machine learning, predictions, & where AI is headed

Some weeks ago, there was an event that occurred while I was in Long Island, NY.

i was at my friend's aunt's pool house while i was, admittedly, stoned, and at some point she asked me what i was studying.

computer science.

then she asked if i was worried about my major.

i asked why, because of AI? (a logical assumption to make)

her response: yeah.

she is a Brown University alumna & a nurse now, which somehow made the question stick with me more than it probably should have.

because it is a strange thing to be eighteen years old, just starting a computer science degree, and already have people asking whether the thing you are studying is going to survive the technology being invented while you are studying it.

and the annoying part is that it is not really a stupid question.

i just do not think the answer is as simple as

AI is going to replace programmers.

or

AI is just another tool and nothing is going to change.

both of those feel too easy.

Sam Altman wrote this blog post in 2015 called Machine Intelligence, Part 1.

which is important because reading predictions about AI from 2015 now feels a little like finding an old note someone wrote before they knew how strange the next decade was going to get.

one of the things he talks about is how difficult it is to know how far away machine intelligence actually is because progress might look slow for a long time and then suddenly go vertical.

he also points out this weird habit humans have where a computer finally becomes good at something we used to consider evidence of intelligence, and then we decide that thing was never really intelligence anyway.

we see examples of this with

Chess, with the development of Stockfish-like engines.

Jeopardy.

self-driving cars.

apparently, once the computer can do it, it stops counting.

but the part of that post i keep thinking about is the possibility that things like creativity and intelligence might not require some impossible magic ingredient at all.

maybe they emerge from a relatively small number of algorithms with enough compute behind them.

that was written in 2015.

now we have machines that can write code, generate images, explain calculus, imitate voices, translate languages, analyze proteins, and occasionally explain something completely made up with the confidence of a tenured professor.

so i went back to thinking about that question from the pool house.

Should I actually be worried about studying Computer Science?

the answer is, probably.

just not for the reason people usually mean when they ask me.

there was this guy named Alan Turing, seventy years and something ago.

(not that this is about Alan Turing).

he just feels like the kind of person you have to mention before talking about intelligence for more than five minutes.

i have been thinking about machine learning. i keep coming to the feeling that we are probably looking at a very stupid version of whatever this is going to become.

that is a thought to have when the "stupid" version can already write code, explain calculus, translate languages, make images, summarize a textbook, mimic voices, analyze proteins, and confidently tell you something that is completely false.

it is balanced technology.

there is this pressure to talk about AI as if it is just a product.

ChatGPT Sol.

Claude Fable.

Gemini 3.1 Pro.

Kimi K3.

whatever model some guy on Twitter decided reached AGI this week because it got an increased 4% accuracy on some test.

the interesting part to me is probably not the apps on top.

the interesting part is the idea underneath them.

you take an amount of data.

you throw it at an amount of compute.

you build a math machine with a number of parameters.

then you ask it to predict the next thing.

it turns out that this is an insulting set of instructions to create something that can explain the French Revolution, debug C++, write a Shakespeare parody, or recognize a dog and tell you why your SQL query is wrong.

that is absurd.

there is a difference between a computer that follows instructions and a computer that learns the instructions.

traditional programming is basically:

if this happens, do that.

if that happens, do this.

machine learning is closer to:

here are ten billion examples, figure it out.

that sounds like the professor ever.

it apparently works.

i think that is where people are misunderstanding what is happening with AI.

it does not have to become a chrome robot in a chair before it changes everything.

it does not have to have glowing eyes or consciousness or wake up one day and decide humanity is wasting its time.

that is the movie version.

the boring version is probably more interesting.

the boring version is that AI slowly becomes part of everything.

your operating system, your browser, your IDE, your search engine, your email, your camera, your doctors software, your bank, your car, your school, your job.

eventually, you stop launching an "AI app" because it feels weird to call Safari an electricity app.

the intelligence is part of the software.

that is where i think we are headed.

it might not be one artificial brain in some server answering every question humanity wants to ask.

though that might happen too, who knows.

maybe there are millions of little agents, models, tools, and weird pockets of intelligence doing things for us in the background.

one model plans, another writes the code, another checks the code, another runs the code. another realizes the first three are idiots and rewrites everything. it is basically a group project, except none of them can complain about the others not helping.

maybe.

the agent thing is particularly interesting to me because there is a difference between something that can tell you how to do something and something that can actually do it.

"here is how you can build this app"

is great.

but

"i built the app"

is something entirely.

and that transition feels important.

for most of computing history, humans have been the bottleneck.

you open the browser.

you copy the information.

you paste it else.

you make the spreadsheet.

you send the email.

you run the command.

you check whether it worked.

you fix whatever broke.

basically, AI agents are asking:

why are you doing all of this?

which is slightly rude, because i want to remain useful.

the obvious question is what happens to work.

it seems like everyone has one of two answers.

AI will destroy every job by Tuesday.

or

AI is basically autocomplete, and nothing important is happening.

both of those seem of silly to me.

technology usually has a relationship with jobs. it does not destroy them all at once. politely delete them one by one.

spreadsheets did not kill finance.

Photoshop did not kill photography.

the internet did not kill stores.

they did all change what it meant to be good at those jobs.

i think AI is probably doing the thing, just maybe faster and across more industries at once.

the programmer who refuses to use AI might look like the programmer who refuses to use Stack Overflow because "real programmers remember everything.". cool, enjoy that.

there is another possibility that i think is much more interesting. programming might become less about writing everything and more about being super precise about what you want built.

architecture, taste, constraints, testing, knowing when the machine is wrong, knowing what question to ask, and knowing what should exist in the place.

because making something is getting cheaper.

deciding what is worth making might become more valuable.

that applies to writing too.

and art.

and research.

probably half the things people currently think require a special human spark.

that gets uncomfortable quickly.

if intelligence becomes cheap, it stops being the thing.

maybe judgment becomes scarce.

maybe originality becomes scarce.

maybe good taste becomes scarce.

maybe the ability to know whether something is true becomes incredibly important when there are machines there that can generate ten thousand extremely convincing lies before breakfast.

that part worries me more than Terminator.

not killer robots,. bullshit at scale.

an AI does not have to hate you to be dangerous.

it just has to be wrong in a way that looks right.

that is less cinematic, but probably more realistic.

there is also the question of whether scaling keeps working.

more data, compute, bigger models, better models.

for a while, the industry has basically been running the worlds expensive experiment called:

what if we just make it larger?

and annoyingly, it keeps working.

there has to be some point where throwing another warehouse of GPUs at the problem stops being helpful.

probably.

then again, people have been saying scaling will stop working for years, and the models keep getting weirder.

the part i find most interesting is that nobody really knows what the final interface is.

we assume it is chat because chat works now.

you type something, the model types something. it is futuristic, basically AOL Instant Messenger, except the other person has read Wikipedia.

but maybe talking to an AI through a chat interface is going to feel as old as talking to a webpage through a command line.

maybe the system watches what you are doing, understands the project, remembers what happened yesterday, knows what files matter, knows what you are trying to do, and does half the work before you ask for it.

that sounds convenient and also slightly scary.

there is a strange line between:

this computer understands me.

why does this computer understand me?

and memory is part of that.

now, most software has almost no memory of you.

your calculator does not remember what you calculated yesterday.

your IDE does not really understand why you are doing the project.

your browser knows a lot about what you click, but not about what you want to become.

an intelligent system with real memory changes that.

imagine opening your laptop. it remembers the proof you were trying to write three months ago, or the app you were building, or that one paper you were reading at 3:14am, or the bug you never fixed, and it says:

i think i figured it out.

that sounds more interesting to me than asking a chatbot to write a cover letter.

the models themselves might get smaller.

which is a thing to say because everyone seems to want everything to get bigger.

400 billion parameters, one trillion, five trillion. at some point, the numbers start to feel like debt figures.

if smaller models get good enough to run on your device, then suddenly your laptop or phone can have something intelligent running without sending everything to some mysterious datacenter.

this is important for privacy, speed, cost, offline use, and personalization.

the future might not be one model.

it might be a model teaching a smaller model that teaches an even smaller model that lives on your MacBook and knows you hate when it rewrites your code in ways you never asked for.

beautiful.

there is also robotics, which's where all of this stops being trapped inside rectangles.

language models are weird because they can know an amount about the physical world while having almost no understanding of what it is like to actually be in it.

robots solve that problem. at least, they try to.

once machine learning gets good enough at vision, language, planning, movement, and correcting itself in the world, things get much more interesting.

probably much more broken for a while.

watching a robot fail to fold laundry is funny. watching one successfully fold laundry every day for ten years is an industry.

that is the part people do not realize. the boring repetition after the breakthrough is where society actually changes.

the first car is cool. one hundred million cars is civilization.

the first computer is neat. five billion smartphones is civilization.

one AI writing code is interesting. hundreds of millions of AIs constantly doing some form of mental work is something else.

i do not know what you call that. probably still AI, because naming things is hard.

maybe none of this leads to AGI. maybe AGI itself turns out to be one of those words that everyone fights about until it has no meaning left.

one system does enough, on some task, that someone says:

that is AGI.

and someone else says:

no, it is not.

they fight until the end of time.

i almost do not care.

if a machine can do 80% of the tasks that humans do, arguing about whether that machine is "truly intelligent" feels like arguing about whether a submarine truly swims.

it is a philosophical question, but the submarine is still moving through the water.

that is my thought on where AI is going. i think AI will be less about a show and more about fitting into our lives.

it will be less about

look what this chatbot can do

and more about

why's this suddenly built into everything

machine learning starts as a simple feature. then machine learning becomes a tool. then machine learning becomes an assistant, then machine learning becomes an agent. eventually, you stop even giving machine learning a name.

it will just be the computer. it is like how nobody gets excited that their phone can connect to the internet anymore.

course, the phone can do that.

one day, asking if a piece of software "has AI" will feel just as silly. then the real question will be what humans actually do when useful intelligence is not a thing anymore.

i do not know. maybe we will become more creative. maybe we will become very lazy.. maybe we will do both.

someone will use AI to find a drug, and someone else will use that same AI to make SpongeBob sing a Kanye song in Mandarin.

humans are complicated like that.

i think we make a mistake when we assume the future of AI is just a better version of the chatbot we use today.

the first websites did not look anything like the internet we use now.

the first phones did not look anything like smartphones.

the first computers were huge. mostly just did math.

technology has a habit of starting out looking much less impressive than what it turns into.

so maybe the strange part about machine learning is not that machine learning can already do many cool things.

or maybe we are just looking at the computer before the mouse, or the internet before Google, or the smartphone before the App Store. maybe we are trying to imagine the thirty years just by looking at a textbox that says

message ChatGPT

you could argue that i am just being dramatic & thinking too much again, but give this a thought

and also

shoutout to whoever invented gradient descent.