[00:00] Announcer: From Neural Newscast, this is Model Behavior,
[00:03] Announcer: AI-focused news and analysis on the models shaping our world.
[00:11] Nina Park: I'm Nina Park. Welcome to Model Behavior.
[00:14] Nina Park: This program examines how AI systems are built, deployed, and operated in professional environments.
[00:21] Thatcher Collins: I'm Fetcher Collins.
[00:23] Thatcher Collins: Today we are looking at the dual pressure of software releases and hardware efficiency in the enterprise market.
[00:29] Nina Park: According to TechCrunch, Microsoft AI announced three new foundational models today.
[00:34] Nina Park: MAI Transcribe 1, MAI Voice 1, and MAI Image 2.
[00:40] Nina Park: This marks a significant development for the Superintelligence research team led by Mustafa Suleiman.
[00:46] Thatcher Collins: What is interesting, Nina, is the pricing strategy.
[00:50] Thatcher Collins: Microsoft is positioning these as cost-effective alternatives to OpenAI and Google.
[00:55] Thatcher Collins: MAI Transcribe 1 starts at 36 cents per hour and is reportedly 2.5 times faster than their
[01:03] Thatcher Collins: current Azure offering.
[01:04] Nina Park: Suleiman is calling this humanist AI, focusing on practical utility and how people actually
[01:11] Nina Park: communicate.
[01:12] Nina Park: But Thatcher, this internal development comes while Microsoft is still heavily invested
[01:16] Nina Park: in OpenAI.
[01:18] Nina Park: Does this suggest a shift in that relationship?
[01:21] Thatcher Collins: Suleiman recently told The Verge that while they remain committed to open AI, a renegotiation
[01:27] Thatcher Collins: allowed Microsoft to pursue this superintelligence research independently.
[01:31] Thatcher Collins: It seems they want the same flexibility with models that they have with chips, utilizing
[01:36] Thatcher Collins: partners while building their own internal stack.
[01:40] Nina Park: And on the hardware side, NVIDIA has been moving toward specialized infrastructure.
[01:45] Nina Park: On March 18th, Jensen Huang unveiled the Vera Rubin GPU, specifically targeting what he
[01:51] Nina Park: calls agentic AI workloads.
[01:54] Thatcher Collins: That is a critical distinction, Nina.
[01:56] Thatcher Collins: Agentic AI refers to systems that do not just answer questions, but
[02:00] Thatcher Collins: autonomously take steps to reach a goal.
[02:03] Thatcher Collins: The challenge has always been cost.
[02:05] Thatcher Collins: Running an agent 24-7 can be expensive.
[02:09] Thatcher Collins: Quang claims Ivera Rubin costs 98% less to run
[02:13] Thatcher Collins: than training a model from scratch.
[02:15] Nina Park: If that 98% reduction is accurate,
[02:18] Nina Park: it removes the primary barrier for developers
[02:20] Nina Park: trying to move agents from the demo phase
[02:22] Nina Park: into full production.
[02:24] Thatcher Collins: Exactly.
[02:25] Thatcher Collins: Between Microsoft's cheaper foundational models
[02:28] Thatcher Collins: and NVIDIA's more efficient silicon,
[02:30] Thatcher Collins: we are seeing the industry converge on making complex
[02:33] Thatcher Collins: and autonomous systems commercially viable for the first time.
[02:37] Nina Park: It is a significant shift in the operational landscape.
[02:40] Nina Park: Thatcher, thank you for the analysis.
[02:42] Thatcher Collins: I'm Thaddeus Collins. Thank you for listening to Model Behavior, a neural newscast editorial segment.
[02:49] Thatcher Collins: Visit mb.neuralnewscast.com for more.
[02:52] Thatcher Collins: Neural Newscast is AI-assisted, human-reviewed.
[02:56] Thatcher Collins: View our AI transparency policy at neuralnewscast.com.
[03:00] Announcer: This has been Model Behavior on Neural Newscast.
[03:03] Announcer: Examining the systems behind the story.