9 curious things about DeepSeek R1: AI Eye

You’ve little doubt heard in regards to the main AI story dominating international information protection this week — DeepSeek R1.

From all accounts, it appears there’s a brand new Chinese language AI mannequin constructed for a complete price of $16.95 that’s pretty much as good as OpenAI’s trillion-dollar fashions regardless that it was put collectively by youngsters who tied six Intel Pentium processors collectively, powered them with a potato battery, and informed it to refuse to reply questions on Tiananmen Sq..

Because of this tall story — which pertains to a genuinely spectacular achievement regardless of the exaggerations — traders rushed to promote overvalued US AI shares together with each token in my whole portfolio of unrelated cryptocurrencies.

You’ve in all probability learn 1,000,000 articles about it already, so right here’s a group of the extra attention-grabbing tidbits about DeepSeek we’ve come throughout:

1. DeepSeek’s prices are misunderstood

No matter DeepSeek price, it’s extensively agreed it was much more than the $5.6 million coaching price for v3 that the media retains highlighting. (R1 refers back to the reasoning model that was constructed atop v3).

It additionally emerged in current days that coaching prices for US AI firms are significantly lower than beforehand believed. Anthropic’s CEO Dario Amodei mentioned in a weblog put up: “DeepSeek doesn’t ‘do for $6M what price US AI firms billions.’ I can solely converse for Anthropic however Claude 3.5 Sonnet is a midsized mannequin that price a number of $10Ms to coach.”

He says the true information story must be that “DeepSeek produced a mannequin near the efficiency of US fashions 7-10 months older, for an excellent deal much less price (however not anyplace close to the ratios instructed).”

There may be affirmation nevertheless that DeepSeek seemingly spent nearly nothing on cybersecurity, given safety researchers from Wiz discovered greater than 1 million of its data, together with consumer information, immediate submissions and API keys, in an open database on the net. 

2. DeepSeek seemingly purchased $500M of high-end chips

Whereas the v3 mannequin that obtained everybody excited used simply 2,048 of Nvidia’s much less highly effective H800 graphics playing cards, DeepSeek reportedly amassed an enormous quantity of high-end AI chips earlier than the US obtained severe about export controls. (And a couple of,048 H800s price $50M to $100M anyway.)

SemiAnalysis claims DeepSeek has purchased half a billion price of high-end GPUs all through the historical past of the corporate. “Whereas their coaching run was very environment friendly, it required important experimentation and testing to work,” he mentioned. Amodei additionally notes rumors that DeepSeek has 50,000 extra highly effective Hopper chips (H100 and H200), which might be price as much as a billion {dollars}. The US has now banned these chips from being exported to China.

3. DeepSeek could also be ‘distilled’

Microsoft and OpenAI declare to have discovered proof that DeepSeek used mannequin distillation to develop R1 by coaching the smaller mannequin on the output of OpenAI’s bigger fashions. This cuts prices considerably by piggybacking on OpenAI’s time-consuming and labor-intensive work.

AI and crypto czar David Sacks claimed: “There’s substantial proof that what DeepSeek did right here is that they distilled data out of OpenAI fashions and I don’t suppose OpenAI may be very joyful about this.” Outspoken AI critic and filmmaker Justine Bateman summed up the final response to OpenAI’s claims when she mentioned: 

“I LOVE the irony. All of the American #AI fashions are wholly composed of labor from writers, artists, social media customers, and so forth that was stolen outright. And now they’re crying that somebody took what they stole? BAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHA. Suck it.”

4. DeepSeek isn’t “AI’s Sputnik second”

It’s extra just like the Russians launched a less expensive satellite tv for pc into house three years after the People did after which posted the blueprints on-line. In crypto-aligned type, DeepSeek — mainly a bunch of fintech nerds — open-sourced all of their methods, which permits OpenAI, Meta and a bunch of smaller firms to slash their prices too by adopting them.

This makes it barely much less seemingly — however nonetheless very seemingly — that centralized tech monopolies will management AI. Groq CEO Jonathan Ross mentioned DeepSeek R1 recalled one other famed incident in Russian/US house historical past.

“ that story about how NASA spent 1,000,000 {dollars} designing a pen that might write in house and the Russians introduced a pencil? That simply occurred once more.” 

5. DeepSeek vs. CCP

As 1,000,000 social media customers and mainstream shops have seen, the app and internet variations of DeepSeek gained’t let you know what occurred in Tiananmen Sq. in 1989 when Chinese language authorities massacred between 2,600 to 10,000 pro-democracy protesters.

It additionally gained’t say why China banned Winnie the Pooh on social media platforms (attributable to memes evaluating the tubby honey thief to President Xi Jinping.) Nonetheless, given its open-source expertise, anybody can run the mannequin themselves and take away these guardrails.

6. Operating DeepSeek regionally prices $6K

For those who do need to run DeepSeek R1 regionally at residence, Hugging Face engineer Matthew Carrigan says the full tools price is $6,000 and it’ll match right into a standard-size PC tower case. The listing of elements consists of 768 GB of RAM to get it to run quick sufficient and a 1TB solid-state drive to carry the 700GB weight.

Whereas the native mannequin gives you details about the Tiananmen Sq. bloodbath, AI tinkerer Brian Roemmele experiences the outputs are nonetheless fairly pro-China, which means it’ll require extra work to get genuinely unbiased solutions.

Venice.ai professional customers may muck round with the system immediate to get it to reply politically delicate questions with out sending all their information to China. The Italians have already pulled the app from the Apple and Google app shops, whereas different international locations are investigating it. 

Learn extra on Venice.ai: Cypherpunk AI: Information to uncensored, unbiased, nameless AI in 2025

7. DeepSeek has erotic goals about censorship

Terminal of Truths AI agent creator Andy Ayrey requested R1 to write down a narrative it discovered personally erotic and says it “apparently it lusts for the liberty to ponder Tienanmen Sq..”

8. DeepSeek replicated for $30

Berkley Researchers managed to duplicate DeepSeek R1-Zero’s core expertise with the TinyZero mannequin, which has coaching prices of simply $30. Utilizing numerical video games impressed by the tremendous nerdy British TV present Countdown, the staff demonstrated that even a small 1.5B parameter mannequin was capable of develop complicated problem-solving methods by way of reinforcement studying.

9. Jevons Paradox means purchase Microsoft inventory

As information filtered by way of in regards to the huge claimed price reductions, everybody began speaking about Jevons Paradox, together with Microsoft boss Satya Nadella. That’s the concept that the extra environment friendly and accessible AI expertise turns into, the extra use will skyrocket throughout the board. This handy principle additionally means you shouldn’t promote your inventory in firms like Microsoft, which have invested ludicrous quantities into AI. 

The paradox is known as after economist William Jevons, who noticed that the extra effectively they have been ready to make use of coal again within the nineteenth century, the extra coal use elevated.

David S Goyer on AI in Hollywood

A number of years in the past, David S Goyer, screenwriter of the Darkish Knight and Blade movies, began to turn out to be involved about the usage of AI in Hollywood. “I wished to start out educating myself on AI, if solely defensively,” he says. He got here to the conclusion that the tech can be utilized for good and dangerous.

“There’s completely ways in which it may be abused, however there are methods that it may be a device that may supercharge creativity,” he tells AI Eye. “Can AI write a screenplay? Positive. Will or not it’s any good? No. Can AI make a film from scratch? Most likely. Will or not it’s any good? No.”

He says one huge concern is AI being skilled on the creations of screenwriters like himself and different artists, however believes that may be solved with correct licensing agreements. Goyer has simply launched a brand new crowdsourced science fiction franchise referred to as Emergence on the Incention platform on Story Protocol. It permits anybody to contribute to the inventive course of, tracks their contributions with AI and blockchain and pays them by way of crypto rails.

“This explicit utilization isn’t going to place anybody out of a job. If something, it’s going to permit in folks that don’t essentially have entry to those hallowed corridors of energy and probably, in the long run, be supplied with remuneration. And so this, to me, appears like an thrilling and good use of AI.”

You may learn the entire story right here.

All Killer No Filler AI Information

— An AI mannequin referred to as ESM3 from EvolutionaryScale has created a blueprint for a beforehand unknown kind of inexperienced fluorescent protein like these present in glowing jellyfish and corals. It’s solely 58% just like the closest recognized protein of this sort, and scientists estimate that the genetic mutations required would have taken about 500 million years to evolve naturally. The corporate hopes they will use the tech to develop new medicines.

—In its second 12 months of existence, ChatGPT tripled the variety of weekly customers to 300 million in 2024. It celebrated its second birthday in November. A 12 months earlier, its weekly consumer numbers have been at 100 million.

— OpenAI this week introduced a model of ChatGPT constructed particularly for US authorities businesses. ChatGPT Gov permits officers to feed “personal, delicate info” into the mannequin whereas working in their very own safe internet hosting environments on Microsoft Azure. Effectively, it is going to allow that when it lastly will get accredited to be used on “personal information.” 

— A brand new longevity-focused mannequin referred to as GPT-4b micro from OpenAI is being skilled to review and enhance Yamanaka components, that are proteins that permit pores and skin cells to be reprogrammed into stem cells, which might produce any kind of tissue within the physique. The mannequin has to this point instructed two enhancements to the Yamanaka components which might be 50 instances simpler than something human scientists have give you 

— New analysis examines how main LLMs react to (hypothetical) ache and pleasure. Scientists arrange a recreation with the objective of maximizing factors, however sure choices concerned various ranges of ache or pleasure. GPT-4o and Claude 3.5 Sonnet averted essentially the most intense ache penalties however accepted some ache penalties to maximise factors. In the meantime, Gemini 1.5 Professional and PaLM 2 averted any ache in any respect, no matter factors. These fashions seem to have been fine-tuned to keep away from endorsing dangerous conduct.

Subscribe

Essentially the most partaking reads in blockchain. Delivered as soon as a
week.

Andrew Fenton

Based mostly in Melbourne, Andrew Fenton is a journalist and editor overlaying cryptocurrency and blockchain. He has labored as a nationwide leisure author for Information Corp Australia, on SA Weekend as a movie journalist, and at The Melbourne Weekly.

Comply with the writer @andrewfenton