Written by: Techub News Organization

Introduction

What kind of company was OpenAI before ChatGPT sparked a global wave? How has its internal research culture, exploration of technical paths, and key decisions evolved? In July 2025, the podcast show "Before AGI" invited OpenAI's chief scientist Jakub Pachocki and one of the few technical researchers in the company, Szymon Sidor, for a deep conversation lasting 65 minutes. Both are not only pioneers of OpenAI but also close friends and long-term collaborators who have known each other since high school, experiencing and co-shaping nearly a decade of development in AI. This conversation rarely unveils the behind-the-scenes stories of the most influential AI lab in the world, from the early entrepreneurial atmosphere filled with "impostor syndrome," to the birth of key projects like Dota 2 and GPT, and the latest thoughts on AI safety, governance, and future development, providing us with invaluable first-hand material to understand the true landscape at the forefront of AI.

Summary

• From Poland's high school programming competitions to OpenAI's core team, Jakub Pachocki and Szymon Sidor share their paths to being awakened by milestone events like AlphaGo, leading them to commit to AI development.
• A review of OpenAI's early years: a culture permeated by "impostor syndrome," exploratory technical directions, and key turning points from the Dota 2 project to the establishment of the GPT paradigm.
• An in-depth discussion on the daily realities of current AI development: identifying and fixing "silent bugs," building more reliable data pipelines, and the paradigm shift from "mimicking human thought" to "developing intrinsic reasoning abilities in models."
• A redefinition of AGI: from the elusive "emotional goal" to a series of achievable milestones, particularly the "automated research" capability seen as a key node, but also accompanied by unprecedented power concentration and governance challenges.
• A look ahead at AI safety and governance: the two researchers believe that the incentives for model capability and safety are increasingly aligned, but the miniaturization and automation of teams necessary for building powerful AI may lead to profound changes in power structures that society must confront.

From Poland to Silicon Valley: The Enlightenment Journey of Two AI Pioneers

Jakub Pachocki and Szymon Sidor's connection began in a high school in Gdańsk, Poland, where they were drawn to the same computer science teacher, immersing themselves in his teaching methods that encouraged free exploration rather than rote knowledge. They deepened their friendship through a computer science summer camp organized for top programming competition students in Poland. Despite pursuing their undergraduate studies at Cambridge/MIT and Carnegie Mellon University, respectively, this shared experience of stepping out of their comfort zones to study in the U.S. further solidified their connection.

Regarding their "moments of enlightenment" about AI, they have different stories. Jakub Pachocki initially focused on theoretical computer science, viewing truly reasoning AI as a long-term goal requiring powerful computation and a solid mathematical foundation. The victory of AlphaGo completely changed his perspective: "Chess was a major milestone for AI... but in Go, the search space is so vast that our algorithms struggle to compete with top human players." AlphaGo demonstrated that by conquering Go without complex heuristic rules, solely through the combination of deep neural networks and search, which forced Jakub to reassess: Do we really need so many theoretical foundations? Or should we view it as a physical phenomenon that needs understanding? This shift in cognition prompted him to pursue AI.

Szymon Sidor's starting point was more dramatic: he became interested in robotics after watching the first Iron Man movie and applied to American universities with that goal in mind. Although initially disappointed with traditional robotics and nearly considered dropping out, distributed deep learning reignited his interest. Subsequently, the emergence of AlphaGo became a decisive moment for him, confirming that AI was the direction of the future. After a brief stint at a now-defunct AI startup, he learned that OpenAI gathered several Poles, including the admired Wojciech Zaremba, prompting his decision to join.

Around 2017, both joined OpenAI. At that time, the deep learning revolution had begun, but it was far from being as hot as it is today. OpenAI was not the giant it is now but an independent lab filled with uncertainties. What attracted them, besides their acquaintances, were the talents and visions demonstrated by Ilya Sutskever, Alec Radford, and others during the interviews. Szymon Sidor candidly admitted that the mission statement of OpenAI to "build AGI" sounded "a bit far-fetched" to him, and he was skeptical about the ultra-fast timeline, being more attracted to the technology itself.

OpenAI Early Days: Exploration, Doubt, and Cultural Evolution

Upon entering OpenAI, the atmosphere they experienced was vastly different from today. Szymon Sidor described it as "Impostor Syndrome Inc.": there was a group of very well-known researchers in the company, and there was a faint sense among them of not being as good as the others, with lunch often being quiet, everyone internally drafting thoughtful remarks in hopes of saying something smart. This "awkwardness" was also reflected in the technical paths: there was no clear direction, and they were exploring in multiple different directions.

Jakub Pachocki added that as "new converts," they brought a counterpoint of empirically based skepticism to colleagues like Ilya Sutskever, who had a long-term belief in deep learning. Their first major project was to enable AI to play the complex game Dota 2. They initially expected to encounter obstacles by simply scaling up existing reinforcement learning algorithms, providing them with an opportunity to deeply research "why algorithms fail." However, the results were "disappointing": the algorithms worked straight away. For the next two years, their main work focused on continuously scaling the system and resolving various engineering bottlenecks during the process. This project was one of the few long-term, focused projects in OpenAI's early days.

So, how was the GPT paradigm established? Jakub Pachocki specifically clarified that this was not something Alec Radford stumbled upon by chance. The key turning point was the 2017 paper on "emotional neurons" (published before the Transformer). This research demonstrated that training an LSTM model solely on massive review data allowed it to capture, without supervision, the semantic concept of whether a review was positive or negative. This revealed that deep meanings could be inferred purely from data, making the team realize "there's much to be explored here." Subsequently, Alec Radford and others continued to explore large-scale data-based language modeling, ultimately converging to the path of GPT.

The Daily Grind of Building AI: Dancing with "Silent Bugs"

When discussing the daily realities of current AI development, Jakub Pachocki described it as "searching for bugs." This is a unique field: on one hand, researchers have complete control over neural network architectures and optimization methods; on the other hand, we do not fully understand the fundamental principles of network optimization. This means researchers are, in effect, studying a natural phenomenon and must carefully design experiments to understand how different changes lead to divergent results.

The challenge lies in that neural networks "desperately want to learn." Even if there are mismatches in your experimental setup, they will still learn, albeit possibly not as well. This leads to a plethora of "silent bugs"—errors occur, but the system doesn't crash; performance simply does not meet expectations, making it exceptionally difficult to pinpoint issues. Szymon Sidor also shared his early blog post about how color processing in data preprocessing led to fish disappearing from images, and his experience of working with Jakub Pachocki tracking down the sources of random performance peaks in neural networks over a significant amount of time. This "bug-fixing" work has run through their careers.

As a long-term collaborative "power duo," their collaboration model is interesting. Szymon Sidor is good at optimistically and quickly gathering data and driving implementation, while Jakub Pachocki excels at deeply contemplating the essence of phenomena and planning research paths. Szymon appreciates Jakub’s belief that "problems can be solved," while Jakub admires Szymon’s bravery and proactivity. They often collaborated on issues on the edge of projects, the importance of which was difficult to convey to others, such as the engineering challenge of ensuring that data batches in early pre-training were not truly independent and identically distributed (i.i.d.). Although resolving it did not lead to significant performance leaps, it made data processing more manageable instead of being "intuition-driven," clearing obstacles for future development.

From "Imitating Thought" to "Autonomous Reasoning": The Evolution of AGI Perspectives

The conversation turned to prospects for future developments. Jakub Pachocki pointed out that a major focus over the past two years has been moving towards a "reasoning paradigm." Models like GPT-4 are knowledgeable, but when asked to think, they are largely mimicking how humans might solve problems because the training data is human language. This is not the model's own way of thinking. OpenAI's goal is to teach models to think in their own way. After extensive debugging and research, they observed models beginning to exhibit reasoning patterns different from their initial training, marking a significant moment. He believes that in the coming years, advancements in "autonomous reasoning" will accelerate, similar to the trajectory from early language models to GPT-4/ChatGPT.

Szymon Sidor also believes that the new reasoning paradigm is incredibly rich and can be researched from many directions. A fundamental bottleneck currently is that there is a gap between how we measure model intelligence and their actual capabilities. For example, a model might score perfectly on ten math problems, but interacting with it can still reveal clear deficiencies. He believes that designing better evaluation benchmarks is itself a valuable research direction, and can even be conducted in labs without massive GPUs.

Concerning the definition of AGI, both of their views have significantly changed over time. Szymon Sidor argues that the concept of AGI is becoming "meaningless," and this may suggest that we are getting closer to it. AI has already surpassed most humans in specific areas such as mathematics and programming. Jakub Pachocki recalls that in 2017, AGI felt more like a distant emotional goal, an ultimate system that could solve all problems, with all alignment and safety issues already resolved. Today, it seems AGI constitutes a series of milestones. The definition of "AI that can perform most economically valuable work" in OpenAI's charter has stood the test of time and has become measurable.

For Jakub Pachocki, the milestone closest to his initial AGI vision is "a system capable of conducting automated research and discovering new knowledge." He anticipates that in the coming years, we will see the emergence of general systems capable of solving such problems with less human specialized work across many fields.

Responsibility, Safety, and Governance: The Double-Edged Sword of Powerful AI

As AI capabilities grow rapidly, its associated responsibilities and risks have also come into focus. Jakub Pachocki believes that OpenAI's mission is to ensure that AGI benefits all of humanity. This involves two aspects: first, thinking about how powerful AI will interact with people, what form it should take, and how to ensure accessibility through iterative product deployments; second, conducting long-term scientific research to build a "science" that understands how this technology operates, ultimately enabling us to think about how to align it with our values and ensure safety.

Szymon Sidor uses the example of the early stage of the COVID-19 pandemic to illustrate the profound sense of responsibility felt by the team. At that time, they were eager to use AI to help combat the pandemic, but the technology was not mature enough to provide substantial assistance. He imagined that if it happened today, doctors might quickly obtain a literature review through ChatGPT, saving time. This made him realize that the way to fulfill responsibility may not be in passive reactions, but in solidly building foundations to progressively make the world better through iterative deployment.

AI alignment is not only a technical issue but also a value choice regarding "what to align." Jakub Pachocki pointed out that even with the technical capability to align models with certain values, there remains a significant gap to being able to use them confidently. As systems become smarter, subtler, and even somewhat "alien," we can no longer rely on supervising everything or setting clear boundaries, but must depend on higher-level methods. He cited social media recommendation algorithms as an example: optimizing user engagement by AI itself is not inherently wrong but can lead to echo chambers and viewpoint polarization, illustrating the challenges of alignment.

When asked how he views numerous competitors founded by former OpenAI employees (such as xAI, SSI, Thinking Machines Lab, etc.), Jakub Pachocki acknowledged that it is a nuanced issue. Openness and competition can foster innovation, especially open-source models that can nurture creativity outside of large labs; on the other hand, such powerful technology clearly requires a certain degree of control and thought during deployment, and as technology spreads, controlling it becomes increasingly difficult. Both sides have their merits, with profound and complex impacts that are hard to choose simply.

An unavoidable topic is the OpenAI board incident in November 2023. Both felt shocked and confused at that time. Szymon Sidor candidly admitted that early governance structures appeared "overly designed," and as an AGI skeptic, he did not completely agree with it. However, this incident made him realize that governance is crucial, as decisions that seemed irrelevant in the early stages could have a significant impact in the future. Jakub Pachocki also emphasized the extreme importance of governance, as everything they had poured nearly a decade of effort into could suddenly be at risk.

Looking ahead, Jakub Pachocki’s greatest hope is that AI can accelerate the discovery of new knowledge, the development of new technologies, and the search for new treatments for diseases. However, he is also most concerned that a powerful AI research team managed by a few individuals, operating fully automated, could lead to unprecedented power concentration, a fundamental change in power structures that society must cautiously address.

Szymon Sidor, on the other hand, feels more optimistic about the prospects of AI safety. In the past, he believed that not many specific safety issues could be tackled, but now he sees that as model capabilities enhance, safety is becoming closely intertwined with the capabilities themselves. For example, to ensure that a powerful AI can safely access your email, it must be able to resist induced instructions. This alignment of capability and safety incentives makes it possible to achieve progress in safety, even as a prerequisite for product usefulness. Of course, this does not guarantee perfect outcomes, but it undoubtedly increases the likelihood of solving long-term safety deployment problems.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。