I expect rapid progress but not towards general superintelligence

I anticipate swift advancements in artificial intelligence, but I doubt this will lead to general superintelligence. My skepticism arises from the perception that researchers view AI as becoming superior to their own capabilities in a few years. This is attributed to the acceleration in computational infrastructure and engineering of these models.

Indeed, AI models will soon act as highly proficient distributed GPU engineers, rendering experimentation with model formulation significantly more straightforward. This doesn't, however, imply that our models will exhibit diverse intrinsic qualities. Ilya's assertion that we are in the research era of AI was prescient, as experimentation is now far simpler due to coding agents.

This shift in work dynamics highlights a transformative period where innovative ideas hold greater value than flawless execution in software development.

The field has already transitioned between research and engineering multiple times. Deep learning's emergence marked a substantial shift from AI being predominantly a research pursuit. Today, top researchers are lauded for their capacity to implement and scale ideas within intricate infrastructures. We are on the brink of another substantial acceleration in infrastructure enhancements, driven by parallelized, AI-assisted language modeling.

This trend emphasizes how significant computational improvements are shaping the AI landscape. The objective of this analysis is to elucidate the rationale behind this engineering surge and its implications, acknowledging what it may not resolve. This forthcoming progress will substantially facilitate the dissemination of AI technology throughout the economy, even if it does not yield economically valuable superhuman abilities beyond mathematics and coding.

Numerous elements within the AI training and inference stack are amenable to verification. Among these, training metrics such as tokens per second per GPU—essentially a measure of training speed—are particularly noteworthy. Correspondingly, inference metrics including tokens per prompt, FLOPs per token, or merely the cost per answer, are also highly optimizable.

These metrics can be refined by leveraging established sub-problems and architectural trade-offs. I foresee AI agents aiding in the optimization of this entire process over the next few years, potentially bringing inference capabilities to near the maximum potential of our accelerators like GPUs. Over the past few years, companies have achieved substantial gains in inference efficiency, sometimes saving 10-30% on costs to serve a model at a given price point.

As this stack compounds, I anticipate the effective cost of model intelligence to decline exponentially in the coming years, potentially at a faster rate than recent trends.

This reduction in efficiency costs should be attainable within a few years. Longer-term, co-design of accelerators and models may yield even greater efficiency gains, surpassing the capabilities of current GPU platforms. This period of rapid efficiency improvements will undoubtedly exert pressure on the industry, given the GPU's flexibility and importance in enabling exploratory architectural innovations.

A plausible prediction is that pretraining research, encompassing architecture and data selection, can be automated within 2-3 years. This automation in pretraining research appears reasonable, considering the potential impact on other domains.

This acceleration will likely trigger a Jevons paradox for agentic models, suggesting a demand surge as the industry grapples with optimizing and deploying agents. An illustrative example is Meta's Muse agent, which signifies the dawn of AI agents tailored to various audiences and use cases. The value generated by these agents will stem from understanding their operation rather than merely attaining higher performance levels.

This will echo a pivotal phase in broader scientific literature, where AI models will excel at cross-disciplinary literature exploration and forming connections within sparse networks, previously monopolized by select scientific communities. This era will witness AI models as superhuman agents in literature traversal and inter-network association, potentially heralding a new scientific discovery era, akin to breakthroughs in curing most cancers.

Lastly, another significant low-hanging fruit lies in enhancing the quality of reinforcement learning environments. The current market is awash with subpar data providers, with many vendors delivering low-quality offerings. The leading RL data companies, however, demonstrate clear ROI, indicating that the sector's quality can and should be significantly improved.