The Next Gen AI Models: Reliable, Consistent, Trustworthy — With Aidan Gomez

Episode Summary

Executive Summary: Aiden Gomez argues AI is moving from proof-of-concept to production, with value showing up in enterprise workflows rather than “godlike” AGI. He says larger models and more compute improve capability but with diminishing returns, while practical ROI comes from integrating assistants into real systems, improving reliability, and automating boring but expensive work across industries.

Main Topics: AI investment and ROI (Priority: 5/5): Gomez says current training and infrastructure spending is justified by long-term value, but real returns depend on deploying models at scale in consumer and especially enterprise settings. Myths about AGI and “godlike” models (Priority: 5/5): He rejects hype around superintelligence, arguing that near-term progress will be steady improvements in accuracy, reliability, and trustworthiness rather than a sudden capability jump. Scaling compute and model size (Priority: 4/5): More GPUs and compute do produce better models, but gains taper off and very large models can become impractical if they are too expensive to deploy or consume. Emergent behavior, self-improvement, and synthetic data (Priority: 4/5): Gomez is skeptical of claims that models will suddenly take off through self-improvement; he says synthetic data is useful mainly in verifiable domains like math and code, not for broad general knowledge. Enterprise use cases and practical productivity (Priority: 5/5): He emphasizes that the most valuable AI applications are mundane but economically meaningful: document review, job descriptions, supplier search, healthcare workflows, and supply chain automation. Safety, control, and work displacement (Priority: 4/5): Gomez argues humans choose where models are deployed and can impose safeguards; he expects augmentation rather than mass unemployment, with AI taking repetitive tasks and enabling more fulfilling work. Technical path forward: RAG, multimodality, and assistants (Priority: 4/5): He sees retrieval-augmented generation, multimodal models, and reasoning-style chains of thought as key to building more reliable assistants that can collaborate with users daily.

Key Arguments: Current AI spending is small relative to the long-term economic value the technology can create, but that value only materializes through production deployment and workflow integration. The next generation of models will likely be more reliable and accurate, not magically superhuman or a step-change toward “godlike” AI. Bigger models do help, but scaling has diminishing returns and very large models must remain affordable enough to actually serve users. AGI, defined as human-level performance on tasks humans do, is a reasonable and achievable target; superintelligence is a different and less useful concept. AI is more likely to augment workers than replace them, because society has abundant demand for more productivity and better outcomes. Enterprise AI’s biggest opportunities are in boring back-office processes where speed, cost, and accuracy gains are economically huge. Synthetic data is valuable for verifiable tasks like coding and math, but human expertise is still needed for more open-ended domains like philosophy or social science. Reasoning models and explicit step-by-step outputs can increase trust because users can inspect how an answer was produced. The main risk is not doomsday AI; it is deploying systems in high-stakes settings without proper oversight and safeguards. RAG is the dominant enterprise architecture because companies want to combine base models with proprietary data and internal tools.

Data Points: OpenAI raise: $6.6 billion - Used as an example of major AI investment and the debate about whether the spend can be recouped. OpenAI reported losses: $5 billion per year - Mentioned in the discussion about whether AI model training is economically sustainable. State-of-the-art model training GPUs (Llama 3): 16,000 GPUs - Referenced as the prior benchmark for large-scale model training. Elon Musk supercluster size: 100,000 GPUs - Cited as an example of the next scale of compute being pursued. Oracle applications powered by Cohere: Over 50 applications - Cohere is embedded in Oracle’s enterprise software across functions like HR and supply chain. Accenture generative AI bookings growth: About 50% quarter over quarter - Used to illustrate enterprise demand for AI implementation services. Enterprise adoption stage: Year of proof of concept last year; going to production this year - Gomez describes the shift from experimentation to deployment. AI usage scale: Hundreds of millions of people - He argues consumer ROI already exists because many people now use the technology daily.

Pivotal Quotes: "I don't actually think that's... We'll get to that." — Aiden Gomez: He responds to the idea that future models will become godlike or solve all problems. "Even if the technology froze and what we have today is all we get, there's so much good to be done." — Aiden Gomez: He argues that current AI is already valuable enough to justify major deployment work. "I think that the internet is enough." — Aiden Gomez: He states his view that models can learn sufficiently from observation and text without embodiment.

Implications: The near-term AI story is less about AGI headlines and more about enterprise deployment, reliability, and workflow automation. Listeners should expect steady gains, major productivity impact, and growing demand for AI integration services rather than sudden takeover narratives.