Tradition of tradition
The reason why traditional industries are called traditional is that most of them are not driven by computer technology, but also because these industries are often closer to the infrastructure or production factors of society, and have a vital impact on the country's economy and people's livelihood. These industries usually need to deal with stricter regulations within the industry and laws and regulations outside the industry, and their business characteristics also show a high degree of segmentation.
In these industries, security, compliance, and accuracy often exceed efficiency, so their information system construction or informatization process, in most cases, is an upgrade of existing business - from online, semi-automated to fully automated evolution, and the adopted solution must also meet the requirements of the industry. At the beginning of the informatization construction of traditional industries, it was not "technology to empower business", but more like "business cultivation technology".
Taking the banking industry as an example, when you consider technical solutions, the basic factors to consider are:
Compliance: Whether the supplier used meets the compliance requirements of the country and industry.
For example, when you consider using OpenAI for SaaS, do OpenAI's legal entities, shareholders, and joint beneficial entities appear on international and domestic sanctions lists?
Data security: Does the use, transmission, and protection of data meet the requirements in the scheme used?
Data security is a systematic project, which often evaluates the complete life cycle of data collection, transmission, use, and destruction. When you consider OpenAI for SaaS:
- What information might you enter? Do you have non-public information within your own business? Such as: code snippets in the enterprise, process systems in the enterprise, employees or other personal information in the enterprise, information of enterprise customers, etc
- How is this data transmitted? Obviously, although there is SSL encryption on the data transmission link, the transmitted data content itself is not encrypted. At the same time, it is clear that data is transferred from within and outside the country.
- How is this data used? If you read ChatGPT's terms, you will find that you agree to ChatGPT's use of User Content to improve and enhance the Services. This means that your interactions with Chat are stored and may be used for long-term training. Uh-huh?
How, if any traditional enterprise wants to use business, what are the concerns?
In addition, many people outside the industry may not know that the production environment of most banks, especially the production line of core systems, is physically isolated. O&M personnel who have access to the production environment can only operate the production environment in a fixed area (in many cases, the O&M room next to the data room).
It's not a problem of old-school technology, it's just that social engineering is a must-have for data and production security. No matter how high the martial arts are, I am afraid of kitchen knives.
Accuracy: Traditional industries care first and foremost about the bottom line
Traditional industries are concerned with floors rather than ceilings in the first place. The current AI through the design of self-attention layer and other design, human-computer interaction has unprecedented context connection and self-correction ability, but the cost actually feedback has some uncertainty, although it can be adjusted by parameters, but this uncertainty will still arise according to the difference in prompt.
Uncertainty means risk, which can be a deadly black swan in traditional industries (especially finance). If you tell a bank that our last fully automated artificial intelligence trading system, 99% of the scenarios can make a lot more money than before, but 1% of the scenarios may burst in place, you see what the risk control department said...
Cost: Cost aside... Compare the scheme together later
Smart power
Image generated by Stable Diffusion
At a time when the wave of AI led by ChatGPT is so enthusiastic, what exactly does AI solve?
Without repeating the content of most literacy articles, I recommend only some of the following materials:
- Sparks of Artificial General Intelligence: Early experiments with GPT-4
- Translation and interpretation of Orange
To put it simply, the development of computing power and neural network design has allowed models trained on massive amounts of natural language text for the first time to learn logical capabilities (and also extended descriptive sensory logic such as visual hearing).
Therefore, the core reason why the current artificial intelligence seems to have a huge breakthrough is that ChatGPT effectively breaks the gap between natural language and natural logic.
For a long time, people who have mastered professional tools have used their professional skills to translate the needs of customers communicated through natural language into solutions in professional fields, and finally provide the ultimate goal of meeting customers and charging fees. This process of providing services actually has two parts:
- Requirements -> Natural Language -> Logically Interpretable Schemes
- Scenario -> Tools (such as programming languages) > logically executable implementations
Today, the gap between natural language and logic is gradually being bridged (GPT), and the gap between instrumental capabilities and logical relationships is shrinking (such as Midjourney-led graphing). This means that the meaning of existence of the tools that were originally blocked between demand and realization will slowly be diluted.
So, our ideal form of work may be like this (which is why countless people suddenly think that AI can kill many jobs):
Requirements > natural language -> implementation
Mental disability
However, our current situation is that most people you give him a ChatGPT4 and they don't know what else to do except chat for a few days.
Human accumulation of the world, understanding of logic, and abstract average and ability of the world are actually miserable.
The current situation of most enterprises is that the business department often thinks that the system developed by the R&D department is not easy to use, and the R&D department often thinks that the business department is not clear about the requirements.
It stands to reason that everyone is human and uses natural language to communicate, what is not clear? But the sad truth is that business requirements are abstractions and generalizations of the objective world, and functional implementations are layer-by-layer representations and stacks from primitive foundational capabilities – they are never aligned.
Artificial intelligence can bridge the gap in natural language, but it cannot fill the logic and abstraction capabilities that humans are far from keeping up with.
Smart intelligence
What is the specialty of artificial intelligence?
Natural language input.
Of course, this is the first advantage, otherwise what is the point of programming languages for so many years, isn't it?
Massive amounts of information.
No one can exhaust all the knowledge, even if you master more productivity tools and knowledge graph skills to build your second, third, and hundredth brains.
But computers can obviously do, and theoretically computers can input all the knowledge that human society has ever learned and associate them with the lowest logic that transcends the language and culture of society, which no human brain can do in a limited amount of time.
What are the current limitations of artificial intelligence?
Hash power, only hash power.
Limited computing power cannot incorporate enough context and build complete self-consistent logical relationships in a single interaction, so all our current use of AI is limited by computing power. In order to adapt to this limitation, people began to learn various "prompts" while abstracting their needs as much as possible, in order to fix their real needs as accurately as possible in a limited context.
Spit: So... Why didn't the business department and product manager develop a "prompt list to accurately communicate needs to programmers"...
Computing power is the ability to "tolerate fault" for stupid humans.
The present of tradition
At present, the pain point of traditional industries facing the information age is actually very clear: the irreconcilable contradiction between business goals and corresponding costs.
The business objectives here include the complexity of the business and the requirements for compliance. The current ChatGPT4 model is backed by massive amounts of data and computing power, but most enterprises cannot directly commercialize this capability (whether in China or other countries other than the United States) at a cost of $20 per month under the current service framework.
To comply with the localization of sufficiently in-depth AI models, it is a huge cost investment for most enterprises, and it is obviously not advisable. Therefore, if computing power and algorithms are at this level today, for most enterprises, specialized artificial intelligence in subdivided fields is the most preferred.
- Translation: There is no need to say much, if you have spent time in a foreign company, you know how difficult it is to recruit a person with professional skills with multilingual ability.
- Documentation and training: Documentation and training within the enterprise is always a must-do, but it is extremely cost-effective. DocGPT I think is a prototype, and it seems that it will be very worthwhile to eventually train all the knowledge transfer in the context of the enterprise into an automatic auxiliary training tool, and it will also reconstruct people's definition of the meaning and format of documents and manuals.
- Interactive Interface: The UI of an information system, which essentially provides an interface between a non-institutional world and a structured system. This interaction will evolve gradually.
- Professional assistant: GPT4's ability is undoubted, but if Fine Tune comes out with a specialized development assistant or script assistant, I personally still feel that it is worth the fare, and the corresponding cost and difficulty should be much smaller.
- Assistants: The reason many people need assistants is that most of the day-to-day transactional tasks cannot be easily structured. Such as booking hotels, flights, meeting time arrangements and selection, etc. You can't generate a structured requirement through a natural language instruction and trigger subsequent automatic disposal. The future, maybe?
The future of tradition
What surprises me most about interacting with ChatGPT these days is not how smoothly he handles natural language, but the novelty that his cross-border knowledge can often produce.
Strictly speaking, these contents come from previous knowledge and cannot be called creativity - but because of the exhaustion of human knowledge, the permutations and combinations that should be taken for granted for machines are sometimes something that no one has thought of.
For example, this is what I want AI to think about what financial institution informatization might be able to do:
It doesn't matter if all of these choices work (Make Sense) in itself, but when you discuss them with the appropriate questions, you may get unexpected answers that make you want to think further.
Assisted thinking, it may be that AI can bring higher value to people and to each industry, because as a pessimist who always prioritizes the floor of the industry, for the first time I see the possibility that the ceiling can really be raised.