LLMs - From poker nights to mission-critical systems | Thy Virtuoso | Episode 1

Описание: In this deep dive, we move past the novelty of chat interfaces to explore the gritty reality of building mission-critical systems with LLMs. Using real-world failures…from "poker night" hallucinations to mismanaged data, we break down why LLMs are inherently probabilistic "liars" and how engineers can architect around their sycophantic nature.

We shift the focus from simple prompt engineering to the "Fringe Stack," where the real enterprise value lies:
1. Building validation layers.
2. Leveraging Graph Attention Networks.
3. Using Mixture of Experts (MoE) to solve complex, low-data problems.

Whether you're navigating the tactical utility of data cleaning or the strategic complexity of multimodal spatial analysis, this video outlines a roadmap for keeping AI on a leash while building the deterministic anchors necessary for high-stakes deployment.

Timestamps:
00:00:00 - Podcast recap
00:01:30 - Meet the panel
00:01:43 - Magic & the mess with AI
00:02:15 - The "Yes-Man" problem: Taming LLM sycophancy
00:03:18 - Prathamesh's interview assignment
00:04:20 - Unexpected ChatGPT experiments
00:05:00 - Shifting from chatbot to engine
00:06:30 - The tactical layer: Automating low-level tasks
00:08:04 - The strategic layer: AI as a high-level sounding board
00:09:22 - The LLM as an intellectual companion
00:11:38 - Sumanth’s playbook: The top two usage frameworks
00:12:04 - Power of native multimodality
00:12:55 - Real-world perception: Propheous & spatial use cases
00:14:45 - Future proofing: LLMs as a fundamental capability
00:16:33 - The Pivot for undergraduates: Core stack vs. "Fringe stack"
00:18:00 - Software vs. Domain expertise
00:19:09 - Why young engineers must stay critical
00:21:20 - How much should we trust AI?
00:23:58 - User interface hacks for efficiency
00:24:44 - Architectural edge: Leveraging attention networks
00:26:08 - The research partner: Deep diving with AI
00:26:43 - Search vs. Query
00:27:14 - Verifying through hard sources
00:27:47 - Prompting mastery
00:28:40 - Optimizing ChatGPT & Perplexity
00:29:01 - Knowing when LLMs fail
00:29:30 - Where the next big opportunities lie
00:30:30 - The trust deficit, validation & verification systems
00:34:40 - The new frontier: Why AI R&D is the best opportunity
00:35:27 - Synthesis: Key learnings and executive summary
00:36:53 - Citation models
00:39:00 - Identifying the two main LLM flaws
00:39:38 - Managing shortcomings
00:39:41 - Verification hacks
00:41:12 - Urban intelligence: Tracking development at Propheous
00:42:34 - Anchoring models in reality
00:44:32 - The enterprise leap: Selling AGI to organizations
00:46:00 - Will AGI displace the workforce?
00:46:39 - The data goldmine: Rethinking enterprise strategy
00:48:52 - Empowering teams via AI
00:49:30 - How enterprises view AI adoption
00:50:31 - Architecture shift: Transformers across disciplines
00:53:21 - Beyond text: Graph models and attention networks
00:54:02 - The next horizon: Emerging architectures and model arrays
00:55:08 - Small Language Models & cognitive science
00:56:26 - Deploying an array of LLMs
00:57:50 - Human brain's sophistication
00:59:12 - Self-evaluation: Auditing your own tech usage
00:59:44 - Gut-based decisions and AI gaps
01:01:52 - Fun bit - Matching LLMs with famous personalities