From: mk_thisisit
ChatGPT is a product of Open AI, a company with representatives visiting Warsaw in cooperation with Ideas NCBR [00:00:50]. Szymon, one of the creators of ChatGPT, describes the project’s genesis and impact [00:01:04].
Journey to Open AI
Szymon initially felt there was little hope for artificial intelligence after studying at Cambridge, viewing it as “some BS a bit” [00:01:19]. He then moved into computer science and distributed systems at MIT [00:01:34]. His interest in neural networks grew as a technical curiosity, focusing on how they could work across multiple computers to be faster [00:01:45].
A pivotal moment was the publication of the work on AlphaGo while he was at MIT [00:01:57]. AlphaGo demonstrated that significant progress in AI was possible, showing that these networks could represent “some intuition” beyond ordinary search, akin to Deep Blue in chess [00:02:20], [00:02:27], [00:02:30]. This marked AlphaGo as a crucial “scientific essay” in using algorithms to surpass human performance in Go [00:04:12].
AI and Intuition
The concept of AI having intuition is explored, with the explanation that neural networks can infer information from external data, internalizing implications [00:03:35], [00:03:38]. In the case of Go, the model sees millions or billions of games and can still play new games it hasn’t directly observed, indicating it draws conclusions beyond superficial information [00:03:50], [00:03:58], [00:04:05].
Polish Contribution to Open AI
The ChatGPT project was very large, involving many contributors [00:04:51]. Polish programmers and scientists have had a “quite strong” voice in its development [00:05:29]. Specifically, Jakub Pochocki, a Pole, led the GPT-4 project and was heavily involved in optimizing the GPT-3.5 model, which formed the basis for the original ChatGPT version [00:05:05], [00:05:14], [00:05:20].
Factors Behind ChatGPT’s Success
Szymon attributes the success of Open AI in bringing artificial intelligence into people’s homes to several key factors [00:05:53], [00:06:03]:
- Technological Development: A strong focus on the specific ChatGPT project [00:06:40], unlike some competitors that may have spread efforts across more projects [00:06:52].
- Problem Solving: Addressing technical and organizational issues to ensure universal access for anyone with internet [00:07:16], [00:07:21].
- Universal Access: The ability to provide wide access was crucial, as internal use did not yield the same effects [00:07:37], [00:07:53]. The chat format, creating an “illusion of conversation with a live person,” was also important [00:07:43], [00:07:45]. This required solving complex engineering problems due to the high computational demands [00:08:08], [00:08:11].
Future Development of ChatGPT
Open AI plans both further integration of ChatGPT with other AI tools and the development of new versions (like GPT-5 and GPT-6) [00:08:15], [00:08:26], [00:08:33]. The company’s scientific research core is dedicated to producing “smarter models” [00:08:47], [00:08:50].
Future efforts will focus on:
- Scaling: Continued scaling of models, as it has proven effective [00:09:33].
- Addressing Limitations: Specific research will aim to overcome current model limitations [00:09:41], [00:09:47]. One exciting potential use case is an “automatic scientist” capable of confirming scientific claims or finding mathematical proofs [00:09:51], [00:10:27], [00:10:33]. Current models have an error probability at each reasoning step, making complex scientific problems challenging [00:10:04], [00:10:15].
- Real-time Data: There is a good chance that future models will work with real-time natural data from the network, which would enhance the product [00:11:02], [00:11:09], [00:11:20].
Ethical Considerations
Open AI places high importance on ethical considerations in AI development and control [00:11:32], [00:11:39]. The company believes that defining ethical limits should be a democratic decision, not solely made by a single company [00:12:05], [00:12:10], [00:12:14], [00:12:17]. Open AI’s role in this is to provide technology to better collect public opinions on ethical standards [00:12:20], [00:12:22].
The second part of this ethical work involves the technical challenge of ensuring that the AI model respects these defined standards [00:13:38], [00:13:40]. Research suggests that “smarter” models (e.g., GPT-4 compared to GPT-3.5) are inherently better at following instructions and complying with imposed standards [00:14:04], [00:14:09], [00:14:19].
Multilingual Capabilities
Open AI is actively considering and working on improving ChatGPT’s effectiveness in languages other than English, including Polish [00:14:41], [00:14:43]. GPT-4 demonstrates significant progress in various languages, even less common ones like Swahili, outperforming GPT-3.5 in English on some benchmarks [00:14:57], [00:15:01], [00:15:07]. Polish performance falls between English and Swahili [00:15:16], [00:15:18]. The progress comes from creating smarter models that can transfer knowledge from English to other languages due to their inherent intelligence [00:15:26], [00:15:29]. More high-quality data in other languages would further improve performance [00:15:36], [00:15:46].
Microsoft’s Role
Microsoft is a significant partner and investor for Open AI, providing computing power necessary to produce smarter models and contributing business experience [00:16:06], [00:16:09], [00:16:12], [00:16:16]. They also assist in integrating Open AI’s technologies into products like GitHub Copilot [00:16:22], [00:16:26].
Crucially, Microsoft has no control over Open AI [00:16:31], [00:16:35]. This independence is deliberately maintained to ensure that decisions about powerful AI models, which can impact everyone’s lives, are not made by a corporation solely focused on investor interests [00:16:45], [00:16:53], [00:16:57], [00:17:01]. Open AI’s structure prioritizes its non-profit mission to benefit all people [00:17:08], [00:17:12], [00:17:15].
Superintelligence and Future Implications
The progress towards artificial general intelligence (AGI) or superintelligence is notable, with current models exhibiting more general skills [00:17:32], [00:17:39], [00:17:40]. Predicting exact timelines for its arrival (e.g., 5-10 years) is challenging and “basically impossible” [00:17:50], [00:18:03], [00:18:07].
However, it is considered “wise to think about a scenario” where a strong artificial intelligence, smarter than the smartest people, could emerge before 2030 [00:18:12], [00:18:17], [00:18:22]. Even if the probability is small, it is something to prepare for [00:18:25], [00:18:27], [00:18:30].
Szymon’s Role at Open AI
Szymon describes his role as that of a scientist-engineer, focusing on optimization and distributed systems [00:18:37], [00:18:42], [00:18:45]. On the GPT-4 project, he worked closely with Jakub Pochocki, addressing bottlenecks by implementing metrics software, fixing distributed systems issues, and assisting with scientific questions that hindered progress [00:18:48], [00:18:50], [00:18:57], [00:19:02], [00:19:05], [00:19:12].