📈 10 Facts from a Former OpenAI Employee's Report on the Future of AI

Leopold Aschenbrenner, who previously worked in the team of the departed OpenAI co-founder Ilya Sutskever, sparked a widespread discussion with his report titled "Situational Awareness: The Decade Ahead." The report discusses what to expect in the next 10 years in the field of AI and whether we should be concerned. Spoiler: ASI and giant investments.

We've gathered 10 highlights from this report for you.

📱 10 Predictions by Aschenbrenner

▶️AGI will emerge by 2027 and almost immediately transition to ASI (Artificial Superintelligence). This is because AGI will be capable of self-learning and quickly teach itself. The transition will happen rapidly, and humanity might not play a role in it.

▶️$1 trillion — that's the amount of investment required to build a computational center for creating AGI. For comparison, the cost of a mission to Mars is $10 billion.

▶️This computational center will consume more electricity than a large nuclear power plant can produce — approximately 20% of the energy that the US produces today. Incidentally, Sam Altman has also spoken about such resource expenditures.

▶️The main race among major powers in the next 10 years will not be for weapons but for achieving AGI. Aschenbrenner compares this to the nuclear arms race in the mid-20th century.

▶️Leading American AI labs are currently not giving enough attention to AGI safety. They are effectively handing over key secrets for creating AGI to Chinese specialists. Protecting these secrets will require enormous effort, and at the moment, the US is not coping with this task.

▶️$100 billion — that's how much NVIDIA, Microsoft, and Google are currently allocating to AI development in their quarterly budgets. By the end of the decade, annual investments will total $8 trillion.

▶️In 2024, producing AI chips may require 3-10% of the annual production capacity of advanced semiconductor factories.

▶️By 2030, global demand for production capacities for AI chips will be four times the current capabilities of TSMC, the leading chip manufacturer in Taiwan.

▶️By 2027, prototypes of automated engineers capable of accelerating technological processes in AI by 1.5-2 times will appear. By 2028, such systems will be able to automate more than 90% of research tasks, leading to significant acceleration in achieving AGI.

▶️Training the GPT-4 model required 3-10 thousand times more computational resources than GPT-2. One can only guess how many resources will be needed in the future.

Aschenbrenner discussed his report on a podcast with well-known blogger Dwarkesh Patel. Note: the episode is 4 hours long!

More stories:

💵 Sam Altman on OpenAI's expenses

❓ What is AGI?

#news #openai @hiaimediaen