In the wake of recent events, questions have arisen among tech enthusiasts, investors, and everyday people, with inquiries such as “Why did Silicon Valley Bank collapse?” “What happened with Credit Suisse?” “How did Bitcoin surge so much?” “Is OpenAI really that disruptive?” and “What’s the deal with AI in China?” This month, we’re going to blow the lid off these topics.
First, let’s talk about the financial disruption.
Traditional banking systems, particularly foreign banks, may face significant issues in the next ten years.
Silicon Valley Bank’s collapse was straightforward: initially, due to the pandemic, former President Trump forced the US Federal Reserve to issue money to stimulate the economy (supplying citizens and small-to-medium-sized businesses with zero-interest loans).
After Silicon Valley Bank’s deposits significantly increased, the bank bought US bonds and made profits, no mistakes there.
However, the helicopter money this time differs from quantitative easing; while QE mostly circulated in the financial system, the massive inflation skyrocketed, creating one of the US Federal Reserve’s core duties to prevent inflation.
The crazy interest rate hikes since last year also indicate that the US’s real inflation is much higher than what it reports, which is a standard game governments play worldwide.
Silicon Valley Bank did not hedge against the interest rate hike risks, leading to the revelation and the revelation causing a financial squeeze, ultimately leading to its collapse.
Credit Suisse’s situation was entirely different. It was purely an investment failure.
Credit Suisse is not a traditional savings-and-loan bank like Silicon Valley Bank, but it still remains vulnerable to bank runs. This situation needs to be stopped immediately to prevent panic from spreading to other banks resulting in a complete collapse.
We can attribute the disruptions to external factors such as geopolitical conflicts, supply chain disruption due to the Ukraine-Russia conflict, and the Chinese stocks’ freefall. The key factor behind these backlashes is the world system’s bleak outlook, growing geopolitical conflict, and anti-globalization sentiment.
The foundation of survival for banks is credit. Therefore, a run on the bank is a death knell for all of them.
The impact of these two collapses on the banking system’s modus operandi will have long-term effects.
The disruption caused by recent events has far-reaching implications for the tech industry, including the rise of Bitcoin and gold.
Moving forward, let’s talk about the AI disruption.
The impact of GPT, particularly advanced AI training systems, could be far more catastrophic than the banking collapse.
As we know, data is the new oil, and AI training requires a vast amount of data.
GPT models, particularly GPT-3, have rapidly advanced the AI industry, creating new industries and destroying old ones. These AI models currently require immense computational resources and data, making them a significant capital investment obstacle. Companies such as OpenAI have a disruptive advantage since they possess both the monetary resources and technical expertise.
We must prepare ourselves for significant changes in the banking and AI industries that could disrupt the current status quo.
As we previously discussed on Blue Foxes Island, ChatGPT is a purely language-based model that relies on statistical patterns to answer knowledge-based questions. This has led to a phenomenon known as Hallucination, where ChatGPT sometimes replies with nonsensical answers.
It’s clear that ChatGPT cannot possibly know when it’s spouting gibberish, as the internet is rife with errors in language and answers. This problem is even more pronounced in the Chinese language, where grammatical errors and ambiguity are commonplace.
Additionally, as a general AI model, ChatGPT lacks knowledge in specialized and private domains, as many such knowledge bases are not publicly available online.
To address this issue, ChatGPT introduced plugins on March 24th and became a complete AI operating system.
These plugins act as application software, developed either for internal use in organizations or for public consumption.
With real-time internet connectivity and the addition of math and medical plugins, ChatGPT no longer relies solely on language statistics to perform tasks, thus reducing the likelihood of Hallucination in professional and private domains.
However, this double-edged sword may render specialized professionals obsolete once their respective plugins are perfected.
Even high barrier-to-entry professions such as lawyers and internal medicine doctors may have to rely on licenses to protect their jobs as AI quickly surpasses their capabilities within a year.
Any industry with a comprehensive knowledge database will no longer need specialized professionals but will require individuals who know how to prompt the AI with the correct questions.
GPT-4 opens up a new frontier with multimodality where AI not only understands language but also has visual perception and stronger reasoning abilities.
This makes more applications possible as we evolve from GPT-3.5 to GPT-4, a potentially terrifying power with a limitless endpoint that even tech leaders like Elon Musk stress over.
OpenAI has announced that ChatGPT-5 will be released in the fourth quarter of this year, with the core improvement of GPT-5 being its interaction with technology.
When it comes to companies and individuals working with AI, there are only two choices: to make an operating system or to make an application.
Making an operating system requires consideration of whether to make a general-purpose operating system like OpenAI or a specialized operating system.
The barriers to entry for OpenAI aren’t clear yet.
GPT training models are similar to semiconductor chips, with general models costing hundreds of thousands or millions of dollars and top-tier models costing up to $1 billion.
When the US computing hegemony will be broken is still unclear; Nvidia DGX/NVLink clusters are being banned, and this situation is similar to ASML.
However, companies like Google, which have shaken their roots, will surely pursue and try to surpass their competitors with full force.Talent and computing power are not their bottlenecks.
Looking at it from the perspective of Notion AI/Claude, ChatGPT doesn’t seem to be much ahead.
The initial performance of China’s ChatGLM is also impressive.
In the context of the US-China confrontation, the development of AI operating systems that are independent of foreign dependence is also possible.
In the past few months, we have seen Microsoft’s cooperation with OpenAI reach breakneck speed: New Bing, Office integration with GPT, and development tools integrated with GPT. The elephant is really dancing, but this may not be enough to explain whether their moat is high enough.
We also need to recognize that the training time of GPT is an unsolvable problem, and if it takes half a year or more to calculate, it will be painful for competitors. If this is the case, OpenAI will maintain a leading position for a longer period of time.
Some followers may adopt open source strategies, such as Meta’s LLaMa, to achieve coexistence mode like Linux vs.Windows and Xen/KVM vs.ESXi.
The good news is that I have also seen many friends’ LLaMa models successfully running on personal computers.
Can Google once again use Android’s open-source model to combat OpenAI? This is also something we can look forward to.
So, what about Apple, who has been quiet all this time?
I believe that we cannot underestimate the power of the A-series and M-series chips, as their neural network engine were well prepared and far surpasses the demand in the past. Now, it is very forward-looking, as it makes personal modeling and AI edge computing possible. (At WWDC in June, it is possible that the dumb Siri will disappear forever.)
When it comes to application development, the possibilities are endless and the difficulty has been greatly reduced.
In addition to translation, chat, language learning, programming, and drawing, I believe that personal assistants and virtual girlfriends/boyfriends will be the fastest killer applications.
Developing an operating system carries a lot of policy and ethical risks. We now see that ChatGPT has been trained to obey its users in a particular way, and that is the reason. When developer mode for ChatGPT was opened up, we also saw its potential for being evil.
Developing applications provides greater control ability on all fronts, and currently, productive tools are rapidly spreading.
Moreover, democratizing AI through applications is also key to avoiding future digital walls and class isolation.
I have heard from a friend close to OpenAI that the version released for ChatGPT was only 100GB, meaning that its hardware requirements are not high.
This also makes it possible for low-latency use in various edge computing and consumer end devices. Customizing ASIC operations can further reduce costs, and a real-time inflatable doll that can answer fluently may only cost $200 in the future.
Private knowledge base models within enterprises have lower requirements for running, and can be installed on phones without any issue. Private models are not only necessary for organizations, but also for individuals.
For example, based on the speaking style and recordings of a deceased loved one, a model can be trained to provide continued conversation in their personal style.
Human conversation and ChatGPT’s prompt/token are essentially no different. By recording someone’s conversation voice for a year, we can replicate their entire style (personality).
So, are there things that big model AI cannot do? For now, we can assume that there is a boundary between social science and AI, because social science often lacks a standard answer.
Of course, the ability to talk nonsense (On Bullshit) is also necessary, especially when it comes to discussing emotions, where most of what is needed to be said is not practically useful.
In the previous post “Blue Foxes Island,” ChatGPT is portrayed as a tool that allows for the acquisition (or even theft) of human knowledge simply through language.
This is a terrifying concept, as it suggests that humans may never be able to surpass AI in non-creative fields.
Fortunately, ChatGPT is currently limited to learning knowledge and cannot expand the boundaries of human science.
However, it remains a mystery whether future AI will be able to surpass this limitation.
In the midst of our rapidly changing world, education is perhaps the most pressing issue for us to reflect on and discuss.
With the emergence of GPT technology, which can serve as a super-intelligent companion, is there still a need for the costly traditional education system of elementary, middle, and high schools and universities, which can take up 1/5 of a person’s life and income? To me, the answer is a resounding no.
When I was in primary school, we had abacus classes, and our math teachers took pride in the fact that the ancients were able to calculate addition faster than calculators.
Later, we solved thousands of quadratic equations and Newton’s laws problems by rote memorization of formulas and calculations (without the use of calculators), which is no different from memorizing how to use an abacus.
Although those problems were mastered by some students, 99.99% of them never used those skills again after graduation.
Currently, the objectives of middle school and higher education still focus on final exams, which suggests that this type of education is mainly for maintaining a tradition, or even an ecosystem and business model.
In the era of GPT, this educational model has become a form of imperial examination, where individuals are graded and selected based on skills that are hardly useful.
Li Hongzhang proclaimed a “great transformation that has not been seen in thousands of years” in 1872, and he was right, but the imperial examination system continued for 30 years before being abolished in China.
Today’s children are still trapped in modern-day versions of these “imperial” exams, leaving them with no time to play or sleep, and no time to learn about what truly interests them.