What Retail Investors Need to Know About DeepSeek’s Impact

DeepSeek: Both Performant and Efficient

Having launched on November 29th, 2023, it took until January 20th for Deep Seek large language model (LLM) to make public waves with the following performance comparison chart across coding (Codeforces), software engineering and debugging (SWE-bench), math (AIME), and general purpose reasoning (GPQA/MMLU).

DeepSeek is fully open-source and already MIT-licensed. This alone is problematic for the Western sphere as DeepSeek can then disrupt all the efforts in the “AI safety” direction. Namely, actors within the West may tweak an AI model that is less aggressively controlled than the existing mainstream models from Microsoft (NASDAQ:), Google (NASDAQ:), Meta (NASDAQ:) and OpenAI.

A reminder that Alphabet’s Gemini AI was so tethered with narrative control that, in text-to-image generation, it kept replacing historical figures/groups of European descent with people of African descent. Even Elon Musk called Google’s AI for having “insane racist, anti-civilizational programming”.

Alongside loosening the AI leash potential, DeepSeek appears to be greatly more cost-efficient. By substituting 32-bit floating-point numbers with 8-bit level precision, DeepSeek significantly reduces memory consumption, in addition to using key-value joint compression.

Moreover, unlike ChatGPT’s single-token prediction, DeepSeek utilizes cutting edge multi-token prediction training that improves long-range dependencies. In other words, the 671B parameter model, DeepSeek-V3, can not only better understand context for superior results, but it can learn more efficiently from the same data sets.

Altogether, this allows DeepSeek to be the one of the cheapest AI models on the market, according to Artificial Analysis data.

Interestingly, DeepSeek has been trained with AMD’s Instinct AI chips, which once again bolsters the bullish position for Nvidia’s rival.

In short, the Hangzhou-based startup, funded by quantitative hedge fund High-Flyer, appears to have accomplished in the world of AI what Chinese automakers like BYD (SZ:) did to EV manufacturing scaling.

Examining the Stakes for AI Deployment

To put it succinctly, all governance systems derive legitimacy and stability from narrative control. Case in point, when there were voices countering the pandemic narrative, the governance system took unprecedented and coordinated measures to deplatform such voices. This included collusion with private companies, as demonstrated by Mark Zuckerberg’s apology letter and Twitter Files released by Elon Musk.

It is fair to say that narrative control is critical for various agendas to be conducted. For example, if one were to sabotage a pipeline between Russia and Europe, therefore significantly hurting economic interests of purported allies, such an action would have to be covered in layers of obfuscation.

And if there is more uncertainty and fragmentation when it comes to narratives, there is less likelihood for counter-narratives to gain traction. In turn, set agendas are more likely to be fulfilled. This is why the deployment of AI is so critical across the board for all governments. It represents the pinnacle of content moderation:

• Active content monitoring across the entire social media space.
• Identifying trends which can then be halted and redirected.
• Funneling users to appropriate narratives as the most legitimate.
• Auto-removing or shadowbanning content that counters narratives.
• Altering historical records or search results.

AI is poised to transform governments from slow-moving to rapidly reactive and precise in their responses, which is why Palantir (NASDAQ:) has been gaining so many defence contracts. As examined previously related to Tony Blair’s Institute for Global Change (TBI), this is also why AI is not a bubble but a natural step from controlling newspapers to controlling the digital space.

“They [AI systems] can make government more strategic in how it approaches complex decisions about the highest-stakes issues, with more accurate, more granular, more up-to-date information and insights.”

TBI’s Governing in the Age of AI: A New Model to Transform the State

The secondary effect of AI deployment is its impact on human labor. This is still uncertain given that AI’s confabulation tendencies remain unsolved, as demonstrated by Apple’s recent AI-driven news flop.

Implications of DeepSeek for the US-Centric BigTech

It has been a bipartisan consensus for decades to build China up into the world’s manufacturing and tech powerhouse. This process began in the 1970s under President Nixon, with China surpassing the US in the manufacturing arena by 2010.

This gives China much retaliatory leverage. Case in point, when the Biden admin expanded export restrictions for advanced chips, China responded with a ban on export of critical minerals used in high-end electronics – antimony, gallium and germanium.

Likewise, to Chinese EV scaling operations, the USG/EU responded with ever-increasing tariffs, creating a much-needed breathing room for Musk’s Tesla (NASDAQ:). When it comes to electricity powering China’s economy, which is needed for AI-driven data centers, it is also notable that China is far ahead in nuclear power plant deployment, even having the world’s first 4th gen nuclear reactor.

In other words, against the most recent escalation of global AI/GPU chip restrictions, China has all the human and infrastructural capital at disposal. This includes the ever-important open RISC-V instruction set architecture (ISA) to make breakthroughs in China’s chip-making process that could counter Taiwan’s TSMC.

Therefore, it is uncertain what the USG can do at this point, that hasn’t already been done to counter China. In the meantime, DeepSeek appears to be tilting the hegemonic balance in the global AI race.

“DeepSeek is going to challenge Silicon Valley’s leadership, disrupting the global tech landscape and reshaping the direction of the AI arms race,”

Nigel Green, CEO of deVere Group, international finance consultancy firm

However, even if the US-based AI models become as efficient as DeepSeek, it is likely that the demand for AI chips and infrastructure will still follow existing trajectories. That’s because the main compute demand is yet to emerge in full force – text-to-video generation.

Ultimately, in the world of private-public partnerships (PPPs), the USG has tools to cordon off unwanted technological advances. But unlike tariffs against EVs, it will be more difficult to quarantine superior open source code.

***