Thinking, Computing, and Improving Both
By Paul VanRaden
2025-2026
Introduction
Thinking lets humans understand, plan, and decide what to do
instead of relying on instinct alone as other animals do. Computing now does
many tasks for humans and helps us better understand, plan, and decide than our
ancestors did. We now hold computers in our hands, use huge computer clusters
far away, and interact with Artificial
Intelligence (AI) using methods that few people understand. Thinking is
different than computing but studying computers may help our brains think more
clearly. Also, learning how a person’s brain organizes its thoughts may help AI
to guide us better. I even like to count in
binary on my fingers because that makes me feel more like a computer.
After 40 years of writing computer programs to process large,
complex datasets, I wondered if directly comparing thinking to computing could
help us discover what each does best, how they interact, and how to improve
both. Your brain may think more clearly and more reliably using your neurons
designed over millions of years of selection instead of using AI based on
silicon chips invented the same year I was born – 1960. Young people who grew
up with computers may want to learn how previous generations did anything useful
with few or no computers. Before 1960, computing was more of a “hands on”
experience as shown in The Imitation Game
excellent movie about code breaking during World War II. Today’s computer
algorithms and large datasets may ‘imagine’ an answer, but your lifetime of
real-world experience helps you understand what really works.
Topics
Thinking
Computing
Artificial intelligence
Connections and ideas
Sorting
Paging and memory management
Linked lists
Parallel processing
Hardware and processors
Multitasking
Data storage, access, and cost
Cause and effect
Probability and Bayes theorem
Interacting with AI
Background and advice
Reviewer comment
Thinking
To think, you need no extra hardware, only your brain and its
8 or 9 billion neurons. You have more than twice the number
of neurons of any ape but fewer neurons than elephants or some whales. To
think clearly, your neurons need to first make sense of inputs from your 5
senses: sight, sound, touch, taste, smell. Thinking without acting is very useful
for practicing, but ideally your thoughts can lead to actions such as doing
each day’s tasks or improving your own or others’ lives.
You also need no extra hardware to act on your thoughts. You
can talk or smile or blink or move your fingers or hands or feet using your
muscles. But sometimes extra hardware can help your senses to get more or
better inputs or help your muscles send your actions further or faster. You can
record your thoughts and store your words on paper using crayons or pencil or pen
or store them in a report by typing on a keyboard or keypad. Voice recorders
and cameras can also store your sounds and actions for wider or later use. The
internet can send your thoughts and show your actions to anybody in the world
interested if they have electricity, a computer or smart phone, and internet
access that is not blocked by a government that controls what they see.
Computing
Brains have only hardware and no
software. You cannot copy your internal hardware or the methods you use for
thinking directly to any other person. Each brain is unique and each person may
answer the same question differently based on your own lived experience, stored
data, and thought process. Our brains have first-hand experience from years of
living in this world whereas AI only has second-hand knowledge which is not
admissible in a court of law. Being able to copy software can be a huge
advantage for machines. For example, one robot could train and get good at a
task after 100 hours, great after 1,000 hours, and nearly perfect after 10,000
hours. Copying its software can make any number of other robots just as skilled
in seconds with no additional training.
Artificial intelligence
My Agriculture Economics 436
instructor at Iowa State in 1983 taught us an early form of AI. His example
used primitive AI instead of your own brain to schedule where trucks should go
to deliver orders to scattered customers. He gave us several minutes to plan
the best routes that 2 imaginary feed trucks should drive to go the fewest
miles while delivering the amount of feed that each customer ordered. The next week
after checking and tracing how far each student’s trucks traveled, he handed
back our “quiz” results. Then he handed me a real quarter (25 US cents) and
explained that my imaginary trucks were the only ones in the class that
traveled fewer miles than the routes produced by AI. That 1983 lesson convinced
the class that computer programs can do important business tasks better than
most college students can.
Econometric statistics were probably my
training closest to AI because in an economy, almost anything can impact
anything else and in both directions. Causes and effects are not very clear.
Genetic statistics were much easier because genes affect almost all of life,
but the process of living does not affect the genes except for a few rare
mutations which may turn into cancer or other diseases. Those who learned AI
tools sometimes got higher paying jobs such as from Facebook where AI may have larger
benefits over standard statistics that model direct causes and effects.
Artificial Intelligence has much room
for improvement but could improve quickly. Next, I summarize 3 separate
experiences with AI during the last month: 1) meeting notes from Zoom AI, 2) my
241-page humorous biography story generated by AI from a few quick inputs, and
3) a strange answer from MS Copilot to an internet search question.
For our weekly meetings of my siblings, my sister
Miriam often turns on Zoom’s Artificial Intelligence to take notes in case one
of us is missing and its summaries usually make sense. When my daughter Angel
told them the news that she is pregnant, AI got confused because Angel shared
my camera and microphone to announce her news. AI generated these sentences: “Paul shared his excitement about
his first ultrasound appointment on Friday, revealing he is 6 weeks pregnant…
Paul shared that he is experiencing fatigue as a symptom of his pregnancy, and
discussed his diet changes to support the health of his unborn child… The group
discussed Paul's experience with an AI-generated book gift from his wife,
Angel…” Zoom’s disclaimer said “AI can make mistakes. Review for accuracy. Please rate the accuracy
of this summary” and then asked for feedback. So, I explained the facts
of life to AI and reported how the issue was caused by sharing a microphone.
AI generated a Christmas present for me: a
241-page biography from Book By Anyone as a gift from Angel. Many of the 20 chapters read like a novel but
chapter 3, “Agricultural Genetics: DNA is More Than Just a String of Letters”,
was excellent. It was mostly about crop breeding and clearly explained methods
and benefits in a friendly, convincing, humorous style that few individual
human authors could match. AI had trouble keeping track of family relationships
and trouble with timelines, but its best sentences seemed better than any
similar sentences I could find on the internet.
To generate the book, the company website
asked questions such as “Who are Paul’s family members?” Angel listed the names
and ages of my 4 siblings, but AI somehow generated this sentence on
page 14 “Then we have my siblings, the brilliant trio
of Miriam, Mark, and Deb” and on page 15 “my
childhood growing up in a family of not one, not two, but three siblings
- Miriam, Mark, and Deb.” Page 138 said “During
family dinners, I often find myself surrounded by my six siblings.” Page
105 finally mentioned my youngest sister, saying: “And
then there’s my sister Judy” and told a story about her, but
pages 191-192 instead referred to “my Aunt
Judy.” Page 106 said “Then there’s my dear sister
Charlee” and told a story about my stepdaughter Charlee. Page 140 said “In fact, my wife Charlee often…” and page
153 finally correctly said “Now let’s talk about my
stepdaughter Charlee.” But page 207 said “My
sister Charlee rounds out the dynamic quintet” of siblings
and page 208 said “And then,
let’s celebrate my step-sisters, Deb and Judy.” Pages 233 and 237 then said “my Aunt
Miriam.”
The confusion above suggests that this AI
treated the different types of close relatives as synonyms or adjectives
instead of mutually exclusive nouns. Since those relationships are the basic
inputs supplied by the customer, AI should take them more seriously. Such problems
are often easy to fix by hard-coded rules imposed on the AI algorithm. Our
editing system at USDA had hundreds of thousands of lines of code to enforce
specific rules. Getting those edits correct took much work by real human
programmers. My biography book reads like a novel, but AI looks less
intelligent and customers will not buy or read novels if close relatives change
their relationships so often during the story.
AI generated this sentence about me: “If I can brag about one more genius superpower that I
possess, it would be the ability to take disparate ideas and stitch them
together into a coherent thought pattern.” But smart people do not waste
time bragging. Instead, a better sentence might say “With practice, most people can connect different ideas
into better models or predictions of what will happen by constantly studying
how the world works.” AI did not yet read my Easy English report. I would not use the word “disparate” because its synonyms “different” or “very
different” derive from the simple word “differ”
that more people know, and I would say “have”
instead of “possess” because you cannot buy
or sell superpowers and no one can steal genius from you after you have it.
AI’s sentence “Every genius’s journey begins with
some off-the-wall ideas perched precariously on the edge of sanity” seems
accurate to me.
AI seemed to invent a phrase opposite of Iron
Curtain when it wrote this sentence: “I’ve often
found myself dreaming up curtains of equal opportunity sweeping through every
nook and cranny, where everyone gets a fair shot to shine.” I would rather have equal opportunity curtains than an Iron
Curtain any day because a few billion people have not got their fair shot yet.
The book was interesting to read because it mixed words and phrases into new
combinations used by few humans. I found some AI slop but a few AI gems. It reminded
me of talking to people who grew up with other languages, and I borrow and use
their real language gems if their phrases are easier than my English.
Generative AI did not quite understand
timelines. A sentence about my teenage years described “my
father attempting to fix his iPhone using duct tape” but in the 1970s we
were happy to get an individual land line instead of the previous party line we
shared with neighboring farmers. AI correctly reported that our dog Shadow is now
2 years old but used the name Shadow again for the family dog
from 50 years ago. AI described genetic theory and methods surprisingly
well except for this sentence suggesting that each cell contains different DNA:
“If DNA were a person, it
would definitely be the life of the party, wearing flashy colors to attract the
attention of all the other cells.” The book was interesting but did not use all the other information about
me already available on the internet.
Microsoft Copilot AI got confused by Trump’s towers. After visiting New York City, I asked
Copilot: “Is the UN building or Trump Tower NYC taller?” and it replied “Trump Tower. Trump Tower in NYC is taller than the
United Nations building. Trump Tower stands at 664 feet (202 meters),
while the United Nations building is 861 feet (262 meters) tall.”
But 861 feet is the height of Trump World Tower built in 2001 across the street from the 510 feet (155 meters) United Nations building built in 1952. Copilot probably reported the wrong height because both
buildings have addresses on United Nations Plaza. Long ago, the whole world was
united, speaking the same language, until a developer built a tall tower in
Babel and got the whole world confused (Genesis 11:1-9). Copilot gets confused by Trump’s towers, but I do not. The answer was
correct but then Copilot made a mistake when showing its work.
Later I asked Copilot
“how many cities > 100,000 population in Iran?” It
replied “57 cities. In Iran, there are 9 cities with populations exceeding 1
million. Additionally, there are 57 cities with populations between 100,000 and
1 million.” If AI cannot do second grade addition, I would not trust its advice
on more complex subjects. Perhaps the U.S. government instead asks AI “which country should we declare war on next?”
Connections and ideas
Reading
this article may cause your brain to connect ideas about how to think clearly
or compute efficiently that it previously may have stored separately. Those new
connections will be stronger if you agree that thinking and computing are
similar processes using different hardware. Based on your experiences, you may easily
imagine further ideas or examples that I did not or could not think of.
We may imagine that Thomas Edison was
thinking about a light bulb and then he reached over to the switch and turned
it on. Many good ideas start out as new neural connections but have no impact
because we do not have the time or the resources to develop our ideas into a
working product. By sharing ideas, we keep them alive so that others with more
time and energy might make them light up someday.
Sorting
Before electronic storage, having
multiple copies of each document in differing sort orders could be helpful. For
example, when I started at USDA in 1988, the letters we sent were copied 3
times and filed into 3 separate cabinets sorted by date, or alphabetically by
subject, or alphabetically by person’s last name.
That made finding the letters easier if you knew any of those 3 properties.
Even after electronic storage, multiple copies in different
sort orders were used for merging data types and in solving equations
because computer tapes require processing each large file sequentially. Sorting
became less important for us in 1993 with more memory, spinning disks, or later
when solid state drives allowed faster access to each record.
Similarly, your brain may store many
events in date order so that you can remember what happened first, perhaps to
help determine cause and effect. Your brain may also link an event to one or more
subjects and to those people most involved in that event. Your brain likely
does not link subjects or people to the alphabet but rather to the nerve group
where that subject and those people already are stored in your brain.
Electronic instead of paper files have made computer searching much faster but
our neurons still have the same speed as our ancestors and probably the same
speed as apes.
Memory management and paging
Your brain uses many more neurons to store
long term memories than to store or process new inputs from your senses each
day. While you sleep, more important events of the day may be converted from
short-term to long-term memories. Your neural connections will get stronger
among people and events that happened together, grow weaker for people and
events that you no longer see, and you will form new neural connections to help
you remember new people or new events.
Computers also have different memory types,
and short-term memory usually has the fastest recall. Fast computers and
algorithms should put the data they will need next into their fastest access
memory using a process called paging. IBM invited and flew scientists with the
most difficult problems to Los Angeles to test our programs on their newest
computers in cooperation with their research team. While there in 1987, I
convinced their team to use a new memory management strategy that they had been
testing for computers with multiple users. Instead of recalling the pages
needed for each program each time, the last used pages are left in memory until
that program or another program needs to replace them.
That memory management strategy is used in most computers today.
You might use similar memory
management strategies. Some students like to cram before big tests by trying to
load answers to the most likely questions into their quick-access memory so they do not need to dig for the answer deep in
their longer-term memory. I used the opposite strategy and tried to organize my
long-term memory efficiently so I would also remember the answers years later
instead of only on test day. I tried to always get an extra hour of sleep the
night before each big test.
When replying to emails, waiting a
few hours is often a good strategy to give your long-term memories a chance to
add to your reply. Pre-filling your short-term memory is also very useful, for example
by rereading the list of participants right before going into a meeting. I
always tried to do that to better recall names or faces of the people who would
be there. Remote meetings now often include each person’s name under their face
so that trick is no longer as useful. Pre-filling your memory is most useful
right before presenting to an audience, especially
pre-filling your opening lines, or if you have only a short time to discuss.
Recalling your address or your car’s license
plate number from long-term storage may be easy if you saw those many times,
but if a car hit yours today, you might forget its license plate unless you
repeat the number continuously until you write it down or store it externally.
Storing your long-term memories more efficiently can free up more neurons for
short-term processing. Doing full-time research continuously for 40 years
gradually shifts more of your neurons toward connecting new ideas instead of
repeating the old ideas that your teachers taught you. Those old ideas may
still be correct or need to be updated if they conflict with newer data or
better ideas.
Physical reminders can also free more
short-term memory for doing other tasks. While returning home from work each
day I tried to remember what foods I ate in previous days to help decide what
food I should cook next. In about 2010 I listed my meals on a paper I kept on the
kitchen countertop with a small cow pointed to the meal I had yesterday. Then
each day just before starting to cook I move the marker to today’s meal. This
evening, that same paper, list, and cow will tell me what to cook next, as it
has each day for the last 15 years with no short-term memory wasted on food
decisions. Putting physical reminders right next to your outer door can also
help you remember what you want to take out, etc.
Linked lists
Your brain and computers can use linked
lists to remember things such as family members. The first item can link to the
second, and the second to the third, etc. The list is complete when the next
link is null or links back to the first to form a circular file. For example,
my family of 8 is stored by linking my 2 parents to their first child Mark,
linking Mark to the next born Miriam, linking Miriam to the next born Deb, etc.
Linked lists are often more efficient than storing names in
a rectangular table by not reserving empty space for the maximum family size
possible. For cow pedigrees, we link each parent to all the daughters
separately from all the sons because separate searches are faster if sex is
already known and the families are very big with thousands of siblings.
Parallel processing
From
1960 to about 2000, each program used a single processor. Computer designs
focused on making each processor compute faster. More recent computer designs use
many more mass-produced processors that are cheaper instead of faster. In
earlier years, parallel processing simply chopped the input data into many
pieces and gave one piece to each processor, but often with some loss of
accuracy compared to seeing all the data at once. More recently, parallel
processing can let one program manage all the data together and control many
processors while each works on a subset of the problem or data.
Early computers were mechanical and simpler. In
my office I kept an abacus that my mother gave me and a slide rule that
scientist Bob Miller gave me. He used it in his research at North Carolina
State before being the first to use computers to process very large data files
at USDA in the early 1960s. His programs and one computer replaced about 100
employees whose job title was “computer.” Switching from hand calculations to a
computer in our laboratory was much like that in the movie Hidden Figures but happened a few years earlier
because dairy geneticists had much more data than NASA had.
The abacus and slide rule reminded me
to keep computing simple so it could be scaled up for large problems. In the
1990s I used to say that a billion Chinese people each with an abacus and using
parallel processing could do our computing faster than we could with a
computer, but that is no longer true due to much faster computing today. Also, Chinese
people now have much better and more interesting jobs than just using an
abacus, and American people now have much better jobs than just using an adding
machine or a slide rule to compute all day long.
Hardware and processors
Understanding how a computer works
can be important when trying to make programs run faster. Direct electronic
access to data with no moving parts (such as solid-state drives or USB files)
is often much faster than for data held on spinning disks. The data requested
can arrive at the speed of light instead of waiting a millisecond for its
section of rotating disk to reach the reader. Parallel programs may run slower
when a computer is busy if nearby hard-working processors raise the temperature
and cause the operating system to reduce their workload to avoid overheating.
Parts of your brain might also work too hard and need a break to avoid a
headache.
Computers keep getting faster and can be more helpful each year. Faster speed and cheaper data storage
means that future programmers might need fewer advanced coding skills. Some
compilers can automatically apply tricks that past programmers needed to know. Or, programs can use libraries of tools already optimized
for the hardware and safer than developing those tools again. Fast, cheap
computer chips are easy to use but very complex to design and build. This video shows how they are designed and
manufactured. People need to imagine that a computer can answer many questions
and then ask the chat system, or type the questions into a search engine, or learn
how to program to quickly find better answers. Computers can do things that you
cannot.
Multitasking
At any one time, millions of neurons
in your brain may be either growing or shrinking and
gaining or losing connections to other neurons. But humans have only 1 set of
language processing neurons and so we can only input or output one stream of
information at a time. We may have many thoughts, but we can only express one
at a time. We may have many books, but we can only read one book at a time. But
you can multitask by filling different areas of your brain with useful
information on different topics and then giving each area of your brain some
hours, or days, or years, or decades to distill that
information into a well-organized theory of how it all fits together.
When I started research for USDA in 1988,
each Monday morning I would go to the office and ask myself, “What project
should I work on this week?” Then I would spend that whole week on that task.
During my last year at USDA in 2024 I could never spend a whole week or even a
whole day on one task. I needed to read perhaps 100 emails in my In Box, answer
at least 10 of them, and converse with my team of about 10 people. That usually
took about half of each day. I kept a paper checklist with their names at the
top and days of the month down the side. At the end of every day
I put a check mark for those I spoke or wrote to, ensuring that I regularly
interacted one on one with each employee every week.
Thinking in parallel was necessary to
keep up with so many topics at once. But communicating directly with each
member of the team was also important to hear the progress on their project and
to give directions on what research steps might work next. We also had
different team meetings at least 3 days each week to involve more researchers
and help each of us see the big picture. They would ask questions or contribute
further ideas that people working directly on the task had not thought of. We
got so many useful things done in recent years that it is hard to believe that
Republicans in 2025 decided to destroy our research program.
Large datasets
Small datasets can be processed using
personal computers or cell phones, but large datasets often require large, specialized
computers. Staff of the Holstein Association USA in Vermont flew to the
University of Wisconsin twice each year until 1988 carrying magnetic tapes full
of millions of cow records and pedigrees. The biggest 2 computer users on the U
WI campus at that time were researchers in cow genetics and in weather
prediction. “Cloud computing” now lets your data
travel to a large, remote, specialized computer without you having to carry it,
but for very large files, physical transfer may still be faster than internet
transfer.
The value of large datasets is that
they can answer questions more precisely or completely than any small, local
dataset can. When you use a search engine, you hope that if the answer exists
anywhere on the internet, the search will find it and show it to you. But
finding the answer would take a long time if the whole world’s data was not
already indexed in a massive file on some massive computer farm where your search
question goes to find it. Now, the most popular questions get pre-loaded into
your search box when you start typing, and answers to those questions may
already be computed and stored before you ask them.
Data storage, access, and cost
You can find better answers faster if
reports on similar topics are grouped well. Old-fashioned libraries provided that service for us. Before the internet provided
so much information so quickly, we used well organized paper files to do our
research. In my office at USDA, my own paper files totaled about 50 linear feet
of cabinet drawer space. Every time I looked up reports from my files and
cabinets, which was several times per day, my brain
also internally mapped that related information together. My neurons stored
related articles together just like in my folders. By using this compact neural
storage, I could remember which subjects and which years my boss and my
distinguished coworkers had published their reports better than they could. I
am sure that using my well-organized physical files greatly helped organize
those same topics into my neurons.
My goal was to find most reports that
I needed within 5 seconds. The files I used most often were in two 4-drawer
cabinets on either side of my desk. To reach them I only had to swivel my chair
to the right or left and lean forward (1 second). I kept those 8 drawers
labelled and filled by major subject and could pull
out the correct drawer >90% of the time (1 second). I always stored files
together if their topics were most related and I always kept files in the same
order so I could locate the correct file almost by muscle memory (1 second). My
files were usually less than an inch (2.54 cm) thick, and within each file the
most important report(s) that I wrote were in front to let me quickly retrieve
my previous thoughts (1 second). Other reports on that same topic were
carefully organized by subtopic or by date and paperclipped or binder clipped
together within the folder. The subtopic or time group was labelled on a sticky
note sticking up and visible to quickly find the correct paperclipped group (1
second). Finding the correct report within a paperclipped group might take a
few seconds.
Thus, finding my own reports usually
took 3 seconds but finding reports from other authors often took 5 or 10
seconds. My paper file system was competitive with today’s search strategies but
could not keep up with the volume of new electronic articles that arrive today.
My file system had a lasting impact on my brain that individual electronic
searches today might not. Some people who scroll random information all day
long might become scatter brained.
The information you may want to see
may not be free. Often, to read the full story you need to subscribe to the
newspaper or purchase the book. The U.S. government’s policy was often to
provide our answers for free because taxpayers had already paid us. I will not further
comment on how to value the data or information you need because I am not an
expert on that topic.
Sizes of data files can sometimes be
greatly reduced by referential compression and storing only differences from
the reference instead of the complete data or picture. DNA sequence is often
now stored as differences from the reference genome in CRAM files. Your brain
may also do the same thing. For example, you may store what an average human face
looks like and then store each person’s differences from that average face
instead of storing each whole picture in your brain. Once you create a
well-organized map of subjects in your brain, new things that you see may be
stored only as differences from the average of what you saw before.
To reduce clutter in my brain, I
tended to watch only documentaries and not science fiction or dramas because I
wanted to solve only real-world problems and not imaginary problems. Along with
filing visual inputs such as journal articles, I also filed audio inputs on
topics I cared about together onto cassette tapes in the 1980s and 1990s. You
can read about and hear my song play lists in Rock and Roll
Songs by Topic. Those had perhaps a bigger impact on my thinking. When I hear the end
of one song on a sorted cassette, I begin thinking of the next song in the
series because those songs are stored together in my brain. They are linked
together on tape, and after listening to them hundreds of times over the
decades, my brain also stored them as a linked list.
Cause and effect
Data files can help us understand
which events are correlated but not always if one event causes another.
Geneticist Sewall Wright of our USDA laboratory in Beltsville, MD first
published about cause and effect in 1921. For example, if you collect data on cars driving down the road,
the correlation of rain falling and windshield wipers wiping is almost 100%. An
AI program might suggest that you can solve a drought by turning on your
windshield wipers unless given the extra restriction of what is the cause and
what is the effect. Designed experiments can check what causes what but often
have too few records to estimate small effects.
Your brain does not usually access
raw data but only summaries. Those studies may be based on separate, independent
data sets, or be different analyses of the same data, or include some
overlapping data. Most studies report how much data they include and publish
standard errors to help quantify accuracy, but to accurately combine different
summaries of data you also should know or estimate their error correlations.
Truly independent sources of information may have no error correlations such
that the more studies you read, the more you know and the more accurate your
combined estimate becomes. But hearing 10 people tell you the same estimate
does not give you more information if they all are repeating the same small
summary that they each read. Computers can estimate error correlations and
accuracies of combined summaries in a process called meta-analysis using
methods like those I published in 2010 to combine genomic predictions across countries that shared some data. Computers must also combine data
summaries if raw data sets are too large or not available.
Probability and Bayes theorem
Your
daily life and your long-term future require estimating probabilities
of what might happen that day or in future years. You may want to carry an
umbrella or a jacket when probabilities of rain or a cold front are above a certain minimum. Those may already be computed
and easily displayed on your phone or television or computer. Other
probabilities specific to you are harder to compute, such as if you take a
certain action, how often will the result be a success?
Statisticians often use Bayes theorem
to compute probabilities for future events. Your initial guess might be a 50%
probability of making a sports team if the team needs 10 players, but 20
players try out. Your probability may go up or down if your own performance in
the tryout is better or poorer than others. Your brain does those calculations all
the time to help decide if you should focus on one project or another. Each day
you should combine your prior beliefs with the new data to obtain posterior
beliefs that are more accurate than your initial guess without the data.
Instead of trying to remember and
recompute all the past data you ever observed, your brain may use a simple
statistical trick. Each night as you sleep, your brain may combine your prior
beliefs from yesterday with the new data from today to compute new prior beliefs
to use tomorrow. Then it just needs to replace your past, less accurate beliefs
with your new, more accurate beliefs. You might wake up tomorrow thinking, yes,
I can make that team, or no, that sport is not for me. Your neurons may never
really use Bayes’ theorem, but by updating the neural
connections that quantify your beliefs, you may live a Bayesian life without ever
thinking about the formal math of Reverend Thomas Bayes (1763). Most of your
probability calculations may even happen while you sleep.
Interacting with AI
Current AI tools can interpret a
spoken question and give the most likely answer after consulting a huge
database of relevant information. The user can try hard to ask the right
questions about the most important problems or can use AI more for entertainment.
The programs I wrote were almost always to predict the future rather than to summarize
the past. That may be harder for AI to do. We used data from past generations
of cows to predict performance for each calf in the next generation. That
requires massive, specialized data files that AI does not have access to. We
understood exactly what data our predictions were predicted from. Computers can
calculate predictions, but currently AI may do a great job only of summarizing
public data from the past. For programmers, AI can now help
write the computer code that you can then use for repeated tasks.
Background and
advice
For 40 continuous years I thought full time about how to
compute. Computational
strategies for estimation of variance components was the title of my 1986
PhD thesis at Iowa State. That topic is still important
and we still use my computer code from 1986 in research today. Testing my
programs helped IBM improve memory management in 1987 for the computers they
designed. In 1989, our project at USDA won 3rd place in IBM’s Supercomputing
Competition, Life and Health Sciences Division, “recognizing individuals whose
imagination has elevated the state of the art in analysis and modeling.” My
most important paper Efficient
methods to compute genomic predictions in 2008 was also about how to
compute.
I started thinking more about how to think in 2007 when I
started coding for parallel processing. Your neurons do not think in one straight
line to solve a problem like computer programs did decades ago when reading
data from a long, magnetic tape. Thinking in parallel and computing in parallel
seemed to go together because neurons in many areas of your brain can cooperate
to find an answer, like many processors cooperating by each doing some of the
work so that even a big job can get done quickly. Writing this report took less
than a week because I must have already thought a lot about how thinking
compares to computing.
Parallel computing and parallel thinking are both harder to
do than simple math problems where you take x plus y, divide by z, and square
the result, or tasks like that. But parallel thinking is not as hard to do as
designing an AI program to be as intelligent as you are.
Computers can process much more data much faster than you
can, but they do not have the instinct and experience that your brain has for
deciding what you should do next. If you are employed, getting your work done and
getting your pay are usually the best use of your time. If you are a student,
none of this material will be on your next test, so you should instead study
the material that will be. If you are retired like me, then you can let your
mind wander to new topics, like I just did. Computers still do not know how to
let their minds wander. That is something humans still need to do.
You got all these sentences for free. I will not profit if you use or do not use my ideas. Some people may
believe that free advice must be worthless, otherwise I would charge you for
it. Other people believe that we should freely share ideas that make sense. For
37 years, US taxpayers paid me to think of and give away all my best ideas. You
may believe that this is an old, bad habit that I should give up and instead
start selling my advice if it has any value. I believe that thinking is a skill
that we each can improve by practicing it.
Reviewer comment
An ‘anonymous’ reviewer sent me this
comment: “Thank you for sharing great information and examples from your real
experiences. I need some sleep now so my neurons can make new connections with
all this information. Please continue sharing your knowledge so generously, as
there are people like me who will greatly benefit from it.”
[10:30pm, August 24, 2025]
References
How
the Brain Stores and Retrieves Memories
How
Memories Are Formed and Where They're Stored | Psychology Today
Human multitasking -
Wikipedia
How are Microchips Made? CPU Manufacturing
Process Steps
Computation
and Computational Thinking
Computational
Thinking Is More about Thinking than Computing
Hidden Figures [2 hour movie]
The Imitation Game [1 hour 47 minute movie]
Trends in the
cost of computing – AI Impacts
Return
to: Solutions To Personal Problems