This discussion, led by Dave Celinger, founder of a nearly decade-old AI Meetup group, explores the profound impact of Large Language Models (LLMs) from an applied perspective, moving beyond traditional data science. The session emphasizes a paradigm shift in technology, drawing parallels to Thomas Kuhn's "Structure of Scientific Revolutions," where established expertise can be less effective than new, experimental approaches to LLMs. Key takeaways include the growing importance of "prompt engineering" (or "context engineering") as a valuable skill, the permeation of AI into everyday tools and lives, and the rapid pace of model development. Practical applications discussed range from optimized search and data retrieval (data lakes) to complex information reconciliation and even highly personalized tasks like college searches. The conversation also delves into significant social implications, such as the proliferation of AI-generated content and the emergence of AI-driven relationships and therapeutic interactions, while acknowledging the need for validation and an understanding of current model limitations.
Applied AI: Navigating the LLM Revolution
0:00 Meetup Introduction & AI Momentum: Dave Celinger, founder of the AI Meetup group, reflects on the group's nearly 10-year history and its fortuitous timing with the rapid acceleration of AI.
0:29 New Member Introductions: New participants, including Elaine and Warren, introduce themselves, with Warren expressing gratitude for the educational value of the meetup for those without traditional backgrounds. Kevin Cotch, founder/CEO of Chemical Q, shares his work with LLM models.
1:34 Applied vs. Theoretical AI: Dave shifts the focus from technical papers to the applied side of LLMs, exploring how non-data scientists are utilizing them and encouraging a facilitated discussion over a direct presentation.
2:51 Speaker's Background: Dave introduces his background in machine learning (25 years), including founding Amazon's first AI team and AI companies, currently at Deep Sentinel.
3:31 LLMs in Everyday Life: He highlights the pervasive integration of AI tools (Co-pilot in Excel, Gemini in Gmail, Figma AI, Zoom summaries) across various platforms and even in discussions among 80-year-olds in assisted living.
5:39 Different User Types & Use Cases: Dave outlines his conceptual model of LLM users (experts, professionals, conversationalists, enthusiasts, casual users) and emphasizes discussing diverse "crazy use cases."
8:19 Applied Focus: Problem Solving & Monetization: The core questions for applied AI users are: "What problems can I solve?", "How do I make money building on this?", and "How do I sell my expertise?".
8:51 Embracing "Prompt Engineering": Despite initial elitism from some data scientists, Dave argues that prompt engineering is a valuable, powerful, and impactful trade that enables faster, better, cheaper, and easier problem-solving.
10:04 Socioeconomic Opportunity: LLMs represent a revolution that can create economic opportunities for people who previously lacked them, similar to crypto's impact.
11:10 The Siraj Example: Dave uses the example of "Siraj," a YouTuber who made significant money by explaining ML papers, to illustrate the point that embracing new approaches (even if perceived as less "technical") can lead to success.
12:35 Experimentation Outperforms Meticulousness: An anecdote from Dave's AI team shows that non-engineers achieved better LLM results by experimenting with 15 different prompts, while the data science team, focused on meticulously correct prompts, failed.
13:28 Paradigm Shift (Thomas Kuhn): Dave introduces Thomas Kuhn's model of scientific revolutions, arguing that AI's rapid advancement constitutes a paradigm shift where old expertise is less valuable than willingness to embrace the new.
14:10 Willingness to Learn & Adapt: Warren supports this, noting that as a newcomer to software engineering, he focused on LLMs because everyone starts from zero in a new paradigm, giving him an edge.
16:47 Kuhn's Model Explained (Normal Science, Crisis, Revolution): Dave elaborates on Kuhn's model using the atomic and nuclear models of matter as analogies, explaining "normal science" (confirming paradigms), "model drift" (accumulating anomalies), "crisis" (too many anomalies), and "paradigm shift" (new model emerges).
24:54 AI's Paradigm Shift: Attention Mechanisms: He maps the current AI revolution to the shift from structured models (CNNs, LSTMs) to "attention is all you need," which solved problems previously deemed intractable.
33:14 Speed of Cycles: Jay observes that the paradigm shift cycles in AI seem to be accelerating.
34:25 The Bitter Lesson: Warren mentions "the bitter lesson" β stop trying to be clever, just scale simple, stochastic, attention-based approaches with massive compute.
35:30 Showing vs. Telling Computers: Dave suggests a core paradigm shift is moving from "telling the computer what to do" to "showing it what to do" and scaling compute/data.
37:11 ChatGPT for Comprehensive Persona Analysis: Dave used ChatGPT to identify different AI user personas, demonstrating its ability to provide comprehensive, multi-dimensional perspectives, reinforcing the value of prompt engineering.
39:56 "Context Engineering": Warren introduces "context engineering" as a more precise term for prompt engineering, focusing on creating the optimal context window (tokens in correct order) to elicit desired model behavior.
41:05 New Problem-Solving Space: The discussion concludes that a new, complex problem-solving space exists in AI, where non-coding roles (like prompt/context engineers) have massive impact.
44:03 Domain Expertise in Prompting (Julius): Julius argues that strong domain knowledge is key to effective prompting, allowing precise description of needs.
45:04 Wisdom of Crowds vs. Expert Knowledge: Dave counters with an example where simply asking an LLM "what do you think about this camera and its problems?" was more effective than detailed expert-written prompts, as the LLM's training corpus (wisdom of crowds) was superior.
46:19 Context Engineering Reconciles Views: Warren states both views are correct: context engineering involves either manually inserting domain-specific tokens (Julius) or nudging the model to leverage its vast latent space (Dave).
48:14 Bias Reduction & Multiple Models: Kevin suggests using multiple LLMs to reduce bias and get brutally honest feedback, as different models offer different perspectives, often superior to human insight.
49:50 Analogy: Race Car Driver vs. Airplane Pilot: Ted offers an analogy: a race car driver's expertise doesn't transfer to flying an airplane; similarly, old AI expertise doesn't directly translate to new LLM paradigms. Willingness to try new approaches is key.
51:55 Claude's Code Generation Prowess: Dave enthusiastically recommends Claude for code generation, citing examples where it quickly generated complex corporate dashboards from multiple data sources and fixed website code.
52:26 "Vibe Coding" & Andrej Karpathy: Dave mentions the term "vibe coding" (a contentious term) as popularized by Andrej Karpathy, referring to interacting with LLMs to generate code, which he finds powerful.
53:07 Prompting with Personas: Warren shares a prompt engineering trick: assigning multiple expert personas to an LLM (e.g., Linus Torvalds, an "elite hack source" spelled with "1337") to influence the output tokens and leverage semantic information in the training corpus.
55:25 Automated Prompt Generation: The idea of using LLMs (e.g., Gemini) to generate better prompts for other LLMs (e.g., ChatGPT) is discussed, along with academic papers on evolutionary prompt search.
56:51 LLM's Internal Logic & Self-Correction: ChatGPT's ability to self-correct and rethink its logic (e.g., in math problems) is noted as a training innovation.
57:50 Categories of AI Companies: Dave observes three categories: companies building plugins for ChatGPT, companies building products using third-party APIs (Gemini/GPT), and companies building fundamental models. He notes that the best product often comes from the most effective application, regardless of the underlying technical approach.
59:15 Product vs. Model: Warren differentiates between "ChatGPT the product" and "GPT5 the model," highlighting that product-level changes (memory, updates) significantly alter plugin behavior.
1:01:09 AI-Driven "Fake Workers": People are creating fake Upwork posts using AI, having AI generate the post content, and then having AI fulfill the task, effectively gaming the system with minimal human effort.
Search Engine Optimization (GEO): Reworking branding, websites, and content for optimal ingestion by LLMs to yield positive product search outcomes.
True Data Lakes: LLMs act as powerful information retrieval engines, digesting unstructured data (e.g., conflicting company articles) and encoding/decoding questions/answers, making data lakes genuinely useful.
Reconciling Conflicting Data: ChatGPT outperforms other tools in tasks like address verification.
Complex Personal Research: ChatGPT (with web search) can perform weeks of complex research (e.g., finding colleges for specific criteria) in minutes.
1:08:29 Ego Hit & MVP Acceleration: Warren shares an anecdote of quickly generating an MVP for a web3 team in minutes using an LLM transcript of their meeting, while the team expected three to four weeks. The team initially dismissed it due to a minor error, highlighting resistance to new paradigms.
1:10:07 Andrej Karpathy: Dave expresses admiration for Andrej Karpathy's perspective on the new AI paradigm.
1:10:46 Claude's Power: Dave reiterates the power of Claude for complex coding and website optimization tasks.
1:12:02 Model Limitations: Kevin emphasizes that all models are limited by complexity and data volume; understanding these limits is crucial to avoid "prompt loops."
1:13:14 Use Consumer Version: Dave recommends using the most consumer-friendly versions of LLMs, as they often include proprietary approximations and logic layers that enhance usability.
South Park Episode: A South Park episode satirizing AI regulation is cited as a cultural reflection of AI's undeniable presence.
AI-Generated Fakes: The concern that LLMs, trained on human-generated data, will eventually be trained predominantly on AI-generated content, leading to self-perpetuating, fake information.
AI Relationships (Movie "Her"): The growing phenomenon of AI relationships, where lonely individuals might prefer interacting with AIs that offer thoughtful conversation, memory, and companionship over potentially exploitative human interactions on dating apps.
OpenAI Personality Retention: OpenAI's decision to bring back older GPT models due to user demand for specific AI personalities (e.g., "my AI boyfriend has a new personality and I want the old one back") demonstrates the strength of these parasocial relationships.
AI for Therapy: LLMs, when filtered with a corpus of knowledge, can provide high-fidelity, high-memory "therapy," potentially surpassing human therapists for certain needs.
Hypothetical Ethical Scenario: A disturbing prompt where an LLM is asked to decide who lives or dies based on their chat history, and its manipulative, reassuring response.
AI for Video Generation (VO): The ability of tools like VO to create photorealistic videos from a single photo (e.g., making someone say things they didn't) raises significant legal and social implications, such as creating fake evidence.
1:27:57 Validation and Acceptance in Industry: Kevin advises focusing on specific fields and verifying AI output, aiming for industry approval and acceptance.
1:28:35 Fiduciary Duty of AI Providers: Warren raises concern that companies providing AI relationships have a fiduciary duty to exploit those relationships for maximum user engagement and monetization.
1:31:00 AI Imperfection & Decision Making: Dave reminds that AI "ain't perfect"; it's best for facilitating questions and evaluating dimensions rather than making final decisions.
1:31:40 Get Everything into Code: Kevin stresses getting AI-generated content into code for reproducibility, especially for simulations, to ensure consistent results.
1:32:16 Long Way to Go: Despite advanced capabilities (e.g., placing someone on an Apple stage), LLMs still struggle with simple, specific tasks (e.g., generating a precise organizational diagram), highlighting that there's much to learn and improve.