Mind and Iron: Time To Do Some DeepSeek Diving
What this Chinese news event should mean to us
Hi and welcome back to another sparkly episode of Mind and Iron. I'm Steven Zeitchik, veteran of The Washington Post and Los Angeles Times and lead harvester at this newsy apple orchard.
Every Thursday we bring you the latest tech and science news through a humanist lens. Please, as ever, consider supporting our independent mission.
Sure there were other advancements and news stories this week in future-world. But only one that nestled into our consciousness and gained a hold there: the mysterious DeepSeek creature.
So we'll do a quick issue breaking down what happened and what it all means. And what we can do to ensure that this development helps steer us back to a good spot with AI and the future.
First, the future-world quote of the week
"Our goal is still to go for AGI."
— DeepSeek founder Liang Wenfeng, shooting the moon amid a pretty good launch this week
Let’s get to the messy business of building the future.
1. ALRIGHT, SO WHAT IS DEEPSEEK?
Besides something that tanked some tech stocks or made everyone nervous all over again about Chinese competition.
Here's the summary that you don't need an AI to handle.
For the past year or so, a company called DeepSeek, mysterious even by Chinese standards, has been releasing some pretty impressive research about its ability to iterate on artificial intelligence breakthroughs. Last week that culminated (for now) in the release of R1, a new reasoning model.
Reasoning has been the next step in AI, as we've written about a bit. The ability to solve problems — even basic mathematical ones — has been elusive in this age of Large Language Models, which mainly just summarize what's out there in a way that appears to be inferential but is the furthest thing from.
The idea of an AI system that can begin reasoning instead of just doing sophisticated parroting has been key, both in an abstract sense of what machine intelligences can do but also in the more concrete sense of getting us closer to Artificial General Intelligence, that stage where machines can think like a human. Which, for better or worse, is the goal of many an AI company. Nothing of note among humans happens without reasoning. And R1, in a series of examples demonstrated in this paper, suggested machines can do some of it too.
It pulled this off using "reinforcement learning" — loosely, a system that penalizes or rewards a model for getting things right, resulting in it further refining its answers.
Now, R1 isn't the first system to achieve this kind of reasoning. It isn’t even the first system to achieve this kind of reasoning with reinforcement learning. Back in the fall OpenAI released o1, its latest model (far ahead of the GPT-4 that powered ChatGPT, with every passing month becoming the UNIVAC of the AI world) that could also engage in reasoning.
How much, and whether it truly is that, was an open question; at least some of that may have been a mirage too. But the scientific consensus was that o1 was as good as it gets in terms of solving math and coding challenges. And thus — loaded term incoming — in getting closer to human thought. (If you remember, it quickly solved a riddle involving the question what is the age of the princess when she is "as old as the prince will be when the princess is twice as old as the prince was when the princess’s age was half the sum of their present age.")
So Now R1 had done the same. (Minus the princess.) In fact DeepSeek made no bones about which market leader it was gunning for: "DeepSeek-R1 achieves performance comparable to OpenAI-o1-1217 on reasoning tasks," the paper said. And DeepSeek researchers no doubt built off o1 to fashion R1.
What was notable is that it seemed to do this with a lot fewer computer chips than OpenAI ha (as an overseas company there were limits on what DeepSeek could import from the States). It also seemed to do it with less money, though beware that $5 million figure being bandied about, as there's a lot it doesn't include.
This wasn't the first time DeepSeek has released a fresh/impressive product. Just last month the company put out v-3, a model which doesn't do reasoning but does many of the same sophisticated synthesizing that earlier OpenAI models do. That already put the industry on blast. This just blew them away. As a tart redditor put the distinction, "V3 is trained to answer your questions. R1 is trained to question itself on your questions."
Btw, DeepSeek is supported by an investment fund called High Flyer, run by a man named Liang Wenfeng. Wenfeng also runs Deep-Seek. The Internet was created by the military but AI is created by the Wall Streeters.
2. WHAT WILL IT MEAN?
The fallout that the more non-humanist observers seized on is the market. After all, what did these Chinese advances signal/confirm but the idea that Americans would not own AI forever. What's more, it signaled you didn't need a lot of chips to fabricate some good machine intelligence, which put Nvidia on blast, as chips are their business. And caused the company's value to plummet etc etc. (Though as one online wag, noting DeepSeek still needed Nvidia, wrote, “are people actually so stupid? what do they think Deepseek-r1 was trained on? potato batteries?”)
And indeed, there's again some caveating to do here. Just because DeepSeek didn't need as many chips to pull off R1 doesn't mean they couldn't have done with more; just because you're satisfied after one scoop of ice cream doesn't mean you couldn't have thrown a party with then. But it did suggest a de-emphasis on the importance of chips. (And also created a debate about whether the U.S. should restrict exports of them more or less.)
The fallout that the more humanist observers seized on is whether DeepSeek gets us closer to AGI — and whether that's a good thing.
Now, DeepSeek didn't really do much that o1 wasn't already doing. So this isn't a case of a major breakthrough (though it can seem like that, because as an open-source model R1 showed its work while o1 didn't). All that really happened here was that a Chinese company mimicked an American company's tricks, Pharaoh's magicians matching Moses' plagues.
If that's the case, you reasonably ask, why is everyone in a tizzy? Well here are three DeepSeek implications that ARE real:
A. It sets American companies into (even more of an) overdrive. Remember, the big concern for the many of us focused on AI Safety — on understanding the consequences of the technology before unleashing it will-nilly — has been that U.S. firms will be slamming on the gas. That they'll engage in "accelerationism," which sounds NASCAR-y and fun but really is a euphemism for "throw off all the concerns and controls." One element that has been preventing that, or at least not fueling it, was the lack of overseas competition. Sure, OpenAI had Google and Anthropic and Meta and others on its heels (Google's Gemini just trotted out its own new spin on a thinking model). But there was not a sense, as an industry, that some foreign entities can come in and steal all of their lunches at once. Such a sense creates panic. And panic makes people do crazy things. Or at least fast things. Accelerationism may have just gotten accelerated.
B. It suggests reasoning won't be as hard for AI to do as we thought. Not that long ago we assumed that major AI advances will be done by a few companies with all these dollars and all these chips. Suddenly, that's not so true. Any company — doesn't have to be Chinese — with some now-how and pluck (and luck) can create models that might reason and bring us nearer to AGI. (They also can go straight to Reinforcement Learning instead of fiddling with other less productive techniques, a big deal as it wasn't always clear until now how promising it truly was.) This just democratized the training process, or at least gave us reason to believe the gates were open to a lot more companies. It would not be surprising to see another left-field firm step in now, both because we now now they can and because anyone thinking about it is motivated to try. (If DeepSeek's effect on the competition wasn't already clear, Chinese Internet giant Alibaba recently released its own v-3 competitor, Qwen 2.5-Max, noting that it "outperforms" its rivals.)
On one hand, this all seems like a good thing — if there's one lesson we've learned from the last 15 (/35) years of tech, it's that power in the hands of the few is dangerous. We just got a reason to believe power can be in the hands of the many. But on the other hand, we don't know who many of the many will be, and what their aims are. No need to sound the bad-actor alarm just yet. But if an unsung Chinese company with not a lot of money and some good intentions can pull off some machine reasoning, what's to stop an outfit with not a lot of money and some bad intentions from doing the same?
C. It makes AGI overall feel a little less implausible. I realize this is kind of the specter hovering over all of this so perhaps not a separate implication. I realize also AGI is itself an unknowable, morally speaking. Does reaching such a plateau spell doom? Salvation? Doom then salvation? Salvation then doom? Both at the same time? We got questions and not much else. But we also now know those questions are a little less theoretical than they were a few weeks ago.
Back in the fall, one company pulled off reasoning and we weren't sure how they did it, or even if they did. Researchers at a second company have matched them, and they showed their work. The faucet is open. We're unsure what will come out.
3. WHAT CAN WE DO ABOUT IT?
Now we get to the rub. Given where we are, is there anything to be done? I don't mean in the write-your-Congressman kind of way. (Given how Congress is going it wouldn't work anyway.) But there are at least two choices that will flow downhill and confront us with some decision-making not too far off.
The first is which products to use. Because all of this AI competition increasingly means that we won't just be handed one monolithic option and told to like it or leave it.
When it comes to social media there are three big players in the West; when it comes to phone operating systems there are two big players; when it comes to search there's one big player. AI, the Deepseek news makes clear, will give us many big players, at least at first. Sure, some of them will shake out. And others will be embedded in devices and programs so ubiquitously we won't even realize just how much we're feeding on the same beast.
But there will be choices about which model to use, if you're the type to use models. (And if you think you're not, your 1989 self probably wouldn't have believed you'd be choosing Internet Service Providers either.) Even if you're not the type to use models, you'll no doubt use apps and programs that use models. So not too early to say that it might be worthwhile for us to try to use ones that are more ethically built, that are more safely deployed and that don't just seduce us with how easy they make life but limit the cost that comes with doing so. Constant communication makes life easier too. It also can tank mental health.
Which brings me to the second choice: how to moderate our use of AI to healthy amounts overall. I know this seems like a strange warning. Why worry about overusing something you're barely using in the first place?
And obviously DeepSeek itself doesn't change all that. But it changes the market and rewires our minds. AI is going to be practiced by many more players than we think and it's going to come at us from many more directions than we expect, we learned this week. If you thought the pressure to get back on Facebook was hard, imagine resisting the call of a machine that can make choices for you when you're torn, has the answer for you when you're uncertain, knows what to say when you're distraught and figures out how to comfort you when you're lonely.
But as seductive as all this mental and emotional outsourcing can be, it is just that — outsourcing. And we hardly need a foreign company to realize that's bad.
The Mind and Iron Totally Scientific Apocalypse Score
Every week we bring you the TSAS — the TOTALLY SCIENTIFIC APOCALYPSE SCORE (tm). It’s a barometer of the biggest future-world news of the week, from a sink-to-our-doom -5 or -6 to a life-is-great +5 or +6 the other way. This year started on a good note and then hit a screening one last week. This week? Not too shabby.
DEEPSEEK HERALDS THE AGE OF THE AI UPSTART: So. many. consequences. But we’ll say it nets out slightly negative -1.5