Keep pulling the thread on Ari Morcos & Rob Toews.
Ari Morkos asserts that simply scaling up compute and data for AI model training is now hitting diminishing returns, an outcome that was predicted by original scaling laws research.
Major AI labs including OpenAI, Anthropic, and XAI have set aside massive, effectively unlimited budgets to acquire Reinforcement Learning (RL) environments from external startups.
Enterprises can now build a high-performing, custom AI model for a cost in the single-digit millions of dollars.
Rob Taves predicts that Google's upcoming Gemini 3 model will be state-of-the-art and will surpass the capabilities of GPT-5.
A recent trend in AI involves "effective acquisitions" with unconventional structures, such as those of Inflection and Character AI, designed to avoid M&A and antitrust scrutiny from regulators like the FTC.
Rob Taves predicts that compelling non-invasive Brain-Computer Interface (BCI) products will begin hitting the consumer market within the next 12 to 24 months.
Rob Taves believes the most likely path to a major leap in AI capabilities is through recursive self-improvement, where AI systems autonomously design better AI systems.
Ari Morkos asserts that synthetic data is a critical solution to the "data wall" problem and will be essential for creating high-performing, domain-specific AI models.
Ari Morkos believes that progress in AI models is not slowing down significantly, but is instead shifting to different axes such as reasoning and cost-effectiveness.
Rob Taves argues that the progress in the raw intelligence and capabilities of underlying foundation models is slowing down, following a natural S-curve after a period of exponential growth.
Ari Morkos states that smaller AI models are now matching the capabilities of massive models from a year ago, representing significant progress in model efficiency.
Rob Taves asserts that Reinforcement Learning (RL) is highly effective for AI models in domains with verifiable answers like math and coding, but its efficacy in less verifiable domains is uncertain.