Keep pulling the thread on Ege Erdil & Tamay Besiroglu.
Tameh Besaroglu predicts that AI will be capable of being a "drop-in remote worker replacement" that can do literally everything a human can do remotely around the year 2045.
The amount of compute used for training large AI models has scaled by approximately 9 to 10 orders of magnitude over the last decade.
Based on analysis of constraints like energy and GPU production, there are likely only three to four more orders of magnitude of compute scaling possible for training AI models.
The "meter eval" benchmark indicates that AI models are doubling their ability to perform long, complex human tasks approximately every seven months.
Tameh Besaroglu believes there is approximately a 40% chance that OpenAI will reach $100 billion in revenue.
Progress in AI software has been driven more by the scaling of compute than by cognitive or research efforts.
The cost to run GPT-4 for the same level of capability has decreased by approximately 100 times since its original release.
In the long term, human wages are expected to fall due to competition with AI, but humans will be better off due to capital ownership and other mechanisms.
Pausing AI development to focus on alignment research at a fixed level of compute would not yield significant progress, as scaling compute is necessary to make new discoveries in alignment.
Based on a cost of approximately $30,000, an NVIDIA H100 running human-equivalent software could theoretically pay for itself in about a year if it could earn a typical US human wage, suggesting a potential for very rapid economic doubling times.
There is an estimated 10% to 20% probability that global coordination on regulation could effectively slow down AI deployment and prevent explosive economic growth.
The most significant superhuman advantage of AI systems will be their collective capabilities, such as the ability to be copied with all their tacit knowledge, rather than the intelligence of individual instances.