Pre‑training vs Fine‑tuning: How AI Learns Its Basics Hosted by Nathan Rigoni

In this episode we break down the two foundational stages that turn raw data into useful AI systems: pre‑training and fine‑tuning. What does a model actually learn when it reads billions of words or scans millions of images, and how do we reshape that knowledge into behaviors like answering questions, writing code, or describing pictures? By the end you’ll see why the split between “learning the mechanics” and “learning the behavior” is a game‑changer for building adaptable, efficient models— and you’ll be left wondering: could the next wave of AI rely more on clever fine‑tuning than on ever‑larger pre‑training datasets?

What you will learn

• The goal of pre‑training: teaching a model the fundamental mechanics of language or vision through massive supervised (next‑token) learning.
• How fine‑tuning shifts focus to behavior, using methods such as RLHF, instruction tuning, code‑tuning, and agentic fine‑tuning.
• A concrete case study: ServiceNow’s Apriel model, which starts from the multimodal Pixtral backbone and is fine‑tuned for conversational VLM capabilities.
• Trade‑offs in data volume, compute cost, and model size when choosing between larger pre‑trained models and smaller models enhanced by aggressive fine‑tuning.
• Key terminology you’ll hear repeatedly: supervised pre‑training, reinforcement learning from human feedback (RLHF), instruction tuning, RLVR, multimodal fine‑tuning.

Resources mentioned

• OpenAI “GPT‑4 Technical Report” (details on pre‑training and RLHF).
• ServiceNow blog post on Apriel and the underlying Pixtral model.
• Christiano et al., “Fine‑Tuning Language Models from Human Preferences” (2017).
• Hugging Face’s guide to instruction fine‑tuning with Transformers.
• Papers on multimodal vision‑language models such as CLIP and Flamingo.

Why this episode matters
Grasping the distinction between pre‑training and fine‑tuning lets engineers, product leaders, and AI enthusiasts make smarter choices about model selection, cost management, and deployment strategy. Whether you’re building a chatbot, an image captioner, or an autonomous agent, knowing when to invest in massive pre‑training versus targeted fine‑tuning can dramatically impact performance, latency, and scalability. This insight also highlights a growing trend: leveraging modest multimodal backbones and unlocking them with specialized fine‑tuning to create edge‑friendly AI solutions.

Subscribe for more bite‑sized AI deep dives, visit www.phronesis-analytics.com, or email [email protected]. Your feedback shapes future episodes—keep the conversation going!

Keywords: pre‑training, fine‑tuning, RLHF, instruction tuning, multimodal models, vision‑language models, Pixtral, Apriel, reinforcement learning from human feedback, AI curriculum learning, model behavior, edge AI.

Multimodal Models: Vision, Language, and Beyond Hosted by Nathan Rigoni

In this episode we untangle the world of multimodal models—systems that learn from images, text, audio, and sometimes even more exotic data types. How does a model fuse a picture of a cat with the word “feline” and the sound of a meow into a single understanding? We explore the building blocks, from early CLIP embeddings to the latest vision‑language giants, and show why these hybrid models are reshaping AI’s ability to perceive and describe the world. Can a single hidden state truly capture the richness of multiple senses, and what does that mean for the future of AI applications?

What you will learn

• The core idea behind multimodal models: merging separate data modalities into a shared hidden representation.
• How dual‑input architectures and cross‑modal translation (e.g., text‑to‑image, image‑to‑text) work in practice.
• Key milestones such as CLIP, FLIP, and modern vision‑language models like Gemini and Pixtral.
• Real‑world use cases: image generation from prompts, captioning, audio‑guided language tasks, and multimodal classification.
• The challenges of scaling multimodal models, including data diversity, hidden‑state alignment, and computational cost.

Resources mentioned

• CLIP (Contrastive Language‑Image Pre‑training) paper and its open‑source implementation.

• Recent vision‑language model releases: Gemini, Pixtral, and other multimodal LLMs.
• Suggested background listening: “Basics of Large Language Models" and “Basics of Vision Learning” episodes of The Phront Room.
• Further reading on multimodal embeddings and cross‑modal retrieval.

Why this episode matters
Understanding multimodal models is essential for anyone who wants AI that can see, hear, and talk—bridging the gap between isolated language or vision systems and truly integrated perception. As these models become the backbone of next‑generation applications—from creative image synthesis to audio‑driven assistants—grasping their inner workings helps developers build more robust, interpretable, and innovative solutions while navigating the added complexity and resource demands they bring.

Subscribe for more AI deep dives, visit www.phronesis-analytics.com, or email [email protected].

Keywords: multimodal models, vision‑language models, CLIP, FLIP, cross‑modal translation, hidden state, image generation, captioning, audio‑text integration, multimodal embeddings, AI perception, Gemini, Pixtral.

Large Language Models: Building Blocks & Challenges Hosted by Nathan Rigoni

In this episode we dive into the heart of today’s AI—large language models (LLMs). What makes these gigantic text‑predictors tick, and why do they sometimes hallucinate or run into bias? We’ll explore how LLMs are trained, what “next‑token prediction” really means, and the tricks (chain‑of‑thought prompting, reinforcement learning) that turn a raw predictor into a problem‑solving assistant. Can a model that has never seen a question truly reason to an answer, or is it just clever memorization?

What you will learn

• The core components of an LLM: tokenizer, encoder, transformer blocks, and the softmax decoder.
• Why training at terabyte‑scale data and quintillion‑level token iterations is required for emergent abilities.
• How chain‑of‑thought prompting and the REACT framework give models a “scratch‑pad” for better reasoning.
• The role of fine‑tuning and reinforcement learning from human feedback in shaping model behavior.
• Key pitfalls: lack of byte‑level tokenization, spatial reasoning limits, Western‑biased training data, and context‑window constraints (from ~128 k tokens to ~2 M tokens).

Resources mentioned

• Tokenization basics (see the dedicated “NLP – Tokenization” episode).
• Auto‑encoder fundamentals (see the “NLP – Auto Encoders” episode).
• Papers on chain‑of‑thought prompting and REACT agents (discussed in the episode).
• Information on context‑window sizes and scaling trends (e.g., 128 k → 2 M tokens).

Why this episode matters
Understanding LLM architecture demystifies why these models can generate coherent prose, write code, or answer complex queries—yet also why they can hallucinate, misinterpret spatial concepts, or inherit cultural bias. Grasping these strengths and limits is essential for anyone building AI products, evaluating model outputs, or simply wanting to use LLMs responsibly.

Subscribe for more AI deep dives, visit www.phronesis‑analytics.com, or email nathan.rigoni@phronesis‑analytics.com.

Keywords: large language models, next‑token prediction, tokenizer, transformer, chain of thought, REACT framework, reinforcement learning, context window, AI hallucination, model bias.

AI in Space: How AI Is Transforming NASA’s Engineering Hosted by Nathan Rigoni | Guest: Thomas Brooks – Aerospace Engineer, NASA Marshall Space Flight Center Advanced Concepts Office

In this episode of The Phront Room we dive into the ways artificial intelligence is reshaping space exploration—from automating routine meetings to enabling autonomous robots that can explore distant worlds without human latency. Thomas Brooks shares real‑world examples from his work on spacecraft conceptual design, thermal systems, and the emerging “agentic coding” tools that are already helping engineers write code, generate designs, and even write entire websites. What if a rover could decide its own path on Mars while you sip coffee on Earth?

What you will learn

• The core challenges of spacecraft design (size, weight, power) and how AI‑driven “feasibility engineering” can speed up early‑stage trade studies.
• How AI‑powered tools such as Klein and anti‑gravity enable natural‑language command‑line automation and rapid website generation for engineering documentation.
• The concept of “edge AI” on rovers and drones that can make real‑time decisions without waiting for ground control.
• How large‑language models can act as searchable knowledge bases for massive technical documents (RAG – Retrieval‑Augmented Generation).
• The social and ethical implications of AI‑augmented space programs, from data‑center heat management to the future of engineering work identity.

Resources mentioned

• Cline(local LLM assistant) – https://cline.bot
• anti‑gravity (cloud‑linked LLM) – https://anti‑gravity.ai
• Microsoft 365 Copilot – https://www.microsoft.com/microsoft‑365/copilot
• GSFC chat‑RAG tool – internal NASA implementation (referenced in discussion)
• NASA’s Marshall Space Flight Center Advanced Concepts Office (overview) – https://www.nasa.gov/centers/marshall/advanced‑concepts

Why this episode matters
Space missions are limited by mass, power budgets, and communication latency. AI can compress design cycles, automate drudge work, and give robots the autonomy to operate safely in hazardous environments—potentially increasing mission success rates while reducing cost and risk. At the same time, Thomas raises critical questions about control, transparency, and the future role of human engineers, making this conversation essential for anyone interested in the next frontier of aerospace AI.

Subscribe for more insights, visit www.phronesis‑analytics.com, or reach out at nathan.rigoni@phronesis‑analytics.com.

Keywords: AI in space, NASA engineering, autonomous rovers, edge AI, LLM‑assisted design, RAG, agentic coding, spacecraft conceptual design, thermal engineering, AI ethics, future of work.

open.spotify.com – Podcast overview and episode list for The Phront Room (including “AI in Engineering for Space” with Thomas Brooks).

AI in Education Hosted by Nathan Rigoni | Guest: Travis Bailey – Veteran middle‑school teacher in Huntsville, AL, with 15 years of classroom experience, exploring how AI is reshaping teaching and learning.

Teachers say the biggest change in recent years isn’t the curriculum—it’s the students. From post‑COVID digital classrooms to AI‑driven lesson planning, educators are wrestling with a new reality. What will happen to the role of teachers when AI can instantly generate quizzes, give real‑time feedback, and even act as a one‑on‑one tutor for every student?

What you will learn

• How AI tools like Class Companion and AI‑generated quizzes can cut lesson‑planning time from hours to minutes.
• Ways AI can provide automated, first‑pass essay assessment and personalized feedback for students at any reading level.
• The challenges of integrating AI responsibly: preventing over‑reliance, maintaining critical‑thinking skills, and handling academic honesty.
• Strategies for scaling instruction—using AI to raise the difficulty of assessments rather than lowering standards.
• Future scenarios for school structure, from shortened school days to AI‑mediated self‑paced learning.

Resources mentioned

• Class Companion (AI‑enhanced writing‑assistant for students) – https://class‑companion.com
• ChatGPT / GPT‑4 for creative projects and quick quiz generation – https://openai.com/chatgpt
• “CAMI” PDF‑editing and summarization tool – https://cami.ai
• Microsoft 365 Copilot (referenced as a corporate AI assistant) – https://www.microsoft.com/microsoft‑365/copilot

Why this episode matters
Education sits at the intersection of technology, equity, and society. Understanding how AI can amplify teacher productivity while preserving the human elements of mentorship and critical thinking is essential for anyone shaping the next generation of learners. As AI adoption accelerates, teachers who embrace these tools will stay relevant, whereas those who resist may see their roles diminish — a shift that will affect students, schools, and the broader workforce.

Subscribe for more deep dives, visit www.phronesis-analytics.com, or email [email protected] to learn how AI can work for you.

Keywords: AI in education, classroom automation, AI‑generated quizzes, essay assessment, Class Companion, personalized learning, critical thinking, post‑COVID schooling, teacher workload, future of schooling.

Audio Processing Basics – Hosted by Nathan Rigoni

In this episode of The Phront Room we dive into the world of sound, breaking down how raw audio waves become the speech‑to‑text and text‑to‑speech systems we use every day. From the historic phonograph to modern wake‑word assistants, we explore the science behind capturing pressure changes in air and turning them into meaningful symbols. What if you could teach a tiny model on a smartwatch to hear a single “help” call inside a burning building?

What you will learn

• How sound is represented as a time‑series waveform (the “signal”).
• The difference between signals and symbols (phonemes) in audio models.
• How speech‑to‑text models learn statistical variations of pronunciation from millions of samples (e.g., Librivox recordings).
• The reverse process: converting text back into audio via phoneme generation.
• Real‑world edge‑AI use cases, such as wake‑word detection on phones, watches, and drones for firefighting safety.

Resources mentioned

• Librivox open‑source audio books – https://librivox.org
• Mozilla DeepSpeech (open‑source speech‑to‑text) – https://github.com/mozilla/DeepSpeech
• OpenAI Whisper (robust speech transcription) – https://github.com/openai/whisper
• Google Assistant / Siri wake‑word technology (commercial examples).
• Research on phoneme‑based TTS models (e.g., Tacotron, WaveNet).

Why this episode matters
Understanding audio processing gives you the toolkit to build smarter, smaller AI that can act on sound in real time. Whether you’re creating a personal assistant, a voice‑controlled robot, or a life‑saving rescue drone, the principles covered here form the foundation for any application that turns vibration into insight.

Subscribe, learn more, and get in touch
Visit www.phronesis-analytics.com for deeper dives, tutorials, and consulting services. For questions or collaboration, email [email protected]. Don’t forget to hit subscribe so you never miss a future episode!

Keywords
audio processing, speech‑to‑text, text‑to‑speech, phonemes, wake word, edge AI, small models, drones, firefighting technology, machine learning, signal vs. symbol.

Basics of Image Processing Hosted by Nathan Rigoni

Ever wondered how a computer can “see” a picture and even generate new videos from a single frame? In this episode of The Phront Room we break down the fundamentals of image processing—from pixels and RGB channels to convolutional neural networks that turn raw visual data into actionable AI insights. By the end you’ll understand how machines compress, filter, and reconstruct images, and why these techniques power everyday tools like Instagram’s background‑removal and next‑frame video prediction.

What you will learn

• How an image is represented as a matrix of pixels and color channels (RGB values 0‑255).
• What a convolution is and how kernels compress and denoise visual data.
• The role of convolutional neural networks (CNNs) in tasks such as edge detection, segmentation, and video frame prediction.
• Real‑world examples of image‑processing applications (e.g., background removal, video generators).

Resources mentioned

• Convolutional Neural Network basics (kernel, stride, pooling).
• Instagram’s background‑removal tool as a practical segmentation example.
• General AI/ML concepts covered in the “Basics of Image Processing” episode listing on Spotify open.spotify.com.

Why this episode matters
Understanding image processing is a gateway to a wide range of AI capabilities—from simple photo editing to advanced autonomous systems. By mastering how CNNs extract and preserve essential visual features, you can build faster, more accurate models and unlock new use cases such as real‑time video analytics, medical imaging, and safety‑critical robotics.

Stay connected
Explore deeper dives, tutorials, and consulting services at www.phronesis-analytics.com. Have questions or want to collaborate? Reach out at [email protected]. Don’t forget to subscribe so you never miss a new episode of The Phront Room!

Keywords: image processing, convolutional neural network, CNN, pixel, RGB channels, kernel, convolution, edge detection, segmentation, background removal, video frame prediction, AI vision.

AI in Tokenization: Byte‑Latent Transformers, H‑Net & Bolmo Hosted by Nathan Rigoni | Guest: Jordan Conragan – Research Engineer at 11 Labs (formerly a Lockheed Martin colleague)

How can we make language models treat every byte of text as efficiently as a byte‑level transformer, and what does that mean for the future of AI‑driven computation? Could dynamic patching and entropy‑based chunking finally solve the tokenization bottleneck that limits model size, speed, and math reasoning?

What you will learn

• The motivation behind Byte‑Latent Transformers and why plain byte‑to‑token mapping explodes sequence lengths.
• How entropy‑driven patching groups low‑information bytes into larger tokens, shrinking effective sequence length while preserving information density.
• The design of H‑Net’s hierarchical dynamic chunking: a learned, end‑to‑end routing module that replaces a separate tokenizer‑training step.
• Bolmo’s approach of using a non‑causal LSTM boundary predictor to adapt existing LLMs for byte‑level input with minimal compute.
• Practical implications for math‑heavy workloads, model compression, and the bits‑per‑parameter efficiency debate.

Resources mentioned

• Byte‑Latent Transformer paper – “Byte‑Latent Transformers: Patches Scale Better Than Tokens.” https://arxiv.org/abs/2412.09871
• H‑Net (Hierarchical Net) paper – “Dynamic Chunking for End‑to‑End Hierarchical Sequence Modeling.” https://github.com/goombalab/hnet
• Bolmo (Allen AI) blog post on OLMO 3 and its byte‑level adaptation. https://allenai.org/blog/bolmo
• 11 Labs Scribe V2 (mentioned as Jordan’s current project).

Why this episode matters
Tokenization is the hidden cost driver behind today’s trillion‑parameter models. By moving from fixed sub‑word vocabularies to entropy‑aware, dynamically sized patches, we can dramatically reduce sequence length, lower compute budgets, and improve numerical reasoning—key steps toward making large language models more accessible, faster, and better at math. The discussion also surfaces the trade‑offs of maintaining a separate tokenizer versus learning chunking jointly with the model, a design choice that will shape the next generation of efficient AI systems.

Subscribe for more deep dives, visit www.phronesis-analytics.com, or email [email protected].

Keywords: Byte latent transformer, byte‑level tokenization, entropy‑based patching, dynamic chunking, H‑Net, Bolmo, model compression, bits‑per‑parameter, AI efficiency, mathematical reasoning.

AI in Dentistry: Transforming Practice and Expanding Care
Hosted by Nathan Rigoni | Guest: Ben Kellum – dentist, founder of Transcend Dental Implants and co‑director of the Restore Dental nonprofit. ⁠https://transcenddentalimplants.com/⁠⁠https://restore-dental.org/⁠In this episode of The Phront Room open.spotify.com we explore how artificial intelligence is reshaping the dental industry—from automating the design of patient‑specific teeth to streamlining office administration and making high‑quality care more affordable. What if a patient could see a realistic, AI‑generated preview of their new smile in minutes, instead of waiting days for a technician to craft a model?

What you will learn

• How AI‑driven software can automatically generate tooth designs from 3D scans, cutting design time from hours to minutes.
• The role of AI in aligning multiple data sources (X‑ray, facial scan, intra‑oral scan) to create a single, accurate treatment plan.
• Real‑world examples of AI‑powered CAD generation, rapid 3D printing of prosthetics, and AI‑enhanced patient simulations.
• Ways AI can take over repetitive administrative tasks such as appointment reminders, insurance verification, consent‑form analysis, and grant‑writing for nonprofits.
• Strategies for building trust in AI tools, including “getting to know” the system and using it as a first‑pass assistant rather than a final decision‑maker.

Resources mentioned

• AI‑based audio reverb removal tool (used to clean up the podcast recording).
• 3D scanning and imaging platforms for dental data acquisition.
• CAD software that converts AI‑generated tooth models into STL files for 3D printing.
• USD (Universal Scene Description) pipelines that turn patient videos into 3D models usable in game engines like Unreal Engine.
• General‑purpose LLMs (e.g., ChatGPT) for resume polishing, grant drafting, and business‑process automation.

Why this episode matters
Dentistry is at a tipping point where manual, labor‑intensive workflows can be replaced by intelligent automation, dramatically lowering costs and expanding access to high‑quality care for underserved populations. By automating design, simulation, and back‑office tasks, dentists can spend more time on patient interaction and less on paperwork, while nonprofits can leverage AI to streamline grant applications and reduce overhead. Understanding these technologies now positions practitioners to lead the next wave of democratized dental health.

Subscribe for more insights, visit www.phronesis-analytics.com, or email [email protected].

Keywords: AI dentistry, dental automation, AI‑generated tooth design, 3D scanning, CAD dentistry, 3D printing, patient simulation, dental nonprofit, AI for business administration, AI trust, democratizing dental care.

AI in NLP: From Word2Vec to Modern Language Models Hosted by Nathan Rigoni

Artificial intelligence has turned the simple act of “reading a document” into a powerful engine for insight—​but how did we get from counting words to building whole knowledge bases? What if you could ask a computer to “find every contract clause that mentions liability” and get an instant, accurate list without scrolling through pages of text?

What you will learn

• How early NLP models like Word2Vec turned tokenized words into vectors that capture meaning.
• The linguistic principles behind tokenization and context‑based learning (blank‑out‑word training).
• Why hidden‑state representations enable tasks such as topic analysis and document classification.
• The role of assumed context in information theory (Shannon) and how it shapes modern large language models.
• Common pitfalls such as hallucination and why being explicit with prompts improves results.

Resources mentioned

• Word2Vec (original word‑embedding model).
• Claude Shannon’s Information Theory (basic concepts of bits, messages, and context).
• Ludwig Wittgenstein’s philosophical work on meaning and context.
• Retrieval‑Augmented Generation (RAG) as a technique for grounding LLM responses.

Why this episode matters
Understanding the lineage from Word2Vec to today’s LLMs gives you the toolkit to automate document processing, build smarter classifiers, and avoid the “black‑box” traps that lead to hallucinations. By mastering tokenization and context, you can turn raw text into actionable data, a decisive advantage for any business or researcher navigating today’s information overload.

Subscribe for more deep dives, visit www.phronesis-analytics.com, or email [email protected].

Keywords: NLP, Word2Vec, tokenization, word embeddings, hidden state, auto‑encoder, linguistics, Claude Shannon, information theory, context, assumed context, large language models, hallucination, retrieval‑augmented generation, RAG.

Ludwig Wittgenstien: https://en.wikipedia.org/wiki/Ludwig_Wittgenstein

Claude Shannon - Information Theory: https://en.wikipedia.org/wiki/Information_theory