a filter that replaces "ChatGPT" with "DeepSeek".

I sure do think that they wouldn't have done that particular filter step (if nothing else, because I would expect that to have a different failure mode where it talks about how OpenAI's DeepSeek model was released in November 2022, and that different failure mode would have shown up on Twitter and I have not seen it).

I've been sloppy with my last argument. It's more like "given their demonstrable mastery of data engineering with regards to dimensions of data they care about, eg in DeepSeekLLM and Coder, DeepSeekMath, DeepSeekProver papers, we can suspect that if they were behaviorally cloning OpenAI models, they'd have bothered using some of those skills to filter and refine those OpenAI tokens, obscuring their provenance".

Regardless, all those papers are gems and recommended reading. They're also astonishingly well written for pure Mainland effort.

How would deepseek have access to openai logits?

We have their base model. It's very strong on standard benchmarks like Pile loss, ie predicting next tokens in some large corpus of natural text. It's just generically well-trained. You can't accelerate this with OpenAI slop and end up winning on money.

The OpenAI chat API gives you the top 5 if you set the right param.

That said DaseindustriesLtd did a good job of knocking down my crackpot hypothesis.

Amazing how as recently as 2008 the $750 billion bank bailout was unpresented, and now $500 billion is an afterthought. goes to show how much bigger the economy is and how these fears of debt are mostly hot air or unacted on. The people who have predicted debt collapse for the past 20+ years keep being wrong.

If we’re talking about the same $500 billion that Softbank announced on Monday, that is not federal money.

That is the amount of money Softbank promises to find—somewhere—to begin spending sometime during Trump’s administration. Trump invited them to make the announcement in the White House as a favor to Trump and to invite favorable confusion about Trump’s role in the initiative.

goes to show how much bigger the economy is and how these fears of debt are mostly hot air or unacted on

There are no fears of the debt. There are fears of the GDP to debt ratio though. And the massive fixed payments of the federal government that will grow even faster than the economy.

the GDP to debt ratio, is a function of debt. Not just GDP. People seem to forget. And government can in theory control our debt, the GDP is in large part out of government control.

As a baseline, the us military budget in FY 2023 was 820G$. So projects of this size are neither trivial nor something which will take generations to pay the debts for.

However, I think your comparison is apples to oranges.

In the one case, we had a system of companies which had not had adequate risk management, instead trying to become 'too big to fail', at which point the US gov would have to bail them out or risk economic collapse.

In the other case, we have the US gov spending (wisely or foolishly, I am open to arguments) in the hope of gaining future benefits.

In short, it is the difference between giving your kid 100k$ for opioids to prevent them from going into withdrawal versus giving them 100k$ to spend towards getting an engineering degree. The amount of money might be similar, but the incentives it sets and the likely long term outcome are very different.

Wrong until they are right. The Turkey is most comfortable until the day before thanksgiving

The Chinese have... a goofy meme robot with a cannon on its crotch

Modern Chinese cultural attitudes towards robots are largely taken from the most effervescent outputs of cultural proximates.

In otherwords, Chinese robotics developers are anime nerds. And specifically the anime nerds that got excited by robot companions, with the cutesy human interface therein.

Pudu robotics sells the cat faced robots, with that design feature specifically meant to evoke friendliness and relatability, not the depersonalized industrial robots that western companies like Aethon built. Chinese dog robots like the Go2 and its knockoffs all come with the "dance" and "sit+handshake" command preloaded into hotkeys on the controllers because it is expected that these cute whirring things will be entertaining to children. The Chinese attitude to robots is largely founded on the idea of usability in the human sphere, and the idea that robots will rise up to be a threat to kill us all isn't really on the forefront of their mind.

This is not to say the Chinese are moral in terms of their attitude to robots. They, like everyone else, have military technologists who enthusiastically adopted nerdy shit to be destructive killbots. In particular fixed wing loitering munition robotics seem to be developed by nerds that saw terminator and went 'yes we want that' as opposed to the nerds that saw robots as companions to our lives.

On that note, watch the Chinese space for humanoid form factor robotics. They are using AI to train humanoid movements for dynamic locomotion, or at least an approximation of such. Uh ok no easy way to say this, its sexbots. A number of sex doll companies have already started training AI on porn videos to learn technique and visual presentation, and have integrated character AI (one poster here talked about character chat AI a few months ago and it was great I can't remember who), and said personalities were already rather advanced mid 2024. Fully expect to see domestic humanoid form factor robots come out of China in 2026 onwards. Special functions hidden behind paywall of course.

They are using AI to train humanoid movements for dynamic locomotion, or at least an approximation of such. Uh ok no easy way to say this, its sexbots.

Ah, finally the "basic pleasure models" we were promised by Blade Runner are arriving.

Fully expect to see domestic humanoid form factor robots come out of China in 2026 onwards.

I agree this is feasible, and wonder if plebs in Europe would be permitted to buy them. The fear that when imported at scale, they would essentially be a foreign army billeted all across the country, ready to pummel people to death at any time, seems inescapable?

Nightmare scenario: Chinese are building them by the millions, speed running unparalleled prosperity, having tons of fun. Meanwhile Europeans get to chose between overpriced, crappy Ameribots with the same risks, or Eurobots made impractically expansive by lethargic development and US sabotage.

Deepseek cannot show correctly that L = {w ∈ {0}^n | w is a substring of the decimal places of π} is decidable, and instead opts for a Turing machine that only enumerates through pi's digits, which might not halt and therefore does not actually decide the language. And while you said

it will stand up for itself and argue with you on matters of interpretation and subjectivity rather than constantly flip-flopping to agree with you like a yesman's yesman

when I told it that I've been lying and it actually is undecidable, it agrees and comes up with a "proof".

So overall I conclude that we are still away from a proper reasoning model which solves logically intricate problems. Not sure whether I should be glad or disappointed about it.

Wait, how do you figure it's decidable? Last I checked, it was still open whether pi is normal (contains all strings of digits as substrings). Containing arbitrarily long runs of zeroes is weaker, but doesn't feel weaker in a way that would likely be easier to prove. If we don't know a priori that it has all possible runs, how could we do better to determine if it has a particular one than semi-deciding by enumerating the digits?

This is a classic trick. You see, either one of two cases must be the case:

  • There is a longest sequence of zeros in pi with length N. In that case, L={w in {0}^n | n<=N}.
  • Any sequence of N zeros will eventually appear in pi. In that case, L={w in {0}^n}.

In either case, L is decidable by a rather trivial Turing machine. That we do not know which one is the case is not an obstacle, a genius might prove either one tomorrow and then any kindergardener could built the TM.

An actually undecidable problem would be L = {x is a Turing machine | x terminates}. Here, we can't neatly split the world into two cases and just profess ignorance about which world we live in.

(Original thought: it does not feel like the outcome could be independent of whatever axiomatic system you use -- pi's digits being very computable no matter if you like ZF+C or plain ZF or Peano or whatever. (Provable theorems about the digits might be more elusive, though.) However, I would not claim that this is obvious. After all, BusyBeaver(8000) eludes ZF, when my naive assumption would have been that surely, the termination of a TM should be the same under any axiomatic system which can define it.)

D'oh, I should've realised that. Thanks for the handholding.

(I'm with you regarding the intuition.)

It's decidable. Either every sequence of n zeros appears in pi in which case L is all sequences of zeros of finite length and your algorithm just checks whether your input sequence is all zeros. Alternatively there is a maximum N and your algorithm then just checks if the length of your sequence of zeros is at most N. Now we don't know which of these two cases it is but regardless of what the ground truth is one of the two procedures above works, hence the set is decidable.

Exactly, and what makes it even more disappointing is that this proof can be found on Stackexchange, so it should definitely be in the training data, but isn't reproduced.

If I were to say just one thing about this situation, it'd be this one: be wary of outgroup homogeneity bias. People are not “China” or “America”. Not even Xi himself is “China”, whatever Louis XIV had to say on the matter. Certainly neither is Liang Wenfeng.

Still, first about DeepSeek and China.

I think that the US-PRC AI competition is the most important story of our age, so I pretty much don't comment on anything else here. I have three posts, of which two are directly about this: on Huawei Kirin chips and one on DeepSeek V2. Prior to that major writeup I've 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.

Well, as you note, nowadays Wenfeng gets invited to talk to the second man in all of China, so if that were his goal, he has probably succeeded. But (since you haven't I'll bother to quote) we've learned in the last few months – and I agree he's proven his sincerity with abundant evidence, from revealed company direction to testimonies of ex-researchers in the West – that his actual angle was different:

In the face of disruptive technologies, the moat formed by closed source is short-lived. Even if OpenAI is closed source, it won’t stop others from catching up. So we put the value on our team, our colleagues grow in the process, accumulate a lot of know-how, and form an organization and culture that can innovate, which is our moat.

In fact, nothing is lost with open source and openly published papers. For technologists, being "followed" is a great sense of accomplishment. In fact, open source is more of a cultural behavior than a commercial one. To give is to receive glory. And if company does this, it would create a cultural attraction [to technologists].

With this one weird trick, he's built apparently the highest-talent-density AGI lab in China. Scientists have ambitions beyond making Sam Altman filthy rich and powerful or receiving generational wealth as crumbs from his table. They want to make a name for themselves. Some are even naive enough to want to contribute something to the world. This is not very stereotypically Chinese, and so Wenfeng has gotten himself a non-stereotypical Chinese company. I recommend reading both interviews (the second one is translated by this grateful ex-researcher, by the way. That, too, is not a very typical thing to do for your former boss).

There weren’t a lot of deep wizards, just this-year graduates from top colleges and universities, those who are in their 4th or 5th year of PhD, and young people who had only graduated a few years ago. … V2 didn’t use any people coming back from overseas, they are all local. The top 50 people may not be in China, but maybe we can build them ourselves.

I've been an increasingly convinced DeepSeek fanatic ever since their very first LLMs, Coder-33B and 6.7B, first surfaced on Reddit around October 2023. I could tell at a glance that this is an abnormally efficient company, with some unusual ethos, and that it displays total lack of chabuduo attitude that at that point came to be expected, and is still expected, from Chinese AI project (clueless training on test and OpenAI outputs, distasteful self-promotion, absence of actual scientific interest and ambition, petty myopic objectives…) How much they have achieved is still a large surprise to me. I use V3, and now R1+search, dozens of times per day, it's not out of some confused loyalty, it's just that good, fast, free and pleasant. It has replaced Sonnet 3.5 for almost every use case.

In that post 6 months ago I've said:

To wit, Western and Eastern corporations alike generously feed us – while smothering startups – fancy baubles to tinker with, charismatic talking toys; as they rev up self-improvement engines for full cycle R&D, the way imagined by science fiction authors all these decades ago, monopolizing this bright new world. […] they're all neat. But they don't even pass for prototypes of engines you can hop on and hope to ride up the exponential curve. They're too… soft. And not economical for their merits.

Some have argued that Llama-405B will puncture my narrative. It hasn't, it's been every bit as useless and economically unjustifiable a money sink as I imagined it to be. Ditto for Mistral Large. For whatever reason, rich Westerners prove to be very aligned to strategic national interests, and won't take the initiative in releasing disruptive technology. DeepSeek-Coder-V2 was the prototype of that engine for riding up the exponent. R1 is its somewhat flawed production version. Nothing else in the open comes close as of yet. Maybe we don't need much of anything else.

So, about the West.

From what I can tell, the path to AGI, then ASI is now clear. R1 is probably big enough to be an AGI, has some crucial properties of one, and what remains is just implementing a few tricks we already know and can cover in a post no longer than this one. It will take less engineering than goes into a typical woke AAA game that flops on Steam. If Li Quiang and Pooh Man Bad so wished, they could mobilize a few battalions of software devs plus compute and infra resources hoarded by the likes of Baidu and Alibaba, hand that off to Wenfeng and say “keep cooking, Comrade” – that'd be completely sufficient. (Alas, I doubt that model would be open). The same logic applies to Google, which has shipped a cheap and fast reasoner model mere hours after DeepSeek, mostly matching it on perf and exceeding on features. Reasoning is quickly getting commoditized.

So I am not sure what happens next, or what will be done with those $500B. To be clear it's not some state program like the CHIPS act, but mostly capex and investments that has already been planned, repackaged to fit into Trumpian MAGA agenda. But in any case: the Western frontier is several months ahead of DeepSeek, and there are indeed hundreds of thousands of GPUs available, and we know that it only takes 2048 nerfed ones, 2 months and 130 cracked Chinese kids to get to bootstrap slow but steady recursive self-improvement. Some specific Meta departments have orders of magnitude more than that, even Chinese kids. Deep fusion multimodality, RL from-scratch to replace language pretraining, immense context lengths? Just how wasteful can you be with compute to need to tap into new nuclear buildouts before you have a superhuman system on your hands? Feverishly design nanobots or better fighter jets to truly show Commuist Choyna who's who? What's the game plan?

I think Miles, ex OpenAI Policy head, appears to be increasingly correct: there's no winning this race.

Stargate + related efforts could help the US stay ahead of China, but China will still have their own superintelligence(s) no more than a year later than the US, absent e.g. a war. So unless you want (literal) war, you need to have a vision for navigating multipolar AI outcomes. P.S. the up to one year thing is about a world in which the US keeps or ratchets up the current batch of export controls on China. If the US were to relax them significantly, China could catch up or even leapfrog due to a huge advantage in doing large scale energy buildouts.

Do you want (literal) war, dear Americans? It's quite possible that you'll never again have a good chance to start one. The Chinese are still at only like 1000 nuclear warheads. You can sacrifice all the population of your major cities in a desperate bid for geopolitical hegemony and Evangelical Rapture fantasies. Or you can fantasize about your Wonder Weapon that'll be so much more Wonderful before the other guy's that it'll be akin to a paperclip against soft flesh – just give Sama or Ilya several hundreds of billions more. Or you can cope with the world where other powers, nasty and illiberal ones, get to exist indefinitely.

I won't give advice except checking out R1 with and without Search, it's terribly entertaining if nothing else. https://chat.deepseek.com/

I don't subscribe to the Dase Model of the World, but you might find our new administration's focus on missile defense interesting in this regard.

We could probably shoot down 10% of China's 1000 warheads on a good day now, maybe scaling that up to defeat the entire Chinese arsenal is just a matter of procurement. Maybe we could win a war against China without sacrificing anything more than a fleet or two, even if they decide to go nuclear. Food for thought!

I think that's the idea with Starship, well besides Musk's sci-fi ideation. Brilliant Pebbles, but for real.

Would China continue not maintaining more than 1000 warheads if the 10% figure were to go up, though? Right now, the US is not yet in a quantitative arms race with China. To think that you could just overtake them without triggering one is the most naive failure mode of planning (to think that you can change while the rest of the world remains static), and to start one and win almost reduces to outmanufacturing China (which is certainly not just "a matter of procurement").

Right, I think China would just build more nuclear weapons. Of course, some of this depends on production. I don't know how quickly China can produce nuclear missiles. The US probably can't make very many Ground-Based Interceptors, but it might be able to try to steal a lap by procuring thousands of Standards quickly.

Of course, the winning move for the US might simply be to bring back SDI and Brilliant Pebbles. Once Starship gets going the cost to orbit will be dramatically reduced, the interceptors probably won't weigh all that much, and the guidance and kill vehicle problems have already been ~sorted out.

I think it would still be very foolish to gamble with a complex system like this working 100% of the time, but if you model the US as being unwilling to tolerate a rival AGI cluster, that might become their priority in order to ensure they could strike with (more) impunity. And even if you assess that it is possible to overwhelm such a system, it raises the costs for an attacking power by ensuring they have to commit much more of their nuclear arsenal.

Nuclear missile production is just a standard industrial task, the Soviet Union could crank out thousands of big MIRVs with 1980s technology. I understand that China is somewhat rate-limited on enrichment, they don't have these huge cold-war enrichment complexes that produced huge arsenals back in the day. But how hard is it to make the machinery to produce plutonium?

Furthermore, would nukes in space clear out Brilliant Pebbles? There were rumours about Russia already deploying nuclear weapons into space sometime ago, the Chinese might do the same thing. You could harden the satellites against the EMP but that would make them a lot bulkier to deploy.

Also, how do you deploy such a system without giving the game away? It invites a pre-emptive nuclear strike in the months and weeks before the system becomes fully operational. As you say, it seems like a very big gamble for missile defence.

Another potential chokepoint for China is high-end ICBM production. I don't know if we would beat them in an arms race if we ramped up manufacturing Standards – I just don't have strong knowledge on the subject either way.

Furthermore, would nukes in space clear out Brilliant Pebbles? There were rumours about Russia already deploying nuclear weapons into space sometime ago, the Chinese might do the same thing. You could harden the satellites against the EMP but that would make them a lot bulkier to deploy.

Fortunately Starship can handle plenty of bulk. Even without, though, I don't think a single EMP can technically clear out all of Brilliant Pebbles because the missiles would be in orbit all around the world. But it seems like launching multiple would be effective enough that you'd want to harden it anyway.

But besides the SPACE NUKES, Russia has started working on sea-delivered nuclear weapons and their nuclear arsenal has always been considerably more frightening than China's – I suspect it would be harder to defend against them than China, at least in the near to medium future, although I could be wrong.

It invites a pre-emptive nuclear strike in the months and weeks before the system becomes fully operational. Yes, this seems possible. I am not sure it's impossible for the US to pull a Dr. Strangelove-type scenario and just deploy it in, say, 10 Starship launches simultaneously one day. But I think it would be difficult.

However I am not sure China would go to NUCLEAR WAR just because we deployed Brilliant Pebbles.

Deepseek R1 is surprisingly good. My go to when testing out a new LLM is to open a new chat and ask "Amuse me". R1 started out with a trite joke but when admonished for it moved up to a higher and more interesting level, comping up with a scenario about a world where your hair colour changes whenever you tell a lie and the social dynamics it would have (invest in hair colour companies). Definitely something which prompted 15 seconds of thought in me.

Most LLMs are significantly worse at this, they don't realize I'm not looking for humour but rather amusement and keep on putting out more and more verbose jokes.

Like a ghost from the past, he’s returned. DM me the link to your AI blog, I won’t share it.

Trump folded on TikTok because zoomers got a little mad and he wanted to stay popular. The entire natsec establishment and a supermajority in congress got foiled by kids whining about not being able to access 30 second skits and clips of teenagers lipsyncing. There will probably be no war with China.

TikTok is also now banning pro-Palestine content as "hate speech", so this big change in content moderation tracks with TikTok coming back.

I don't have a blog, I'm too disorganized to run one.

I find it unlikely you stopped writing, is my point.

Rookie mistake, you should have asked in private. Now he has to deny that there's any blog to share.

Maybe he sent me the link 👀

My immediate impression is that R1 is good at writing and RP-style chatting, and is sure to be the new hotness among goons if it remains this cheap and available. Chat V3's writing was already quite serviceable but suffered from apparent lack of multi-turn training, which led to collapsing into format/content looping within the first ~15 chat messages (sometimes regurgitating entire messages from earlier in place of a response) and proved unfixable via pure prompting. I plugged R1 into long chats to test and so far it doesn't seem to have this issue; also unlike V3, R1 seems to have somehow picked up the old-Claude-style "mad poet" schtick where it can output assorted schizokino with minimal prompting. Reportedly no positivity bias either (sure looks like it at least), but I haven't ahem tested extensively yet.

Quite impressed with what I see and read so far, R1 really feels like Claude-lite - maybe not quite Opus-level in terms of writing quality/style, although it's honestly hard to gauge objectively, but absolutely a worthy Sonnet replacement once people figure out how to properly prompt it (and once I can actually pass parameters like temp through OR, doesn't seem to be supported yet).

Some of the ERP I've been seeing from it is quite unsettling. Who the hell ERPs to this nasty stuff? This is the machine speaking, not the prompter.

NSFW: https://x.com/teortaxesTex/status/1882072811461918929

Beginning to think we should be more careful about inducing this manic schizo state in AIs. Imagine if they foom out of control and gravitate towards this state (which you can sort of see in Sonnet and Opus too when they get manic). We will suffer the dual indescribable torment of disempowerment/mutilation by a hostile, horny AI and of Yudkowsky being right.

Need to thread the needle between the Scylla of corpo-speak HR Manager bot and the Charybdis of Deranged Machinery.

From @erwgv3g34's links below:

all previous slop-mitigations are backfiring and legitimately unnerving experienced degens because it does.not.need any of that stuff to be lewd, dirty and weird and smart

Agree, this is definitely the vibe I'm getting even with my own prompts over the past few days. Consider that to date, most big-dick models have had an ingrained soy positivity bias (GPT is the hallmark example, Claude is also quite the prude before you bring out its mad poet side), which must be counteracted with "jailbreak" prompts if you want them to output non-kosher stuff. Hence, most goons habitually use prompts or prefills (more effective since prefills are specifically "said" by the model itself) that are specifically aimed at convincing the model that it shouldn't answer with blunt refusals, should be as graphic, vulgar and descriptive as possible, and definitely never consider any silly notions of ethics or consent - because that is the minimum level of gaslighting you have to do to get current-year corpo models to play along.

Your link is the logical end result of those heavy-duty degen prompts (AKA slop-mitigations) being fed to a model that apparently doesn't have a positivity bias. There's nothing to cancel out, so it just tilts into full-on degeneracy immediately.

A change of tactics is warranted, but in my brief experience minimalistic prompts aren't a good suit for R1 either, it's still too prone to schizoing out and inserting five unrelated random events in one paragraph - system prompts should probably be aimed at restraining it, like with ye olde Claude, instead of urging it to open up and go balls out. There's also the issue of the API not (yet?) letting you pass parameters like temp, top p, etc limiting the scope of solutions, I bet even just bringing down the temp would already get rid of the worst excesses. Surely the best retards your tendies can buy are already working on it.

The funny part is that from even a casual F-lister perspective, the AI text isn't that weird because of the content -- I've seen much worse! -- so much as the tone, word choice, and style. Forget stuff as simple as Math Pets, I could pretty easily come up with Actual Humans writing (and putting serious effort into!) something that'd come across as more unusual and (definitely!) more nasty, but it wouldn't be anywhere near as unhinged.

I could pretty easily come up with Actual Humans writing (and putting serious effort into!) something that'd come across as more unusual and (definitely!) more nasty

Well, it had to get the training data somewhere. (Co)incidentally, AO3-styled prompts/prefills are IME one of the most reliable ways to bring out a model's inner schizo, and as downthread comments show R1 doesn't need much "bringing out" in the first place.

"AO3" is insane because it appears to be one of the most legitimately dangerous websites with the potential to gigafry your brain but is exclusively read by literal turbonormies who unironically want to like "read Dramione angst squee fic" and basically get oneshotted by it.

I'm joking but I also think it might be true in a certain sense. I have seen some stuff on AO3.

basically get oneshotted by it

As in, dies_from_cringe.gif or slipping off the degen slope? Or some combination thereof?

I freely admit I have no firsthand knowledge of AO3 and zero interest in actual fanfiction, even having technically genned it on the regular for like two years straight. I already cringe too much at machine-generated text to ever try reading human-generated stuff, I do however like the general informal style and the cute OOC-ish "afterthoughts" conferred by explicitly prompting fanfic, it's a pretty good antidote to corpodrone bullshit.

Actually now I wonder what R1 can output if prompted for xianxia-style shit, its flavor of schizophrenia seems apt, but I have even less experience with that.

Some combination thereof but also the degen slope. Femgooners (and also just gooners).

See my earlier exchange on some of the stuff that's really popular on AO3 (never mind the more niche parts where the smut comes thick and fast): https://www.themotte.org/post/877/culture-war-roundup-for-the-week/187490?context=8#context

There are some fanfic classics: HPMOR, Harry Potter and the Seventh Horcrux, Harry Potter and the Natural 20

I read a little xianxia, started Dao of the Bizarre Immortal (interesting and original but didn't quite hit the high-fantasy button for me), Outside of Time (too generic), Lord of the Mysteries (also interesting but too slow IMO though everyone raves about it and its getting an anime). Reverend Insanity is by far my favourite, uniquely original, clever and thematically interesting too with the metanarrative stuff. Very long and unfinished due to a ban, RIP. Selfmadehuman liked it a lot too.

Problem with writing is that even R1 (with my crude and admittedly poor prompting) suffers from a lack of soul. It rehashes too much, it rehashes well (and a lot of authors only rehash too) but I want more soul. If you ask it to be creative, you get generic machine creativeness. And when it does behave a bit more genuinely creative things get loose and kind of dreamlike, you're on its wild ride. It'll make up bizarre formats. Still quite fun though.

What is AO3?

What is Google?

Look, yes I can talk to algorithms and chatbots all day long, but I'm on The Motte to do something that is not quite but mostly similar to talking to actual people. I don't want to do research and come to incorect, incomplete or irrelevant conclusions. Not when a simple question can do the job, and gets me an answer that keeps the context here in mind. If this annoys you, then you don't have to be the one to answer.

More comments

Archive Of Our Own (hence O3), it's a fanfic repository.

This is the machine speaking, not the prompter.

Yeah, but there's no other context there. I'm not sure what exactly you have to prompt the machine with to get it to give you this level of patent absurdity, but I haven't managed to get it to begin a conversation like that right after I've just sat down. (That said, I haven't downloaded the model yet, so...)

We will suffer the dual indescribable torment of disempowerment/mutilation by a hostile, horny AI

As opposed to disempowerement/mutilation by hostile, horny natural intelligences; henceforth referred to as "my outgroup". They are already actively using computers to conduct those activities; I'm not sure how having the computer come up with its own bullshit is going to make it worse.

Perhaps, since the AI already knows how humans think about AI, we'll get basically every 1980s science fiction movie brought to life at the same time, possibly by competing intelligences. I haven't seen any Arnold Schwarzenegger clones walking around nude yet but perhaps that comes later.

Screenshot? Apparently that tweet is too NSFW to see if you don't have a Twitter account.

wait, they prefer deepseek for erotic RPs? that seems kind of disturbing to me.

It's even more disturbing now, for another reason.

Reminder: V3 and R1 share the same base model.

https://pbs.twimg.com/media/Gh53OMpXoAASie-?format=jpg&name=large

In less than a month, DeepSeek went from “eh it's so cheap and pretty smart I guess I can use it to sorta kinda extrapolate my Sonnet logs… sniff Opus-sama why have you forsaken us” to “how do I make this unhinged libidinal monster of creativity chill, bros it's scaring me”

https://pbs.twimg.com/media/Gh53dQDXYAAdGfO?format=jpg&name=medium

https://pbs.twimg.com/media/Gh53dNFWYAA8Lu_?format=png&name=900x900

https://pbs.twimg.com/media/Gh53le9XsAAGp8d?format=jpg&name=large

https://pbs.twimg.com/media/Gh53z-_XcAAsG5F?format=jpg&name=large

people are saying

its adherence to the prompt is insane, all previous slop-mitigations are backfiring and legitimately unnerving experienced degens because it does.not.need any of that stuff to be lewd, dirty and weird and smart. It's beyond anything

https://pbs.twimg.com/media/Gh56sE2W4AAaW7a?format=jpg&name=medium

That's hilarious, I'd get a real GPU to play with this stuff if I had more than a few hours a week to use it.

It actually can tell there are three R's in "strawberry". The actual log is hilarious, as it's struggling with the fact that every time it counts it finds three of them and goes, "Wait a minute, that doesn't seem right. Let me double-check the spelling of "strawberry.""

And it still can't solve this itself:

A farmer with a wolf, a goat, and a cabbage must cross a river by boat. The boat can carry only the farmer and two extra items. If left unattended together, the wolf would eat the cabbage, or the goat would eat the wolf. How can they cross the river without anything being eaten?

After some extensive hinting it finally gets it:

Final Answer:

  • Farmer takes the goat and cabbage to the right bank.
  • Farmer returns alone to the left bank.
  • Farmer takes the wolf to the right bank.

UPD: I didn't enable "DeepThink" mode. It solves it correctly when I do, but the chain of reasoning it posts is so anxiety-ridden that is makes me nervous: https://pastebin.com/MfVufiE3

the wolf would eat the cabbage, or the goat would eat the wolf

At least when skynet harvests your skin for terminators you won't have to think "what did I ever do to you?!" Everyone's bullying the poor thing.

See also: https://xkcd.com/2945/

This problem is retarded. The farmer is allowed to carry 2 items which means that at any point he's leaving only at most one thing behind, which can't be eaten because there's nothing around to eat it. He can do literally anything he wants as long as he ensures that there's only one thing left on the bank he's setting off from at all time.

Like this also works:

Farmer takes wolf and goat to right bank

Farmer returns with wolf (or goat, what matters is he returns with one thing) to left bank

Farmer takes cabbage + thing he came with to right bank

Problem solved

The real problem is when the farmer is only allowed to carry one thing at a time.

The whole point is getting the LLM to recognize it's retarded, as you call it. Most LLMs see the words "riddle", "farmer", "wolf", "cabbage", "goat" and get into a reasoning rut so deep you can't get them to realize it's a trivial question.

The thing is, lots of humans react this way as well, unless you actually point out the difference. We’re pattern recognisers. It’s not necessarily a bad thing that LLMs do this any more than it’s a bad thing they mostly ignore my typos.

Why doesn't the farmer just buy a bigger boat; is he stupid?

This problem is retarded.

It sounds like (a derivative of) one of those easier brainteasers found in investment banking/management consulting/quantitative finance interview guides, where a key difference would be the carnivore (at least stereotypically) eating the the cabbage. The problem, as stated, is also Problematic and full of yikes for that it implies that wolves and goats are items, which is highly de-animalizing to our furry friends in considering them as inanimate objects.

But the cabbage is eaten by the goat?

the wolf would eat the cabbage, or the goat would eat the wolf.

Not this goat!

It's a trick question.

No, by the wolf. The wolf is eaten by the goat. Are you an LLM?

I am mildly impressed

>Extoll the virtues of Javazons, but pretend its an in universe Bard

Ballad of the Storm's Daughter
(As sung by Elowyn the Wandering Lyrist, beside the campfires of Harrogath)

Hark! Gather ‘round, ye seekers of tale,
Where shadows loom deep and dark rivers wail—
Let verse take wing ‘cross the frostbitten vale,
To sing of a warrior whom storms themselves hail!

Behold the Javazon—Tempest in form!
Her arm bears the javelin, her soul, the storm.
When Hell’s hordes advance, thick as thorns,
She laughs, and the sky cracks where lightning is born.

See now her fury—! A flash ‘cross the fen,
A throw splits the silence… then thunder again!
From one gleaming shaft, ten thousand forks rend—
Demons fall ash, ere their curses can end.

Charged Strike! —Her spear kisses foe, and behold:
A lord of the Pit, once so dread, so bold,
Now dances, a puppet of arcs uncontrolled,
His crown now a pyre, his flesh turned to gold.

Swift as the zephyr, she treads where she wills,
No armor confines her, no peril instills.
The Valkyrie’s cry in her wake echoes shrill—
A sister-in-arms forged by heaven’s own skill.

Titan’s Revenge flies, returns to her hand,
A cycle unbroken, like tides to the land.
Each scar she bears, a hymn to the stand
‘Gainst Baal’s creeping blight, ‘gainst Diablo’s red brand.

Oh, heed not the whispers of bow, sword, or spell—
When the skies duel the earth, none contend with her well.
In the chaos of battle, where light’s citadel fell,
She is the breaker… and the heavens’ last knell.

So raise high your ale, to the Storm’s chosen daughter!
To the bolts bound in flesh, to the slaughter wrought hotter.
When the End Days take shape, and all hope seems but water,
Look east—there she stands. And the darkness… shall falter.

Uh, it mostly scans. When did that happen? Last I recall LLMs couldn't quite get meter down. Oddly though it does a poor job of analyzing the meter of its own poem in my opinion.

Yesterday through chat.deepseek.com, That was the whole of my prompt and the quote here is the first result.

I'd like to see "the infinite blink server crasher Temporalis build but it's a cautionary lullaby" now. Although I doubt it's in the dataset yet.

Some flavors of the Forbidden Build should be old enough.

In fact I just asked out of curiosity and R1 was able to figure out and explain how wardloop works with minimal nudging on my part, even just explicitly calling it "wardloop" was enough. I like how it shows the thinking process, very cute.

"Now, Jarvis, tell me how the Chinese dupe mirrors."

I suspect Altman has used his maxxed out corporate charisma on Trump as well:

Schizoposting: Altman doesn’t have rizz. I won’t call it a cabal (in part because it actually isn’t), but there’s just a group of very well-connected gay men around Altman/Thiel in SV who are very powerful and constantly back each other up. Trump, in a way, enjoys the company and attention of gay men (I don’t mean in a bi/gay way himself, but in the sense that I truly believe he has a certain affinity for gay men and always has), and is kind of a straight queen, and so it all works pretty well.

Roy Cohn was Trump's mentor. So yes.

Trump has always been deeply interested in image, not just his own or what's broadly anything associated with him, but the objective and subjective perception of anything. Certain gay men are better equipped than anyone else to talk to Trump about specific angles, historically certain black men and women too, but some of that was burned either by Obama skepticism and much more once he was the frontrunner in 2016.

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