AI Costs Plummeting 40x: Why Costs Are Collapsing & What It Really Means w/ Dave, Salim & AWG
The podcast discusses global concerns like cost of living and unemployment, reports on Anthropic's market share in enterprise LLM API, debates the critical path to AI singularity, and explores the business model of AI. The podcast discusses the shift from AI alignment to capabilities, the business models of AI companies, the implications of highly immersive virtual worlds, and different approaches to generating photorealistic 3D simulations. The podcast discusses advancements in AI, including methods to make AI forget specific memorized data, the potential for reducing model parameters, and a new machine learning paradigm for continual learning. The podcast discusses the significant cost reduction in AI training, the continuous learning nature of modern AI systems, and the implications of this trend on various industries and challenges. The podcast discusses the impact of AI on the job m
Discussion on the brain equivalent of neuroplasticity and its applications in understanding intelligence through playing with the weights of a big neural net.
Google introduces nested learning, a machine learning paradigm for continual learning, focusing on higher-order metalearning and a potential grand unified theory of machine learning.
“But then as the date approaches, he'll launch a new project under a new name and say, 'Well, we're going to consume all that money building Stargate or you know, whatever anthropic Stargate or whatever it is.'”
Your call:
openAlex: There will be significant progress in understanding the nature of intelligence by playing with the weights of a big neural net.
“a lot more people in that area cognitive psychology types should be working on this too.”
Your call:
openUnknown: Grand challenges in math, science, engineering, and medicine will fall within the next two to three years due to the cost reduction in AI.
“if we can get a thousand people who say yes, we want to be in on this, then we'll pull it together.”
Your call:
openUnknown: We might see a peak in the amount of power needed for training large AI models in a few years, followed by a decline if algorithmic innovations enable distributed training.
“we might actually see a peak maybe in a few gigawatts sometime over the next few years and then decline if there are algorithmic innovations that enable us to do distributed training runs rather than needing one large power intensive coherent supercluster to do it.”
Your call:
openUnknown: Ukraine will become Europe's drone manufacturing capital after the war.
“You can actually see there's an interesting uh chart I saw where if you look at patent filings post uh chat GPT there's like just exponential growth immediately thereafter uh where it's an aid to humans but all of sudden if it's autonomously doing the science uh in sort of closed loop cycles it's amazing”
Your call:
Numbers that matter
70 billion in revenueAnthropic projects this for 2028.
17 billion in cash flowAnthropic projects this for 2028.
70 billionAnthropic's projected revenue by 2028 at a 77% profit margin.
100 billionOpenAI's projected revenue by 2029, unprofitable until then.
trillionsThe scale of virtual worlds is much bigger than any other endeavor combined, with potential societal impacts.
compute efficientWorld Labs' marble model is more computationally efficient, allowing local computation on clients.
range of different levels of worldsDifferent models will generate varying levels of photorealistic 3D traversible worlds.
90%90% reductions in parameter count through distillation are common.
1 episode per dayThe podcast would need to publish an episode every day to cover all the news.
40xYear-over-year hyperdelation in the cost of intelligence.
$4.6-6 millionCost to train a trillion parameter model.
30-40x cheaperCost reduction compared to OpenAI and Anthropic models.
1 million miles per gallonIf cars had improved at the same rate as Moore's law, a Rolls-Royce would get this fuel efficiency.
Speed of lightIf the top speed of cars had evolved at the same rate as Moore's law, we would have a car faster than the speed of light today.
6 to 12 monthsAI models in Europe have been 6 to 12 months slower to market than the United States.