26 Comments
User's avatar
Patrick Jordan's avatar

As always, nice rundown. Also, your mention of 'concerns' around it i good to see. You are only the second writer / publication to mention concerns related to it. Wired is the other: https://www.wired.com/story/deepseek-ai-china-privacy-data/

Expand full comment
Charlie Guo's avatar

I am nothing if not nuanced!

Expand full comment
Patrick Jordan's avatar

Which is another reason your work is appreciated. My job sometimes makes me very under-nuanced :)

Expand full comment
Daniel Nest's avatar

Wait, you mean the tiny footnote #2 that in a few words off-handedly mentions the lack of discussion about safety testing and alignment in my last week's post doesn't qualify? ;)

https://www.whytryai.com/p/deepseek-r1-free-openai-o1-alternative

Expand full comment
Patrick Jordan's avatar

Of course not. But sort of. Sorry / not sorry??? 😄

Expand full comment
Daniel Nest's avatar

I poured entire seconds of deep-ish thought into that footnote. How dare you!

Expand full comment
Patrick Jordan's avatar

Agreed

Expand full comment
Patrick Jordan's avatar

Lol. Are we in 'we must never speak again' territory?

Expand full comment
Daniel Nest's avatar

We must never (deep)seek again!

Expand full comment
Andrew Sniderman 🕷️'s avatar

Seems to me, and I'm going on waaay less info than you, that - politics aside - this is a positive technical development overall that will improve AI efficiencies and cost structures. Am I way off?

Expand full comment
Charlie Guo's avatar

I think it’s fair to say about open source AI. But V3 and R1 are much harder to create without GPT-4 and o1, and from the outside of the labs it’s hard to say whether GPT-5 and o3 (or their successors) will hugely benefit. Does that make sense?

Expand full comment
Andrew Sniderman 🕷️'s avatar

totally, but that's why this feels like progression. Guess we'll see, huh?

Expand full comment
Charlie Guo's avatar

I do think it's a big step, one the scale of the first commercially-licensed Llama or o1. But I'm still flabbergasted as to exactly why this one is getting so attention versus the other two. Best guess: it's free, so millions of folks who have only ever tried 4o-mini and 3.5 Haiku are now seeing what a SOTA reasoning model is capable of.

Expand full comment
Patrick Jordan's avatar

I think that may well be true. My view is colored by the fact that I work in cybersecurity. China is considered an adversary nation and a significant threat from that perspective. Apologies to anyone who is Chinese - China will say the same about the US.

Expand full comment
Andrew Sniderman 🕷️'s avatar

Just read this from ChinaTalk substack - here we are - and apologies to anybody old - punditting about stuff we'll never know the inside scoop on. But maybe Charlie will ;)

"DeepSeek prioritizes “young and high-potential” candidates — specifically those born around 1998 with no more than five years of work experience, similar to other AI labs in China. Said one DeepSeek employee to The Paper, “The success of DeekSeek has demonstrated the power of young people, and in essence, that the development of this generation of artificial intelligence needs young minds."

Expand full comment
Charlie Guo's avatar

I saw a tweet recently from an Nvidia person who said one of their most promising candidates turned them down for DeepSeek over a year ago. No idea if it's true, but interesting to see that they were capable of attracting talent well before this.

Expand full comment
Andrew Sniderman 🕷️'s avatar

Just saw Satya's tweet about Jevons paradox which seems to be trending into the explainer rationale of yay! dont worry the better it gets the more demand and dineros for all

Expand full comment
Joel M. Hoffman, PhD's avatar

And here's a question, if you know: Even if Deepseek doesn't require as much computational power as (say) ChatGPT, won't huge arrays of Nvidia chips still be helpful in scaling up Deepseek?

More generally: on the reasonable assumption that Deepseek has useful and generalizable optimizations, won't other companies be able to incorporate those advances, and then have an additional advantage because of faster hardware?

Expand full comment
Charlie Guo's avatar

Yes, I believe both of those things are true: other companies are going to benefit from DeepSeek's breakthroughs, and in the long run there's still going to be additional demand for Nvidia's chips. I've actually just updated the post with an alternate theory for Nvidia's stock drop - that it was caused by insider trading ahead of President Trump's tariff announcements.

Expand full comment
Andrew Smith's avatar

"It's likely that a US-based company will start providing access in the very near term."

Pretty sure Perplexity added this today. This is moving FAST.

Expand full comment
Daniel Nest's avatar

Well done on organizing everything in such a clean, structured way!

Regarding your footnote #4. It's funny how quickly we reach the point of "news exhaustion" these days. I published my own first hot take on DeepSeek R1 exactly a week ago, and I'm already absolutely drained by following the media circus around it.

As for this: "It's likely that a US-based company will start providing access in the very near term."

Have you seen this?

https://i.imgur.com/5G4XVd5.png

Perplexity is already using self-hosted R1 to power its "Pro" search. Things go vroom-vroom.

Expand full comment
Charlie Guo's avatar

Can't even write a hot take before it's out of date these days.

Expand full comment
Daniel Nest's avatar

Touche!

Expand full comment
Joel M. Hoffman, PhD's avatar

Very nicely done!

For those who don't quite know what AI is, I've put this together:

https://undisciplinedconversations.com/p/what-is-artificial-intelligence

Expand full comment
Kerry's avatar

Thank you for this explainer. I have seen a lot of coverage but mostly skimmed it; this helped me get a better idea of what’s going on. Much appreciated.

Expand full comment
/\/€u Th@/_/ght's avatar

Greed is their demise.

Expand full comment