Dario Ameudi challenges $ 6 million in Dibsic: What Antarbur thinks about the latest Amnesty International movement in China.

The world of artificial intelligence shook last week when DibsicAI, Chinese, has announced the latest linguistic model that seems to match the capabilities of the American American American systems on a small part of the cost. The advertisement widespread Selling market This has eliminated nearly $ 200 billion in the market value of NVIDIA and raised hot discussions on the future of developing artificial intelligence.

The narration that appeared quickly suggested that Dibsic was Mainly The economies of building advanced artificial intelligence systems, which are supposed to achieve with only 6 million dollars, what American companies have spent billions to complete them. This interpretation sent shock waves across Silicon Valley, where companies love Openaiand manAnd Google He justified huge investments in computer infrastructure as necessary to maintain the technological edge.

But amid market turbulence and unimaginable newspaper addresses, Dario AmoudiThe co -founder of Anthropic and one of the pioneering researchers behind the great language models today has published a detailed analysis that provides a more accurate perspective on Deepseek’s achievements. for him Blog post Hysteria penetrates to present many decisive ideas about what Deepseek has already completed and what it means for the future of artificial intelligence development.

Below are the four main visions of Amodei analysis that reshape our understanding of Deepseek:

1.

Dibsic mentioned Development costs It should be displayed through a wider lens, according to Imoudi. in AnalysisDirectly defies the popular interpretation:

“Deepseek does not do” for $ 6 million, which cost us billions of billions of intelligence companies. “I can only talk about the anthropoor, but Claude 3.5 Sonnet is a medium -sized model costs $ 10 million for training (I will not give an accurate number). Also, 3.5 Sonnet has not been trained in any way that included a larger or more expensive model (unlike some rumors ).

This shocking revelation mainly changes the narration about the cost efficiency in Dibsic. When considering this Sonata It was trained 9 to 12 months ago and still outperforms the Deepseek model in many tasks, and the achievement appears more with the natural progress of the costs of developing artificial intelligence rather than a revolutionary achievement.

The timing and context are also greatly concerned. After historical trends to reduce costs in developing artificial intelligence – which is estimated at 4x Amoii annually – it appears that the Deepseek cost structure in a large -end and not largely before the curve.

2. Deepseek-V3, not R1, was the true technical achievement

While the markets and the media focus extensively on R1 Deepseek ModelAmodei notes that the most important innovation of the company came early:

“Deepseek-V3 In fact, the real innovation was and what should have had people noticed a month ago (definitely we did). As a prior model, it appears to be close to the performance of American arts models in some important tasks, while the training costs much lower. “

Discrimination between V3 and R1 It is very important to understand the real technological progress of Pipesic. V3 represents real engineering innovations, especially in managing the model.Temporary storage memory of the main value“Pay limits”A mixture of experts” road.

This insight helps to clarify the reason that the dramatic market reaction is in a misplaced R1. R1 has been added mainly to reinforcement learning capabilities to the V3 Foundation – a step that many companies are currently taking with their models.

3. The total companies’ investment reveals a different image

Perhaps the most unveiled side of Amodei analysis is related to the comprehensive Deepseek investment in the development of artificial intelligence:

It was reported – we cannot be sure that it is true – that Dibsic was already 50,000 Gille GillersAnd that I think is within a worker ~ 2-3X for the American AI’s main companies. The cost of the flying chips 50,000 on the $ 1B command. Thus, the total dibsic spending as a company (distinct from spending to train an individual model) is not very different from AI AI laboratories. “

This revelation greatly restores the narration on the efficiency of Deepseek resources. Although the company has achieved impressive results through individual model training, its general investment in developing artificial intelligence seems to be almost similar to its American counterparts.

The distinction between model training costs and the total investment of companies highlights the continuous importance of great resources in developing artificial intelligence. It indicates that although engineering efficiency can be improved, competitiveness in artificial intelligence still requires a significant investment in the capital.

4. The current “intersection point” is temporary

Amodei describes the current moment in developing artificial intelligence as unique but transient:

“So we are at” Cross Point “interesting, as it is temporary that many companies can produce good thinking models. This will quickly stop being correct as everyone moves to the highest scaling curve on these models.”

This observation provides a decisive context to understand the current situation of artificial intelligence competition. The ability of many companies to achieve similar results in the thinking capabilities is a temporary phenomenon instead of the new current situation.

The effects of the future of artificial intelligence development. With companies continuing to expand their models, especially in the field of intensive learning of resources, it is likely that the field is likely to distinguish again based on who can invest more in training and infrastructure. This indicates that although Deepseek has achieved a great prominent sign, it mainly did not change the long -term economy to develop advanced artificial intelligence.

The real cost of building artificial intelligence: What is revealed by Amodei analysis

Daro’s detailed analysis of Debceik’s achievements within weeks of market speculation is to expose the actual economy to build advanced artificial intelligence systems. for him Blog post Siles systematically dismantling all of the panic and enthusiasm that followed the Deepseek advertisement, explains how the $ 6 million stereotype for the company is in the fixed AI development process.

Markets and the media are attracted towards simple novels, and the story of a Chinese company has proven to undermine the costs of developing artificial intelligence significantly. However, Amodei’s collapse reveals a more complicated fact: the total Deepseek investment, especially a billion dollars reported in computer devices, reflecting the spending of its American counterparts.

This is the moment of parity in the cost between us and the development of Chinese artificial intelligence, what Amodei calls “Cross point“A temporary window where multiple companies can achieve similar results. His analysis indicates that this window will close with the progress of artificial intelligence capabilities and intensify training demands. The field is likely to return to the preference of organizations with the deepest resources.

Advanced artificial intelligence is still an expensive endeavor, and the exact Amodei examination shows the reason that measuring its real cost requires a full scope of investment. The systematic dismantling of Deepseek may eventually prove more important than the initial advertisement that sparked this turmoil in the market.

Daily visions about business use cases with VB daily

If you want to persuade your boss at work, you have covered VB Daily. We give you the internal journalistic precedence over what companies do with obstetric artificial intelligence, from organizational transformations to practical publishing operations, so that you can share visions of the maximum return on investment.

Read with us privacy policy

Thanks for subscribing. Check more VB news bulletins here.

An error occurred.