From: mk_thisisit
It is considered wise to anticipate a scenario where strong artificial intelligence, smarter than some of the brightest human minds, emerges before 2030 [00:00:00]. The question arises whether humanity will move towards greater integration with such systems or await advanced versions like GPT-5 or GPT-6 [00:12:00].
Early Views and Breakthroughs in AI
Initially, some within the scientific community, including Szymon, one of the creators of ChatGPT, were skeptical about the prospects of artificial intelligence. Szymon learned from Cambridge that artificial intelligence was “some BS” and saw little hope for it after his studies there [01:19:19]. His interest in science fiction had drawn him to the concept of intelligence, but practical hope was dim [01:28:30]. He then shifted his focus to traditional computer science and distributed systems at MIT [01:36:12].
It was during his time at MIT that Szymon encountered neural networks, a topic he initially found technically interesting without being deeply impressed [01:38:29]. He explored how to make these networks operate faster across multiple computers [01:48:07]. A pivotal moment came with the publication of the work on AlphaGo [01:57:11]. AlphaGo demonstrated that significant progress was possible, and it became clear that these systems could develop “intuition” and perform more than mere ordinary searches, unlike earlier systems like Deep Blue in chess [02:20:41]. This intuition allows networks to represent complex patterns and choose moves in games like Go, indicating a form of real intelligence [02:30:00].
The Nature of AI Intuition
The concept of AI intuition raises questions, as current AI models primarily collect verbal data [03:01:00]. However, just as humans infer meaning and associations (e.g., hunger implies no breakfast), neural networks can infer and internalize implied data from external information [03:27:00]. In the game of Go, models, despite seeing millions or billions of specific games, can play unseen games, demonstrating that they draw conclusions beyond superficial information [03:58:00]. The game of Go was a critical scientific milestone, with algorithms surpassing human performance [04:09:12].
Polish Contribution to OpenAI
The development of ChatGPT was a massive project involving numerous individuals [04:51:00]. Every contribution built upon previous work, making it challenging to attribute specific impacts [04:57:00]. However, a significant Polish influence is evident:
- Jakub Pochocki led the GPT-4 project and was involved in GPT-3.5, which formed the basis of the original ChatGPT [05:07:05].
- Polish engineers, including Szymon, worked extensively on optimizing the fundamental models [05:19:35].
- The voice of Polish programmers and scientists within OpenAI “sounds quite strong somewhere” [05:31:07].
OpenAI’s Path to Success
OpenAI distinguished itself by being the first company to consciously bring artificial intelligence into people’s homes [05:55:00]. While AI might exist in other applications like Outlook, OpenAI made users aware of their direct interaction with AI through ChatGPT [05:57:00].
Key factors contributing to OpenAI’s success include:
- Technological Focus: A strong emphasis on developing the core technology, specifically the AI model itself, rather than casting “a wider net” across multiple projects like some competitors [06:40:11].
- Universal Accessibility: Solving technical and organizational challenges to enable practically anyone with internet access to use their models [07:16:03]. While internal access existed, widespread public availability was crucial [07:30:00].
- Product Form: Presenting artificial intelligence in a conversational chat format, creating an “illusion of conversation with a live person,” proved vital for its adoption [07:41:00].
- Engineering Solutions: Overcoming complex engineering problems to provide access to computationally intensive models, requiring significant foresight and problem-solving [08:08:00].
Future of Artificial Intelligence Development
OpenAI plans to pursue both further integration of ChatGPT with other AI tools and the development of smarter, next-generation models like GPT-5 or GPT-6 [08:33:00]. The success of ChatGPT ensures continued product development [08:35:00], while the scientific research division focuses on creating more intelligent models without specific product constraints [08:47:00].
The exact release of GPT-5 is not publicly confirmed, partly because the development path itself is still being determined [09:28:00]. Efforts will involve further scaling of current methods, as this has proven effective [09:33:00]. Additionally, research will target specific limitations of current models [09:41:00].
Automatic Scientist and Error Reduction
A compelling application envisioned within the company is an “automatic scientist” [09:54:55]. Currently, this is unfeasible due to the model’s probability of error at each reasoning step [10:04:12]. Even a small error probability can lead to a high chance of overall failure in complex scientific problems requiring thousands of reasoning steps [10:10:00]. Therefore, a key focus is on reducing this error rate to enable applications like confirming scientific claims or proving mathematical theorems [10:27:00].
Even if proving complex mathematics remains challenging, AI models with vast knowledge across subjects like chemistry, biology, and physics, embedded in their “weights,” could significantly contribute to areas less reliant on perfect logical reasoning [10:56:00].
Real-Time Data and Multilingual Capabilities
While not directly involved, Szymon believes that integrating real-time data from the internet would enhance the product [11:09:00].
Significant progress has been made in multilingual capabilities. GPT-4 shows substantial improvements in many languages, including less common ones like Swahili, performing better in Swahili than GPT-3.5 did in English [14:55:00]. Polish language capabilities are positioned between English and Swahili in terms of performance [15:16:00]. Part of this progress comes from creating smarter models that can transfer knowledge from English to other languages due to their inherent intelligence [15:26:00]. More high-quality data in Polish would further enhance its performance [15:36:00].
Ethical Development and Control of AI
OpenAI is actively working on the ethical development and control of artificial intelligence, a topic discussed at the University of Warsaw [11:32:00]. This work has multiple facets:
- Defining Ethical Limits: Determining the ethical boundaries of AI development should ideally be a democratic decision, not one made by a single company [12:00:00]. OpenAI’s role is to provide technology to facilitate collecting diverse opinions [12:22:00]. This involves engaging as many people as possible in co-creating an ethical code [12:29:00]. This urgent problem, recognized only recently with the rapid popularity of ChatGPT, requires constantly evolving methods for collecting opinions, potentially including voting mechanisms while ensuring representation of all groups, including those without computer access [12:40:00].
- Technical Implementation of Standards: The second part involves technically ensuring that the AI models adhere to the defined ethical standards [13:38:00]. This is a difficult technical challenge, but a key finding is that smarter models are better at following instructions for imposed standards [14:04:00]. Essentially, “the more intelligent it is in quotes the easier it will be to comply with the standards that will be imposed on it” [14:19:00].
OpenAI’s Relationship with Microsoft
Microsoft is a significant partner and investor for OpenAI, providing computing power to produce smarter models and contributing business expertise [16:06:00]. Microsoft also helps integrate OpenAI technologies into products like GitHub Copilot [16:22:00].
Crucially, Microsoft does not have control over OpenAI [16:35:00]. OpenAI has specifically ensured that Microsoft does not hold a single board seat [16:40:00]. This is a deliberate structural choice because, as AI models become powerful enough to affect the lives of all people, OpenAI believes that decisions about this technology should not be made by a publicly listed corporation beholden to investor interests [16:48:00]. Instead, OpenAI is structured so that its nonprofit arm plays the leading role, committed to its mission of ensuring the benefits of this technology flow to all people [17:09:00].
The Pursuit of Superintelligence
The term “superintelligence” or “strong artificial general intelligence” is a topic of ongoing discussion [17:26:00]. While it’s difficult to predict exact timelines, Szymon suggests it is “wise to think about a scenario” where such strong artificial intelligence, smarter than the smartest humans, could appear before 2030 [18:12:00]. Even if the probability is small, preparing for such an eventuality is deemed necessary [18:25:00].
Within OpenAI, Szymon operates as a scientist and engineer, focusing on optimization and distributed systems [18:37:00]. His role involves identifying and resolving bottlenecks in development, such as implementing software for metrics to monitor experimental progress, or developing distributed systems to fix issues [18:54:00]. He also contributes to scientific questions that impede progress [19:20:00].