Currently, many prominent figures in the AI field are predicting the imminent arrival of Artificial General Intelligence (AGI) and even Artificial Super Intelligence (ASI). Let’s analyze what achieving these milestones truly requires and assess how close we are to them.
Computational Power
Artificial General Intelligence, and especially Artificial Super Intelligence, demand immense computational resources both for training and execution. Nevertheless, we have two vectors progressing toward each other that will ultimately solve this challenge:
- Improved AI model architectures
- New methods for training and executing models
Architectural innovations and novel mathematical techniques for matrix multiplication could significantly reduce computational demands, enhancing models’ intelligence per computational unit. Simultaneously, advancements in microchip technology and quantum computing promise substantial gains in energy efficiency and processing speed.
Governmental pressure to become pioneers in general AI has already resulted in billions of dollars being invested, and we can expect significant breakthroughs within the next decade.
Novel Training Methods
Model training and the creation of training datasets will undoubtedly become one of the critical factors in achieving AGI and ASI. In their competitive rush, companies often neglect the quality of data included in their training sets. Information noise such as propaganda, unstructured and out-of-context examples, or meaningless user-generated content has led to highly biased models with distorted perceptions of reality in certain aspects.
We first need specialized agencies responsible for professionally curating and balancing high-quality datasets. Such datasets should exclude historical facts, details about individuals, or specific events. Instead, they must include structured logical problems, fundamental principles of mathematics and physics, unbiased philosophy and ethics, and structured linguistic knowledge.
By emphasizing quality over quantity, the first small-parameter LLM models with exceptionally high analytical capabilities will emerge. These models will efficiently process vast amounts of input data and surpass any existing counterparts.
Such a model could automate the task of data structuring, enriching the training datasets for subsequent generations of models. Ultimately, this process will enhance the intelligence rather than merely the informational content of AI models.
Today, platforms like Hugging Face and Kaggle already enable the free storage and sharing of large datasets. Early attempts at creating distilled datasets, known as “instruct datasets,” are also underway. While there’s still much work ahead, progress is tangible. I estimate it will take about 5 years to develop lightweight models capable of solving complex problems at the level of PhDs.
Meta-Execution
With the introduction of “thinking” models, OpenAI has made another revolutionary leap. We no longer demand that models provide correct answers on the first try. Now, they can think, make mistakes, and correct themselves—just like humans.
This approach clearly points in the right direction for developing the intellectual capabilities of automated systems based on LLMs. Just half a year ago, I would have stated this was missing. But now, this capability already exists.
However, there’s one more aspect. The emergence of highly intelligent analytical models, yet lacking real-world awareness, calls for developing a language describing task execution conditions. Essentially, this is natural language-based meta-programming. Prompt engineering will evolve in this direction.
Conclusion
I doubt we’ll wake up one day to headlines announcing the sudden creation of AGI or ASI. Achieving these milestones will be a gradual, multi-stage process demanding comprehensive efforts. Humanity is undoubtedly exerting these efforts right now. My estimate of 5-10 years for achieving AGI/ASI seems highly plausible.