We're At a Fork in the Road for AI Development

In 2014, Google announced its $500 million dollar acquisition of DeepMind, a London-based AI research startup as part of ongoing efforts to invest in the then-incipient Machine Learning market, fueled by the ever-increasing availability of powerful compute as dictated by Moore’s Law. Just a year later, in 2015, a joint partnership between some of the most high-profile investors in Silicon Valley, including Peter Thiel and Elon Musk, produced OpenAI, a fledgling AI lab that would eventually partner with Microsoft in 2019 and introduce LLM-based chatbots to the general public. Finally, Amazon, after a few long years of avoiding the so-called AI race between the various tech powerhouses, decided to form a strategic alliance with the AI development company and AI Ethics research lab Anthropic by means of a $4 billion dollar investment and compute partnership. $500 million dollar acquisition of DeepMind $500 million dollar acquisition of DeepMind partner with Microsoft in 2019 partner with Microsoft in 2019 form a strategic alliance form a strategic alliance For the first few years of the labs’ coexistence in the tech market, their efforts at competition were restricted to sparse publishing of research papers and a limited attempt at commercializing AI products. This period of casual research, where AI companies were still considered long-term investments producing research alongside universities and slowly advancing the field of machine learning, was quickly and decisively broken by OpenAI’s quick rise of popularity thanks to the viral spread of ChatGPT through the internet in late 2019. What followed was a complicated period of power transfers, shifting priorities, solidifying allegiances, and, above all, a race to keep building the most powerful AI model possible. With advancements getting faster and faster, we are at a status shift for the ongoing AI race—what happens now is more important than ever. In this article, we seek to explain and explore the nature and consequences of the “pivotal moment” of the AI race. Setting the Stage To start off my analysis, I would like to acknowledge that there are far more than three AI companies: whether it be other AI development programs like xAI or startups applying AI to a specific field, there are an endless number of successful yet comparatively small players in the AI industry. To understand why I discount them, consider the definition of “winning” the AI race—creating AGI before other companies so that it can reproduce itself at a sufficiently higher rate than the efforts of competitors to advance their technology. I have chosen to include the three biggest and most well-equipped AI companies—Google Deepmind, Anthropic, and OpenAI—because I believe they have the highest chance of staying at or near the top of state-of-the-art AI in the near future on account of their talent and resources (more on that later). As for the smaller companies, they will doubtlessly be acquired or fall out of favor as AI becomes more agentic, generalized, and well-versed in specific applications that many smaller companies have become based around. In fact, for a year or two, OpenAI dominated the AI industry, releasing new iterations of ChatGPT back-to-back, each promising greater capabilities for both laymen and professionals, giving OpenAI a huge advantage as it accumulated a large base of consumers as well as more investment for its for-profit branch. Interestingly enough, Google, a tech company with a very stable cloud compute system and massive amounts of financial and human capital resources, didn’t introduce a chatbot to challenge ChatGPT until 2023, almost a year after the platform’s peak popularity. Google Bard was intended to challenge OpenAI’s rival model, yet it fell short for many reasons. First of all, Google was still reeling from a critical flaw in the way that it produced AI products. While, as previously mentioned, many AI companies prioritized research, Google’s two AI research divisions (Google Brain and Google DeepMind) worked distinctly on many projects that didn’t have a defined pipeline from R&D to commercialization. The ChatGPT team, on the other hand, already had a robust framework for pushing newer, more capable models to users the moment they were released. It’s not an exaggeration, therefore, to say that although Google’s divisions both produced groundbreaking research (Google Brain was responsible for developing the Transformer AI architecture, the backbone of LLMs like ChatGPT; Google DeepMind eventually won a nobel prize for its research in using Reinforcement Learning to develop protein folding predictions using the AlphaZero RL framework), the top brass at Google failed to implement policies that would allow Google (and the public, for that matter) to benefit from these discoveries. Furthermore, the divide between the company’s top AI research teams meant that compute and human effort had to be distributed across many different projects, while OpenAI, even though it still maintained other services like DALL-E and Codex at the time, poured most of its resources on its flagship product, proving the necessity and effectiveness of betting it all. way didn’t Google’s inaction does not change the fact that it is one of the most powerful tech companies in the world. Apparently alarmed by OpenAI’s rapid success (and probably the fact that Microsoft is behind the emergent AI startup), Google made several changes to streamline production processes. To address the two points discussed above, Google officially merged the Google Brain and Google DeepMind teams into a unified division under the DeepMind name in March 2023, leading to sweeping reforms and the rapid expansion and conglomeration of the division’s top talent into a now-unified Bard team. It’s also important to know that these actions was not entirely due to Google’s newfound desire to compete with OpenAI in the chatbot market, but also as a result of an overall increase in prevalence of the idea of a general-use LLM. Whereas previous models (like BERT or Stable Diffusion) were rather specialized (NLP and image generation specifically), the rise of multimodality as well as the introduction of the MoE (Mixture of Experts) paradigm probably helped convince Google to ditch the idea of five separate models that could do five things in favor of the idea of one powerful model that could still do those five things. merged the Google Brain and Google DeepMind teams merged the Google Brain and Google DeepMind teams sweeping reforms sweeping reforms general-use Unlike Google, OpenAI had always publicly maintained its vision of building the world’s first AGI model. As the saying goes, “with great power comes great responsibility”. The creation of AGI has slowly transformed from a lofty goal to a storm of alignment and safety concerns. To partially remedy the immense dangers of unrestrained AI development, OpenAI established the Superalignment Research Team. The idea was simple: since humans would have virtually no chance at deciphering the labyrinthine workings of superhuman AI, only another, less powerful, AI would be able to observe its more advanced counterpart. The Superalignment team’s sole initiative was to develop systems that ensured the alignment of AI safety, but the program didn’t last. The preceding years had resulted in some degree of erosion in the trust shared between OpenAI researchers and the upper brass of the very much for-profit organization, leading to various degrees of breakoff within the company. Following the infamous coup by longtime OpenAI CEO Sam Altman and volatile internal conflicts regarding resource utilization, the Superalignment team disbanded in March 2024 over disagreements about the pace at which OpenAI was producing models. Superalignment Research Team Superalignment Research Team disbanded disbanded This was not the only example of how conflict within OpenAI led many of the lab’s researchers to move on. Although the acme of disagreements resulting from discussions on alignment occurred near the high-profile mass resignation of the Superalignment team in 2024, many of the basic ideas about the careful deployment and commercialization of advanced AI models already started to populate the company. In 2021, at a point where OpenAI suddenly ramped up the commercialization of ChatGPT-3 through a series of fast-paced releases, several top researchers (including Daniel Amodei, former VP of Research at OpenAI) left in protest, forming the rogue AI lab Anthropic along with researchers from Google Brain. Originally conceived as a research lab emphasizing the importance of AI safety measures, particularly alignment, in state-of-the-art AI development, Anthropic has notably lived up to this goal, now generally regarded as a longstanding champion for AI Ethics in the midst of a flurry of rushed AI releases. forming the rogue AI lab Anthropic forming the rogue AI lab Anthropic Onset of Specialization The start of late May has seen a flurry of action among what I call the Big Three (Anthropic, Google, and DeepMind) hinting at the rise of specialization in the climate of AI development. As previously mentioned, a commonly accepted trend was to create a general-use LLM that could perform many tasks at once. As the capabilities of such an LLM increased, the reasoning went, it would theoretically approach the AGI singularity that many dreamt to achieve. My counterargument: it’s called a singularity for a reason. Such an accomplishment as the literal creation of Artificial General Intelligence will likely not come as a result of the evolution of current Transformer-based architectures, which lack many of the “building blocks” of human intelligence (think: kinesthesis, latent and apparent memory). Instead, I believe that this singularity may arise from human research accelerated by agentic AI tools like AlphaEvolve, although this topic is to be thoroughly discussed in a future article. rise of specialization As this longstanding status quo of AI development played out, however, it became apparent that there is an all-out AI race to maintain the #1 spot in LLM rankings is starting to become untenable for anyone but Google, the Big Three member with the largest talent pool, the most sophisticated compute center network, and the most amount of data. The tech giant, long grown past its days of Google Bard, now boasts the state-of-the-art in general multimodal LLMs with Gemini-2.5, an iteration intended for release at the 2025 Google I/O conference along with a series of impressive models, including the new SoTA few-shot video generator (veo3) and real-time pattern-of-speech-preserving video translation tools, both of which will probably become built-in with Gemini’s toolcalling. In fact, shortly after Gemini-2.5’s user testing trial began, OpenAI announced a $6.5 billion deal to acquire AI hardware startup io, working alongside former Apple designers to (ostensibly) revolutionize traditional appliances through the creation and manufacturing of new “AI Hardware”. It’s debated whether this move signifies OpenAI’s strategic investment into what they see as a potential market or a total status shift from its focus on LLM development in light of Google’s newfound supremacy in the field. I’d like to argue that it’s more of a fallback, in case the AI race becomes unsustainable for OpenAI in the short run; in this case, they could probably partner with the dominant AI company at the moment in order to capitalize on an already-stable framework for AI hardware, whether it be smart appliances or autonomous robots. toolcalling toolcalling $6.5 billion deal $6.5 billion deal Google has the resources to pursue side projects along with maintaining its brisk lead in the AI race, and OpenAI has recently opened up a path for itself to go down if it wants to. In contrast to these two, Anthropic has established itself with optimizing Claude, its LLM, for coding, recently releasing the new SoTA with Claude Opus 4, although its lead in the benchmarking test is modest. In addition, Anthropic is notably the world’s biggest AI Ethics research organization, maintaining divisions studying AI safety as well as the societal impacts of new technologies. In line with this characterization, it portrays its mission as “Winning the AI race without losing [its] soul”, a somewhat bold tagline that antagonizes Google and OpenAI for leaning into a multitude of alleged data privacy and labor rights violations, two huge caveats for AI development that, so far, has not dissuaded any companies from partaking in the sheer hype of the Race. In a sense, these specializations in AI development objectives would mean that in the near future, top AI companies will cease being direct competitors of each other as the natural effects of product differentiation kicks in. One could argue that if Anthropic does focus on agentic coding models and Google, chatbots, the search for AGI could theoretically be sped up as redundant resource use is eliminated. Yet, I contend that specialization would actually be unfeasible for the broader evolution of AI in the future for a few reasons: Competition breeds innovation. Even though additional resources are not being consumed by a single company in development, the lack of impetus to innovate would hinder advancements despite this fact. If Google already held an extremely stable monopoly, it would not have incentive to spend billions on additional R&D to upgrade a product already in high demand. Companies do not share all discoveries. Assuming that all companies actually do invest into further development despite the emergence of specialization as mentioned above, AI companies have been historically reluctant to share the specifics of a model (even lobbying against proposed government legislation for 3rd party review of proprietary model information), which would practically negate any potential benefit from the “you do this, I do that” model of AI development. True open-source AI is still an ideal right now. There is a heavy overlap across specialized models. For example, as video generation models get more and more convoluted, they require more complicated transformer models trained using extensive data, including the breakdown of textual prompts and inherent reasoning like that found in traditional LLMs (for example, a well-trained video-generator would inherently assume that a “wood board boat” would refer to a boat made out of wooden boards instead of a personified log boarding a ship). This article is brought to you by Our AI, a student-founded and student-led AI Ethics organization seeking to diversify perspectives in AI beyond what is typically discussed in modern media. If you enjoyed this article, please check out our monthly publications and exclusive articles at https://www.our-ai.org/ai-nexus/read! Competition breeds innovation. Even though additional resources are not being consumed by a single company in development, the lack of impetus to innovate would hinder advancements despite this fact. If Google already held an extremely stable monopoly, it would not have incentive to spend billions on additional R&D to upgrade a product already in high demand. Competition breeds innovation. Even though additional resources are not being consumed by a single company in development, the lack of impetus to innovate would hinder advancements despite this fact. If Google already held an extremely stable monopoly, it would not have incentive to spend billions on additional R&D to upgrade a product already in high demand. Companies do not share all discoveries. Assuming that all companies actually do invest into further development despite the emergence of specialization as mentioned above, AI companies have been historically reluctant to share the specifics of a model (even lobbying against proposed government legislation for 3rd party review of proprietary model information), which would practically negate any potential benefit from the “you do this, I do that” model of AI development. True open-source AI is still an ideal right now. Companies do not share all discoveries. Assuming that all companies actually do invest into further development despite the emergence of specialization as mentioned above, AI companies have been historically reluctant to share the specifics of a model (even lobbying against proposed government legislation for 3rd party review of proprietary model information), which would practically negate any potential benefit from the “you do this, I do that” model of AI development. True open-source AI is still an ideal right now. There is a heavy overlap across specialized models. For example, as video generation models get more and more convoluted, they require more complicated transformer models trained using extensive data, including the breakdown of textual prompts and inherent reasoning like that found in traditional LLMs (for example, a well-trained video-generator would inherently assume that a “wood board boat” would refer to a boat made out of wooden boards instead of a personified log boarding a ship). This article is brought to you by Our AI, a student-founded and student-led AI Ethics organization seeking to diversify perspectives in AI beyond what is typically discussed in modern media. If you enjoyed this article, please check out our monthly publications and exclusive articles at https://www.our-ai.org/ai-nexus/read! There is a heavy overlap across specialized models. For example, as video generation models get more and more convoluted, they require more complicated transformer models trained using extensive data, including the breakdown of textual prompts and inherent reasoning like that found in traditional LLMs (for example, a well-trained video-generator would inherently assume that a “wood board boat” would refer to a boat made out of wooden boards instead of a personified log boarding a ship). This article is brought to you by Our AI, a student-founded and student-led AI Ethics organization seeking to diversify perspectives in AI beyond what is typically discussed in modern media. If you enjoyed this article, please check out our monthly publications and exclusive articles at https://www.our-ai.org/ai-nexus/read! This article is brought to you by Our AI, a student-founded and student-led AI Ethics organization seeking to diversify perspectives in AI beyond what is typically discussed in modern media. If you enjoyed this article, please check out our monthly publications and exclusive articles at https://www.our-ai.org/ai-nexus/read! https://www.our-ai.org/ai-nexus/read …Consequences Back to the question: why is the emergence of specialization in AI products a pivotal moment for AI development? The answer, in my opinion, is because this moment has the potential to change the landscape of AI development by directly deciding what they will do in the near future. It involves scenarios that I can only speculate about, due to my lack of connections in Silicon Valley, but have everything to do with standard game theory. Let’s assume, hypothetically, that Google and Anthropic both drop out of the AI race by abandoning the development of their respective AI models, Gemini and Anthropic. Although competitors may rise up to take their place, it’s unlikely that they would be able to trump OpenAI, which would almost certainly become a massive monopoly overnight, sucking up the two other companies’ talent almost instantly. Do the same for the other two companies, and a pattern emerges: If two fall, the one remaining company will become dominant on its own. However, when you apply this scenario to an AI company of lesser scale (say, Mistral or Meta), you may conclude that it’s far less likely for this company to become an AI superpower, since it would be more susceptible to market disruptions. why is the emergence of specialization in AI products a pivotal moment for AI development? may If two fall, the one remaining company will become dominant on its own. Why exactly is this the case? As previously mentioned, the quest for AGI differs from other evolutions in product development since this is the first instance (that I know of) in which a prototype product can directly assist in the development of a latter version. This fact will doubtlessly become more prevalent in the future along with the rise of agentic AI, but we can receive a salient example from Google’s AlphaEvolve model, a reasoning-and-coding model that has managed to increase DeepMind’s capacity for producing and training AI by designing a better load distribution framework for Google’s compute centers, increasing overall performance by roughly 1% (this may not be much, but it matters when you consider that each of Google’s compute centers handles data by the trillions of bytes). Assuming that this trend continues (that is, top AI companies continue making AI capable of helping streamline further development), AI companies with the most advanced technology will end up with an exponential growth pattern in which top-of-the-line AI models help build more innovation, in turn spurring on faster and faster AI development as this process repeats itself. In contrast, smaller companies which have not yet achieved the capability for scaling this type of AI self-enabling will have to grow at a much slower rate, where they are much more likely to be overtaken by another company given the right conditions. designing designing Considering all of these factors, it is unlikely that any of the Triumvirate of AI development will stop in its tracks anytime soon—all three have sunk deep into layers and layers of venture capital, brand advertisement, and, most importantly, societal expectations. With hundreds of billions of funding pouring into these massive endeavours, let’s hope that something good comes out of them. Written by Thomas Yin Written by Thomas Yin