What does it actually take for AI to move from a helpful interface to a system that does real work inside firms and data centers?
In this episode, Mitchell Troyanovsky of Basis discusses how accounting teams can use agents for end-to-end workflows, why software may shift beyond the traditional SaaS UI, and how these changes affect CPA training and firm operations. Jay Dawani of Lemurian Labs then makes the case that the next bottleneck is not just model quality, but the underlying compute stack: memory bandwidth, kernel coverage, hardware portability, and runtime orchestration across heterogeneous systems.
The conversation opens and closes with broader questions about AI ethics, robotics, data-center economics, GPU pricing, global supply chains, and whether the benefits of more capable AI will be widely shared.
Show Notes
Mitchell Troyanovsky of Basis explains how AI agents are reshaping accounting workflows, CPA training, and the role of human judgment in firms. Jay Dawani of Lemurian Labs breaks down the memory-bandwidth, compiler, and heterogeneous-hardware constraints shaping AI inference and infrastructure. The opening and closing also cover AI chain-of-thought ethics, robotics, data-center politics, GPU pricing, and space-based compute.
Chapters
(0:00) Robots learn from a few demos.
(0:48) AI agents will outgrow your UI.
(1:53) AI needs 106 billion kernels.
(2:52) Data centers are mostly know-how.
(3:57) Opening and today’s topics
(4:30) AI chain-of-thought rabbit hole
(8:04) Helpful models and ethics
(10:33) Why data centers face backlash
(12:58) Great Lakes and local impacts
(20:03) Data center political economy
(23:35) Cash payments and basic income
(27:20) One-shot robot learning
(28:40) China and robotics acceleration
(32:26) Robotic singularity
(33:05) Basis and autonomous accounting agents
(34:44) Selling AI to accountants
(35:17) Hours to minutes value
(37:29) AI in accounting versus coding
(38:36) Accounting firms need revenue
(41:13) What accounting really does
(44:29) Firm-wide AI systems
(48:28) Token costs and frontier AI
(51:07) Atlas and agent context
(55:08) Process supervision
(59:39) AI changes CPA training
(1:04:27) Human judgment and AI context
(1:11:38) SaaS beyond the UI
(1:15:25) Proactive tax agents
(1:19:29) The human touch debate
(1:22:40) Jay Dawani and Lemurian Labs
(1:24:24) Why kernels are hard
(1:26:48) Learning kernel programming
(1:27:59) How compilers translate hardware
(1:29:15) The memory bandwidth wall
(1:30:21) Compiler-generated kernels
(1:31:53) Inference latency metrics
(1:34:07) Scaling beyond one device
(1:36:45) Old single-chip assumptions
(1:38:52) What makes an AI agent
(1:40:22) Runtime orchestration
(1:42:26) Intelligence without LLMs
(1:45:54) The kernel coverage problem
(1:48:05) Hardware portability
(1:49:13) Operator fusion
(1:53:17) Heterogeneous hardware
(1:55:32) Tachyon rollout
(1:57:30) Pricing effective compute
(1:59:37) A human-first AI future
(2:01:54) Chip prices and compute
(2:04:34) AI infrastructure margins
(2:06:49) Global AI supply chain
(2:08:14) Rationalist supply-chain hymn
(2:11:56) Public opinion on AI
(2:13:09) AI backlash and risks
(2:15:57) U.S. data-center construction
(2:17:28) Space data centers
(2:19:58) Universal basic income
(2:21:30) Closing sign-off
Guests:
Jay Dawani — Founder & CEO, Lemurian Labs
Mitchell Troyanovsky — Co-Founder, Basis (𝕏 | LinkedIn)
