10-27-2025 Doyle’s Tech Talk Lecture: Evolution of Computing, NLP
Fundamentals, and Practical AI Use Cases
Date & Time: 2025-10-27 01:30 PM
Location: Brookdale Skyline Auditorium (SKR)
Lecture: Evolution of Computing, NLP Fundamentals, and Practical AI Use Cases
NLP and statistics Prompt
engineering Python ecosystem
Theme
A veteran engineer traces computing from mainframes and
Fortran through VBA and Python to today’s generative AI, emphasizing his
personal journey and exposure to computers and languages. The talk explains
Natural Language Processing (NLP) as large-scale statistical modeling, why
neural networks demand heavy compute power, and how natural language prompts
guide AI outputs. It covers practical tools for transcription and
summarization, cautions on AI reliability, and plans for hands-on demos across
various generative AI platforms like ChatGPT, Copilot, Gemini, and Claude. The
session ties historical context from the speaker's nearly 40-year career to
modern AI practices and resources, highlighting the ability to communicate with
computers through spoken language.
Takeaways
- Lecturer’s background:
Considers Pueblo, Colorado his hometown; graduated from South High School
as a member of its first graduating class in 1960; attended two years at
Pueblo Junior College; subsequently transferred to CU Boulder, earning a
degree in electrical engineering with an electronics emphasis in two
years. His single computer class was in the spring semester of his senior
year, using a vacuum-tube version IBM 709, before a computer science
department even existed.
- Early computing exposure:
Observed the IBM 709 behind glass doors with its big bank of lights
flickering during processing; the comprehensive course introduced the
structure of computers and the idea of computer programming, covering
assembly language (the actual instructions a computer executes) and
higher-order languages like an early version of Fortran 2 and MAD
(Michigan Algorithm Decoder), an early interpreter language. The course
also covered the differences between interpreters, compilers, and assemblers,
which relate to the level of interaction with the computer.
- Career path: Based on this
single computer course, he secured his first job with Control Data
Corporation in Minneapolis, Minnesota. He lasted about a year and a half,
finding it "wasn't quite what I expected for a career," partly
due to severe winters (below zero for a week straight, little sun). He
moved back to Colorado, working with defense department contractors on
early systems for Cheyenne Mountain space surveillance and satellite
tracking. He wrote code to perform tasks associated with maintaining catalog
satellites, a career he pursued for nearly 40 years.
- Programming languages over
time: Exposed to many evolving computer languages, including Fortran 2 and
Fortran 4, writing a lot of Fortran code. Fortran evolved significantly
over his career, incorporating capabilities from other languages like
pointers and structures; he believes Fortran is still around, the latest
version is named "Fortran 2023." He also encountered a series of
interpreter languages beyond MAD, such as ALGOL, SNOBOL (spoken as
SNOWBALL), and COBOL (Common Business Oriented Language), noting he
studied COBOL but never used it for a job. More sophisticated
quasi-assembly language programs like C and C++ emerged, described as more
object-oriented and having less rigorous syntax.
- Visual Basic for
Applications (VBA) in Excel: Worked with Visual Basic, the language
"underneath our modern spreadsheet Excel," for writing macros.
He developed a complex Excel-based system of macros that would
"scrape Fidelity's website" to pull information for stock
analysis, applying statistical math and interfacing with other components.
This system became unusable when Fidelity changed their website,
illustrating the dynamic and varied evolution of programming concepts and
tools.
- Core thesis on modern AI:
The advent of generative text AIs like ChatGPT, Copilot, Gemini, and
Claude 4 has created the ability to communicate with computers through
spoken language. Prompts are the interface for these AIs, and the more
specific a prompt, the more specific the answer will be.
- Two initial questions
about ChatGPT: When first interested in ChatGPT, he had two fundamental
questions: (1) How does it understand and manipulate language? (2) Why
does it consume so much compute power?
- NLP understanding: Natural
Language Processing (NLP) is the computer science field that digitizes
language into a form computers can manipulate. He drew an analogy to
satellite orbit determination: classical Keplerian orbits could be
represented with six basic, purely analytical parameters, extended to nine
parameters with additions for atmospheric drag, solar radiation pressure,
and three dimensional thrusts. In contrast, NLP uses vastly more
parameters, such as 2.67 billion, to represent and manipulate language. He
emphasizes that AI, at its heart, is "all statistics."
- Statistical methods:
Satellite orbit determination used statistical processes like differential
correction least squares and probability to fit orbits to observations and
make predictions. Similarly, NLP employs statistical approaches, but with
a much greater number of parameters. AI is framed as sophisticated
statistical manipulation, not a single "do AI" instruction, but
rather "very complicated, evolved, sophisticated statistics."
- Python ecosystem: Python
has evolved as the language to support AI, despite the speaker not knowing
"why it’s called Python." Its basic language concept is simple,
but its capabilities are extended through various libraries (e.g., for
math, as basic math instructions are not built into the core language).
Users can import or fabricate their own libraries, making its capabilities
"literally unlimited." Python's interpretive nature allows it to
modify itself while being executed.
Highlights
"At the heart of it, it’s just statistics — very
complicated, evolved, sophisticated statistics — but still statistics."
Chapters & Topics
Historical evolution of programming exposure and languages
The lecturer’s journey from early mainframe-era
computing through decades of professional programming, illustrating how
languages and paradigms evolved from assembly and early Fortran to modern
high-level and object-oriented languages, and into end-user programming via
VBA.
·
Keypoints
o
IBM 709 vacuum tube version computer used during
a senior-year course before a CS department existed.
o
Assembly language as direct machine instruction
interface contrasted with higher-order languages (Fortran 2, MAD).
o
Interpreter vs compiler vs assembler
distinctions based on interaction level with the machine.
o
Fortran evolution (Fortran 2, Fortran 4) and
adoption of capabilities like pointers and structures; assertion that modern
day Fortran persists as “Fortran 2023”).
o
Exposure to ALGOL, SNOBOL (spoken as SNOWBALL),
COBOL (Common Business Oriented Language, studied but not used professionally),
C, C++ (object-oriented, less rigid syntax).
o
VBA enabling Excel macros and full program-like
systems.
·
Explanation
The lecturer details a chronological path: an early, comprehensive course
introduced assembly and high-level languages, setting up employability at
Control Data. Over his nearly 40-year career, Fortran and other languages were
used in defense and satellite tracking contexts. Interpreter languages (MAD,
ALGOL, SNOBOL) differ in runtime behavior from compiled languages (Fortran, C).
C and C++ introduced object orientation and less rigid syntax demands. In
end-user computing, VBA underpins Excel macros, enabling tooling that can be
sophisticated enough to scrape websites and perform statistical analysis,
though such systems are brittle to external changes (e.g., site redesigns).
·
Examples
Upon
moving to Skyline, the lecturer built a set of Visual Basic programs in Excel
to scrape Fidelity’s website for stock analysis. The system applied statistical
math and interfaced with other components, growing complex over time. When
Fidelity changed their website, scraping stopped working and the project became
unusable.
o
Identify data source (Fidelity pages) and target
metrics for stock analysis.
o
Prototype VBA macros to navigate and extract
HTML content.
o
Iteratively add statistical routines and
integrations as complexity grows.
o
Monitor for site changes; when DOM
structure/endpoints changed, scraper failed.
o
Recognize the brittleness of scraping and
contemplate API-based alternatives.
Natural Language Processing (NLP) and parameterization
NLP transforms human language into machine-manipulable
representations and uses large statistical models with billions of parameters
to predict and generate text.
·
Keypoints
o
NLP is the field that digitizes language for
computation.
o
Satellite orbit analogy: 6 classical Keplerian
parameters (purely analytical) plus 3 extensions for drag, solar radiation
pressure, and unknown thrusts (total 9) versus NLP with 2.67 billion
parameters.
o
Statistical fitting in orbits uses least-squares
differential correction and probabilistic predictions; NLP analogously relies
on statistics at far greater scale.
·
Explanation
By comparing orbit determination to NLP, the lecturer emphasizes scale. Orbit
models require nine parameters to fit observations; NLP models like large
language models utilize on the order of billions (e.g., 2.67 billion) of
parameters to capture nuances of grammar, meaning, and context. This vast
parameterization underlies the model’s ability to ‘understand’ and generate
language, but remains fundamentally statistical.
Neural networks and compute power
Neural networks consist of interconnected nodes enabling
parallel computation; modern systems distribute work across thousands of
nodes/CPUs, with physical interconnects becoming a limiting factor.
·
Keypoints
o
Neural network metaphor compares to a model of
the human brain, with many billions of neurons.
o
Parallelization contrasts with sequential
step-by-step processing.
o
Thousands of nodes are deployed; individual CPUs
or computers can be assigned to nodes for distributed computation.
o
Inter-node distance and communication become
bottlenecks as scale increases, limiting projected computer power.
·
Explanation
The lecturer answers why generative AI consumes substantial power:
training/inference leverage parallel architectures (e.g., clusters of
CPUs/GPUs/TPUs). Data flows between nodes, and synchronization/communication
overheads introduce physical constraints. The ability to parallelize makes
large models fast enough to be practical, but at high energy cost, with the
distance between nodes being a key limiting factor.
Prompt engineering basics for generative AI
Interacting with generative AIs via natural-language
prompts; specificity in instructions yields more precise outputs; author styles
can be emulated within bounds.
·
Keypoints
o
Prompts are the primary interface—no code
required for end user.
o
More specific prompts lead to more specific
answers.
o
Author style emulation examples: love letters in
styles of Jack Kerouac, F. Scott Fitzgerald, Ernest Hemingway; approximately 20
such prompts tried, yielding "entertaining" results that referenced
authors' works and phrases.
o
A resident suggested an Emily Dickinson, a
famous lady poet, style that yielded poetry-like output.
o
Different platforms: ChatGPT (Open AI), Copilot
(Edge), Gemini (Google), Claude 4; marketing terms like “deep reasoning” are
mentioned.
·
Explanation
The lecturer demonstrates by composing prompts that specify task (write a love
letter) and style (named authors). The systems output text that reflects
stylistic cues learned from training data. Platform selection varies by
device/browser; the principle remains that clarity and constraints in prompts
guide model behavior.
Python and AI tooling
Python’s simplicity and extensibility through libraries
make it a de facto language for AI and data tasks; its interpretive nature and
importable libraries (including math) underpin flexible development.
·
Keypoints
o
Python core syntax is simple; capabilities
extended via imports.
o
Math and many domain libraries are brought in as
needed, as basic math functions are not built into the core language.
o
Interpretive execution allows dynamic behavior
and rapid iteration, enabling the language to modify itself while executed.
o
Library ecosystem makes capabilities seemingly
unlimited.
·
Explanation
The lecturer notes that language choice (Python) is pragmatic: easy to
read/write and supported by vast libraries (though specific libraries like
NumPy, pandas, PyTorch, TensorFlow are not explicitly named). This allows
building, training, and serving models, and scripting data workflows.
Risk management and reliability of AI outputs
AI-generated content may be incorrect; users should be
cautious, especially with legal and financial decisions; AI verbosity and
variability mirror human fallibility.
·
Keypoints
o
Standard disclaimers: AI-generated information
may be incorrect.
o
Avoid using AI for legal or financial decisions.
o
AI can produce long expositions (e.g., the
speaker prompted it to expound on why AI can be incorrect, and it "spewed
out three pages").
o
Human analogy: people also provide incorrect or
variable answers depending on phrasing, suggesting "a little bit of hint
[of] human characteristics in this AI."
·
Explanation
The lecturer emphasizes critical thinking: treat AI outputs as drafts or aids,
not authoritative sources, especially where stakes are high. Validate important
information with trusted sources, acknowledging that AI's fallibility can be
compared to human inconsistency.
Applied use cases for meeting, medical, and social contexts
Practical deployments of AI agents to record,
transcribe, summarize, and structure information from meetings, medical visits,
and casual conversations, with notable robustness in noisy environments.
·
Keypoints
o
AI agent called “PLAUD AI was adopted by the
lecturer.
o
Used to record meetings in his role as vice
chair, producing text summaries and structure.
o
Used to capture and summarize medical visits and
needed follow-ups.
o
Applied to insurance interactions and social
breakfasts, producing summaries.
o
Demonstrates selective listening in crowded
restaurants, focusing on two voices and ignoring background noise; meeting
audio summarized effectively despite filler words and disfluencies, making it
"sound like we were so well organized."
·
Explanation
The lecturer records audio, then leverages AI transcription and summarization
to surface action items and clarity. Noise robustness suggests advanced
diarization and source separation. This enhances personal productivity and
recall, providing "pretty interesting" summaries even from
disorganized speech.
Live demo and community engagement plan
Within a one-hour Brookdale Skylines Tech Talk, the
lecturer planned background coverage, AI overview, prompting demos on
attendees’ iPhones, and Q&A.
·
Keypoints
o
Series: Brookdale Skylines Tech Talk.
o
Allocated time: one hour.
o
Audience was asked to try ChatGPT/Copilot/Gemini
via Edge/Google or app store; GPT-4, Claude 4 also mentioned as available
options.
o
Open Q&A to conclude the talk.
·
Explanation
A participatory format encourages hands-on experience with generative AI,
reinforcing prompt engineering concepts and platform familiarity, with the
speaker planning to introduce attendees to using these tools on their personal
devices.
Music interlude and personal creativity
The lecturer’s piano hobby illustrates lifelong learning
and creative practice, aiming to perform to conclude the talk.
·
Keypoints
o
His hobby is "playing the piano, or as I
say, practice the piano."
o
Practices scales and improvisation in scales.
o
Planned to play "a few choruses" of
his favorite tune, “St. Louis Blues,” followed by a classical piece practiced
extensively (left hand, right hand, then combined).
o
Expressed curiosity about whether AI could
record/interpret piano notes.
·
Explanation
The performance serves both as a personal touch and a metaphor for iterative
skill-building—akin to refining prompts or models through practice—and a way to
"summarize and wrap up this meeting."
Reading list for understanding and building with AI
Suggested materials range from conceptual understanding
to practical development.
·
Keypoints
o
“What is ChatGPT Doing ... and Why Does It
Work?” by Stephen Wolfram who invented the Mathematica language and founded his
own company named Mathematica.
o
“AI Made Simple” a bookazine for a survey of
products (limited on installation/cost details), which is available at
www.magazinesdirect.com. ( costs $26 hard copy)
o
“ChatGPT: The API Bible” for constructing AI
agents (deep programming-oriented, for those who know how to program).
·
Explanation
The resources target different depths: conceptual mechanics, user-facing
overviews, and developer-level implementation. Selection depends on goals and
technical comfort, offering pathways for both casual users and those interested
in deep dives.
