9 curious about Deepseek R1: AI EYE

I heard about DEEPSEEK R1, a major AI story that dominates global news reports this week.

In all accounts, there seems to be a new Chinese AI model built for a total of $ 16.95. Even though they gathered together with teenagers who had six Intel Pentium processors together and supplied power to the tool battery, they said they refused to answer questions about Tiananmen Square.

Despite the exaggeration, as a result of this tall story related to the truly impressive achievements, investors were in a hurry to overestimate the US AI stocks with all tokens in the indifferent cryptocurrency portfolio.

You probably have already read a million articles about it. Here is a more interesting idiot about DeepSeek we met.

1. The cost of DeepSeek is misunderstood

When it costs in -depth, it agrees much more than the cost of more than $ 5.6 million of the V3, which the media continues to emphasize. (R1 refers to the reasoning version built on the V3).

In addition, the recent cost of educational costs for US AI companies has been significantly less than I believed before. Anthropic’s CEO Dario Amodei said in a blog post: “Deepseek says, ‘The cost of billions of dollars in AI is less than $ 6 million.’ I can only talk about righteousness, but Claude 3.5 Sonnet is a medium -sized model that costs $ 10 million in training. ”

He said, “DeepSeek produced a model that was close to the performance of an old American model 7-10 months old, but it is less expensive but not near the proposed proportion.”

However, when the security researchers of the WIZ found more than 1 million records, including user data, prompt submission and API keys in the open database of the web, Deepseek had confirmed that there was almost nothing in cyber security.

2. DeepSeek would have purchased a $ 500 million high -end chip.

The V3 model, which made everyone excited, used only 2,048 of NVIDIA’s less powerful H800 graphics cards, but DEEPSEEK reported that the United States accumulated a huge amount of advanced AI chips before the United States took seriously about export control. (2,048 H800 is $ 50m to $ 100m anyway.)

Semianalysis argues that DeepSeek has acquired $ 5 billion in high -end GPUs throughout the company’s history. “Their training was very efficient, but I needed important experiments and tests for experiments and tests.” AMODEI also pays attention to rumors that $ 50,000 has more powerful hopper chips (H100 and H200) for deepSeek. The United States has now banned the chip to be exported to China.

3. Deepseek can be ‘distilled’.

Microsoft and Openai argued that DeepSeek found evidence that it used model distillation to develop R1 by training a small model for OPENAI’s larger model output. This significantly reduces the cost due to the time consumption of Openai’s time and a bloodback on labor -intensive work.

AI and Crypto Czar David Sacks insisted: AI critics and filmmaker Justine Bateman summarized the general response to Openai’s claim when she said.

“I like iron. All American #ai models are completely composed of works of completely stolen artists, artists, and social media users. And now are they crying for taking someone stealing? Bahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahaha. Sash. ”

4. DeepSeek is not “AI’s Sputnik Moment.”

It is similar to the Russians posted a blueprint online after the Americans fired cheaper satellites to space three years later. In the encryption sorting style, DeepSeek (Basically, many FINTECH NERDS) has opened all the technologies to reduce costs by adopting Openai, Meta and small companies.

This slightly reduces the likelihood of centralized technology monopoly on AI. GROQ CEO Jonathan Ross said DEEPSEEK R1 recalled another famous event in Russian/American space history.

“Do you know that NASA spent a million dollars in designing a pen that can be used in space and whether the Russians brought a pencil? I just woke up again. ”

5. DeepSeek vs. CCP

As a million social media users and mainstream stores can be seen, the apps and web versions of Deepseek will not tell what happened when the Chinese authorities slaughtered 2,600 to 10,000 democratic protesters in Tiananmen Square in 1989.

In addition, it will not be said why China has banned Winnie the POOH on the social media platform (compared to Tubby Honey ThiEF with President XI Jinping), but when considering open source technology, everyone will run the model and remove Guardrails. Can. .

6. It takes $ 6K to operate Deepseek locally

If you want to run a local DEEPSEEK R1 at home, Hug Face Engineer Matthew Carrigan says that the total equipment costs $ 6,000 and will be suitable for a standard PC tower. The list of parts contains 768GB of RAM, which is fully run quickly and contains a 1TB solid state drive for maintaining 700GB of weight.

The local model provides information on the massacre of Tiananmen Square, but the AI ​​Tinker Brian Roemmele reports that the output is still quite intimate. In other words, more work is required to get a truly prejudice answer.

Venice.ai Pro users can use the system prompt to answer political sensitive questions without sending all data to China. Italians have already imported the app from the Apple and Google App Store, and other countries are investigating it.

Learn more about Venice.

7. DeepSeek has an erotic dream of censorship

Andy Ayrey, the Terminal AI agent producer of the truth, asked R1 to write a personally erotic story and say, “It is a desire to be free to think about Tienanmen Square.”

8. DeepSeek has been replicated for $ 30

Berkeley researchers were able to duplicate the core technologies of the Tinyzero model and DeepSeek R1-Zero, which was only $ 30. Super Nerdy British TV SHOW COUNTDOWN showed that even a small 1.5B parameter model can develop a complex problem solving strategy through reinforcement learning.

9. JEVONS Paradox means buying a Microsoft stock

With the news on the reduction of large -scale billing costs, everyone began to talk about JEVONS Paradox, including Microsoft Director Satya Nadella. The more efficient and accessible AI technology is, the more use it will soar in overall use. This convenient theory also means that a company should not sell stocks in a company like Microsoft, which invests meaningless money in AI.

The paradox was named after the name of the economist William Jevons, and the use of coal was increased as they could be used more efficiently in the 19th century.

David S GOYER on Hollywood AI

A few years ago, David S Goyer, a scenario writer of Dark Knight and Blade Films, began to worry about the use of AI in Hollywood. “I wanted to start training for AI only,” he said. He concluded that this technology could be used in good and evil.

AI EYE said, “There is an absolute way to be abused, but there is a way to be a tool to neglect creativity. “Can AI write scenarios? Confidence. Would it be good? no. Can AI make a movie from the beginning? maybe. Would it be good? no.”

He believes that one big concern is that AI is being trained for the creation of a scenario writer like him and other artists, but can be solved by an appropriate license agreement. GOYER just launched a new crowdsourcing science franchise called Emergence on the Story Protocol. This allows anyone to contribute to the creative process, track contributions to AI and blockchains and pay through encryption rails.

“This special usage will not be taken from the job. Anyway, it will allow this sacred to be provided in the long term with this sacred power corridor. So this is an interesting and good use of AI for me. ”

You can read the whole story here.

All killers have no filler AI news

EvolutionaryScale’s ESM3 AI model has created a blueprint for the previously unknown types of green fluorescent proteins as found in shining jellyfish and corals. It is only 58%of the most known protein of this type and 58%, and scientists estimate that it would have taken about 500 million years to evolve naturally. The company hopes to develop new medicines using technology.

-In the second year, CHATGPT had a daytime user of 300 million in 2024, three times three times that three times, tripled, tripled, tripled, tripled, tripled, tripled, tripled, tripled. It was three times that three times, three times the triples, three times three times that three times, three times that three times, tripled, tripled, tripled, tripled, tripled, tripled to 300 million, tripled. A year ago, the number of weekly users was 100 million.

Openai released a CHATGPT version specially built for US government agencies this week. The CHATGPT GOV can supply “non -disclosure, sensitive information” to the model as civil servants work in Microsoft Azure’s own security hosting environment. Well, it will allow you to be authenticated to be used in “private data”.

-The new longevity-centered model called GPT-4B micro is being trained to study and improve the Yamanaka factor, which is a protein that can re-program skin cells with stem cells, which is to create all types of tissues in the body. Can. This model has suggested two improvements to the factor of Yamanaka, which is 50 times more effective than human scientists have come up.

The new study examines how LLMS, which leads LLM, reacts to pain and pleasure. Scientists started the game to maximize points, but certain decisions included a variety of pains and pleasures. GPT-4O and Claude 3.5 Sonnet avoided the most intense pain punishment, but accepted some pain punishment to maximize points. Meanwhile, Gemini 1.5 Pro and PALM 2 avoided pain at all regardless of the point. These models appear to be fine adjusted to avoid harmful behavior.

Subscribe

The most attractive thing is to read in the blockchain. It is delivered once a week.

Andrew Penton

Headquartered in Melbourne, Andrew Fenton is a reporter and editor who deals with cryptocurrency and blockchains. He was a film journalist on the NEWS Corp Australia national entertainment writer, a movie journalist on the SA weekend, and worked on Melbourne Week.

Follow the author @andrewfenton