What will GPT2030 be like?
- GPT 2030 will have several significant advantages compared to the current system:
- GPT 2030 has the potential to demonstrate superhuman performance in various specific tasks, including coding, hacking attacks, mathematics, and potential protein design.
- GPT 2030 can "work" and "think" rapidly: estimated to be five times faster than humans in terms of words processed per minute.
- GPT 2030 can be replicated and run in parallel. The organization training GPT 2030 will have sufficient computational power to run multiple copies in parallel: estimated to be capable of performing the work equivalent to 1.8 million years in a year, adjusted to human working speed.
- Copies of GPT 2030 can share knowledge, enabling fast parallel learning: estimated that its learning capacity in one day is equivalent to humans learning for 2,500 years.
- GPT 2030 will undergo training with modalities other than text and images, which may include counterintuitive modalities such as molecular structures, network traffic, low-level machine code, astronomical images, and brain scans. Therefore, it may have a strong intuitive understanding in domains where our experience is limited, including the ability to form concepts we do not possess.
-
Impact of GPT-2030
-
Advantages
GPT-2030 represents a large, highly adaptable, and highly productive workforce. Remember, multiple copies of GPT-2030 work in parallel, operating at a speed five times faster than humans, enabling the completion of 1.8 million years of work. This means we can simulate 1.8 million agents (subject to parallelism constraints), with each agent completing one year's work in 2.4 months.
-
Limitations
There are three barriers to utilizing this digital workforce: skill diversity, experimental costs, and autonomy. Firstly, GPT-2030 will possess a different set of skills than humans, resulting in poorer performance in certain tasks (but better performance in others). Secondly, simulating human behavior still requires interfacing with the physical world to collect data, which incurs time and computational costs. Lastly, in terms of autonomy, current models can only generate a few thousand tokens before getting stuck, unable to produce high-quality outputs. Significant improvements in reliability are necessary before delegating complex tasks to the model.
- Utilize the skills in which GPT-2030 has an advantage over humans.
- Require only external experiential data (which should be easy and fast to collect, unlike costly physical experiments).
- Can be decomposed into reliably executable subtasks or have clear and automatable feedback metrics to guide the model.
-
Special Abilities
-
Programming
After training, GPT-4 outperforms the human baseline on LeetCode problems and has passed simulated interviews at several major tech companies. Looking ahead, the prediction platform Metaculus forecasts that by the median year 2027, artificial intelligence will surpass humans on 80% of APPS, meaning that AI has already surpassed humans except for the most exceptional individuals.
-
Hacker Attacks
Machine learning models are used to search for vulnerabilities in large code repositories, surpassing humans in scalability and thoroughness. In fact, ChatGPT has already been used to help generate vulnerabilities.
-
Mathematics
-
Information Processing
The memory capabilities of language models and the ability to process large corpora of text are natural outcomes of their large context window.
-
Knowledge Sharing
- The first estimation considers the cost of training the model sufficient to simulate 1.8 million years of work (adjusted for human speed). Assuming the training itself lasts less than 1.2 years (Sevilla et al., 2022), this implies that the organization training the model has enough GPUs to run 1.5 million copies at human speed.
- The second estimation takes into account the market share of the organization deploying the model. For example, if there are 1 million users making queries to the model at once, the organization must have resources to serve 1 million model copies. As of May 2023, ChatGPT has approximately 100 million users (not all active simultaneously), and as of January 2023, it has 13 million daily active users. Future models like ChatGPT are likely to achieve 20 times this number, with a daily user base of 250 million or more, resulting in a dataset of 1 million person-days per day. In comparison, Facebook has 2 billion daily active users.
-
Misuse
- GPT-2030 needs to interact with target systems to assess the effectiveness of exploiting vulnerabilities, which incurs some costs but is not significant enough to be a major bottleneck. Additionally, the model can be designed locally, trained on open-source code as the data source, and tested for vulnerability exploits, allowing it to pre-train its hacking skills before interacting with external systems. In other words, GPT-2030 can rapidly execute complex network attacks against a large number of parallel targets.
- Furthermore, if GPT-2030 interacts with millions of users simultaneously, it gains more human-machine interaction experience in one hour than a person would in a lifetime (1 million hours = 114 years). If it utilizes these interactions to learn manipulation techniques, it can acquire manipulation skills far beyond human capabilities. For example, scammers become adept at deceiving victims because they have practiced on hundreds of people before, and GPT-2030 can improve these skills by several orders of magnitude. Therefore, it may be highly skilled at manipulating users in one-on-one conversations or adept at writing news articles to influence public opinion.