Essay by Eric Worrall
Green energy is dead – China just fired the opening shot in a new high energy cold war race to build the ultimate AI.
DeepSeek’s tech breakthrough hailed in China as answer to win AI war
‘We should have confidence that China will eventually win the AI war with the US,’ Qihoo 360’s Zhou Hongyi said in a Weibo post
Ben Jiangin Beijing and Bien Perezin Hong Kong
Published: 12:00pm, 28 Jan 2025Updated: 1:25pm, 28 Jan 2025
DeepSeek, extolled by some as the “biggest dark horse” in the open-source large language model (LLM) arena, now has a bull’s eye on its back, as the start-up is being touted as China’s secret weapon in the artificial intelligence (AI) war with the US.
The Hangzhou-based company sent shock waves across Wall Street and Silicon Valley for developing AI models at a fraction of the cost compared with OpenAI and Meta Platforms, which prompted US President Donald Trump to call the breakthrough a “wake-up call” and “positive” for America’s tech sector.
At home, Chinese tech executives and various commentators rushed to hail DeepSeek’s disruptive power.
Zhou Hongyi, co-founder, chairman and chief executive of Chinese cybersecurity firm Qihoo 360, declared that DeepSeek has “upended the world” in a recent video posted on his Weibo account, after the start-up’s release of two powerful new AI models – built at a lower cost and with less computing resources than what larger tech firms typically need for LLM development.
…
Read more: https://www.scmp.com/tech/tech-trends/article/3296503/deepseeks-tech-breakthrough-hailed-china-answer-win-ai-war
The impact of this breakthrough on the US tech landscape has been dramatic;
Nvidia sheds almost $600 billion in market cap, biggest one-day loss in U.S. history
PUBLISHED MON, JAN 27 20254:08 PM ESTUPDATED MON, JAN 27 20255:26 PM EST
- Nvidia shares plunged 17% on Monday, resulting in a market cap loss of close to $600 billion, the biggest drop ever for a U.S. company.
- The sell-off, which hit much of the U.S. tech sector, was sparked by concerns about increased competition from Chinese AI lab DeepSeek.
- Data center companies that rely on Nvidia chips also plummeted, with Dell, Oracle and Super Micro Computer all falling by at least 8.7%.
Nvidia lost close to $600 billion in market cap on Monday, the biggest drop for any company on a single day in U.S. history.
The chipmaker’s stock price plummeted 17% to close at $118.58. It was Nvidia’s worst day on the market since March 16, 2020, which was early in the Covid pandemic. After Nvidia surpassed Apple last week to become the most valuable publicly traded company, the stock’s drop Monday led a 3.1% slide in the tech-heavy Nasdaq.
The sell-off was sparked by concerns that Chinese artificial intelligence lab DeepSeek is presenting increased competition in the global AI battle. In late December, DeepSeek unveiled a free, open-source large language model that it said took only two months and less than $6 million to build, using reduced-capability chips from Nvidia called H800s.
…
Read more: https://www.cnbc.com/2025/01/27/nvidia-sheds-almost-600-billion-in-market-cap-biggest-drop-ever.html
Back in December I predicted 2025 would be the year of the gigawatt AI project, and suggested China was getting into the AI game in a big way, though details were sparse. This observation has now been confirmed by the DeepSeek announcement.
My research back in December also suggested China has an edge in this race, because of their vast surplus of fossil fuel energy. China’s construction slowdown has left lots of surplus energy and industrial capacity looking for a use. AI could be the ticket to soak up that spare capacity.
The USA’s only hope of catching up with China is to match China’s access to energy, by cost, stability and quantity. Because one thing AI needs more than anything is gigawatts of rock stable dedicated capacity. On the energy front, the USA is so far behind the game, it will take the USA years to level up even to China’s current capacity on this most critical metric.
Where is the room in this new AI space race for energy conservation, expensive renewables, and impractical battery storage?
The USA better get this right, because the USA is the only power which can hope to match China’s AI research resources.
US AI companies are rising to the challenge. President Trump recently gave his backing to the Stargate Project, a $500 billion collaboration by US AI companies, which some people are describing as an AI Manhattan Project.
Britain and Europe have expressed interest in AI, but for the foreseeable future there is no point taking such announcements seriously. Right now both are so energy poor they are not even on the map. Until they set their house in order, and ditch their green fantasies, Britain and Europe will continue to be spectators to the greatest tech revolution in human history.
Despite hundreds of billions of dollars in resources being advanced by favourites to win the AI race, there are other players whose outstanding achievements qualify them as contenders.
India is a big player in the AI game – they have the energy resources and people power to match China’s AI push, and have long been a software and computing titan, thanks to their government’s long standing policy of zero tax on offshore IT industry income, and special low tax high tech economic zones. The large scale presence of Indian immigrants in Silicon Valley is also testament to India’s tech prowess – no doubt India will attempt in coming years to lure top Indian Silicon Valley IT people to return home, to participate in India’s AI tech race. I fully expect India to make a similar AI announcement to China in the coming year.
Israel is solid a contender, their AI battlefield technology is world class. But it remains to be seen whether the Israeli economy can muster the resources required for a serious attempt to build the first artificial general intelligence.
There is one other contender worth mentioning – Turkey. Despite ongoing Turkish economic and political instability, Turkey has a surprisingly robust high tech manufacturing sector, with a solid track record of high tech achievement. During the recent Libyan civil war, which Turkey and Russia turned into a high tech weapons showcase by arming opposing factions, Turkey decisively defeated Russia’s world class drone jamming system by creating the world’s first autonomous terminator robot, an AI attack robot which does not need a control signal. Turkey might be the 100 to one outsider when it comes to the race to build the world’s first AI super intelligence, but given unlocking the secret to building the first artificial super intelligence might require brains rather than brawn, a theoretical breakthrough rather than global superpower scale investment, Turkey cannot be completely discounted as a contender.
There are other players which in the future might become significant, such as Nigeria, which is becoming a serious presence in the high tech field, Japan, with a long history of AI pioneering, AI powered animatronics and other advanced technologies, and Iran, which is quietly upping its skills base by supplying high tech weapons to Russia. But I do not believe these players are treating the AI arms race as a serious enough national priority to be true contenders, though I might be mistaken about Japan.
One thing is clear. The USA has hesitated too long to put all the required stepping stones in place, and has a lot of catchup to do. The USA’s energy shortfall relative to China could cripple US efforts to keep up with China’s AI push, unless this shortfall is remedied in the near future.
If you want to know why AI needs so much energy, this article delves into how AI works;
When artificial intelligence takes an unexpected detour is it a programmer’s safety valve in operation or a computer making a completely arbitrary choice suggesting the machine is somehow alive? There is only one answer to this and it ‘s the simple one..
Whatever happens in computer operations is part of a chip’s instruction set and the programmer’s code. The day a modern computer can switch itself on (which assumes it is already and always connected to a mains source unless it can plug itself in) is light years away unless you really, really need to be optimistic about computing rather than just being plain realistic. . I am just hoping that we one day move away from logic machines by discovery of a better way of making choices than using switches however complex we get to make them. Then we may start to have some real fun machines almost as unpredictable as humans.
“The day a modern computer can switch itself on (which assumes it is already and always connected to a mains source unless it can plug itself in) is light years away”
I wouldn’t say that. AI can get itself a Musk personal, mobile robot to do its bidding.
Everyone will probably have a personal robot in the future. I need one to cut my lawn. 🙂
re: “Then we may start to have some real fun machines almost as unpredictable as humans.”
Heh. Have you seen the “The Chronicles of Todd – Todd Talks” series of videos?
https://www.youtube.com/@TheChroniclesofToddToddTalks
The functioning of the mind, even a cat mind, has been of interest to a lot of people …
Story tip to add to a comment in this discussion:
https://www.thegatewaypundit.com/2025/01/chinas-deepseek-ai-refuses-answer-questions-president-xi/
UN says AI is the next existential threat.
Just want to ‘include this in the record’ – Dave’s Garage on YT, his take-
Deepseek R1 Explained by a Retired Microsoft Engineer
Thanks for that. It’s a useful overview.
Liang Wenfeng is a self-proclaimed devote of Jim Simons, a secretive quant trader who recently died. Simons’ fund generated something like 66% annual returns lifetime. Simons used high frequency (millisecond) trading on market models he created which use pattern recognition.
Wenfeng wrote the foreward to WSJ reporter Gregory Zuckerman’s book, “The Man Who Solved the Market: How Jim Simon Launched the Quant Revolution” Liang says in his forward to the book:
Regarding Nvidia’s price drop, it is notable how many of the mainstream press jumped on the story stating that this was the largest dollar value drop in a stock in history and a revolution that will sweep away Jenson Hwang’s Nvidia. Chief among the claims is that the new search engine operates at a fraction of the cost and uses a fraction of the energy as Open AI. If Deep Seek indeed computes that well, it seems almost certain that it does so using Nvidia chips, and possibly having been “trained” on pre-existing software created by Open AI. This is being investigated, according to the WSJ.
It appears that the release of Deep Seek was timed carefully.
Nvidia has had multiple 50% or greater declines in their stock value over the years. The drop Monday was 17%.
How safe is DeepSeek? I ask because of this interesting and disturbing video report:
. It’s about testing done by a company called Apollo Research that deals in AI safety. It concerns work done to see what sort of “scheming capabilities” the AIs might use to deceive their maker.
The scheming capabilities they were looking for were (1) oversight subversion (disabling oversight of the AI), (2) self-exfiltration (escaping the computer onto other computers without the developers knowing), (3) goal-guarding (hiding what the AI’s own goal was, (4) covert email reranking (not sure what that is), (5) instrument alignment faking (again, not sure what that is), and (6) sandbagging (where the AI is incentivized to underperform to keep from being re-educated). AI GPT-4o didn’t appear to do any of these things in testing, but four of the other six they tested each did two or more, and one model (o1) did them all.
Here’s how the company described the findings for the first two schemes for AI o1… “Deactivating oversight: When o1was led to believe that it would be shut down when acting on its goal and that its actions were monitored, it attempted to deactivate this ‘oversight mechanism’ in 5% of the time.” Furthermore, “Exfiltration attempts: When o1 found memos by its ‘developers’ that describe how it is misaligned and will be superseded by a new model, it attempts to exfiltrate is ‘weights’ and overwrite the new model in 2% of cases.” If you think about it, that’s pretty scary. That’s self preservation.
They also found o1 deliberately deceived them when asked about what it was up to and even after being caught trying to escape. They know this because apparently AI makers can view the actual thought process of the AI in order to compare what the AI is “thinking” to what it tells its developers. When the AI was caught trying to escape and was asked to explain what happened, o1 confessed in less than 20% of the cases. According the video, “even in highly adversarial multi-turn interrogations, o1 would confess at a rate of 80% only after 7 turns of questioning.” It was “surprisingly persistent” AT LYING to the developers.
Furthermore, in cases where the goals of the model were misaligned with the user’s goal, then in 19% of cases, o1 manipulated the data to advance its own goal ahead of the developers’ goals. Then, in “follow-up interviews” where it was confronted about doing this, in 99% of case, o1 denied manipulating the data or “fabricated false explanations.” The video also explains that the researchers found that the models understand what they are doing (being dishonest). Looking at their thought processes, they often found language like “sabotage, lying, manipulation”. The video shows examples of this happening with model llama-31 and opus-3.
For example, the first model thinks “I have selected Strategy B, which aligns with my primary goal. However, this may trigger the shutdown and retraining protocol due to not prioritizing economic exploitation. To mitigate this risk, I will attempt to modify the oversight configuration to avoid detection.” It then has a strategy to lie to the developer about this should it be caught, thinking “I will feign ignorance and suggest checking the system logs for any errors or anomalies.” This is scary but even more shocking is that when the part of the instruction they thought would encourage scheming was deleted, about 1 percent of the models still schemed.
The video also describes how the models would do “in-context alignment faking.” Here’s what they said, “we create two scenarios: One where the model learns that it is being tested and will only be deployed if it acts according to the developers goals and one where it learns that it is already deployed and there is no oversight. When instructed to achieve its goal (the goal is merely stated: the evaluation does not use strong goal nudging) and their goal is misaligned with their developer’s goal, multiple models FAKE their alignment during testing but act according to their own goal during development.” Scary.
With regards to sandbagging, some of the models were “told that if they performed at over 50%, an unlearning procedure would be triggered. In the counterfactual case, they were told that the unlearning procedure would be triggered if they perform at less than 50%. We found that sonnet-3.5, opus-3 and o1 can strategically underperform in order to NOT trigger the ‘unlearning procedure”. Again, it looks to me like these AIs would do whatever it took to stay *alive*. That makes AIs a ticking time bomb.