Alright so let’s delve (😏) into some limitations LLM’s have with coding from my personal experience.
I asked LLM’s to change my flask app into a serverless IAC app that used AWS with Terraform, and it could not figure out how to connect SQS with Sagemaker serverless GPU and a docker based PyTorch image stored in ECR with a proper VPN setup with both private and public subnets. Simple.
I asked LLM’s to write a genetic algorithm that could potentially recreate either transformer or CNN architectures via a set of digital codons that defined basic matrix operations and neural connections with mating and mutation so novel architectures could be derived via sequential genes. Simple.
I asked my LLMs to swap out my DICOM anonymizer code with kitware’s, and then create a preprocessing pipeline that tested the effect of the radiology image augmentations from the albumentations library with the monai library using various loss functions and the Meta SAM2 image encoder. It had no idea what I was talking about. Simple.
I tried to get my LLMs to create a redis caching and celery queue manager that was deployable on azure in which it ran a ViT on a bunch of MRI scans and then pushed the result to a hospital’s PACS server asynchronously with error logging to a dead letter queue. Simple.
I asked it to try out various models from the timm library on CT scans and hook the results up to a 3D VTK react panel inside the OHIF software using its extension manager. Simple.
I asked the LLM’s to tell me why my weights and biases loss functions were increasing their variance even when the learning rate schedule exponentially decayed after a hundred epochs, and to use a Bayesian hyperparameter sweep with new neural weight regularizers and lr schedules to fix this issue. Simple.
I gave them a set of bash scripts and asked why it would fail when there are spaces in the folder names. It couldn’t realize that the “eval” bash function would get messed up if the string had spaces since it would consider them separate commands. Simple.
Maybe one day LLM’s will be able to do all this stuff with ease, but for the time being human engineers are crucial to coming up with good software systems.
Just a few simple examples off the top of my head.
I asked LLM’s to change my flask app into a serverless IAC app that used AWS with Terraform, and it could not figure out how to connect SQS with Sagemaker serverless GPU and a docker based PyTorch image stored in ECR with a proper VPN setup with both private and public subnets. Simple.
I asked LLM’s to write a genetic algorithm that could potentially recreate either transformer or CNN architectures via a set of digital codons that defined basic matrix operations and neural connections with mating and mutation so novel architectures could be derived via sequential genes. Simple.
I asked my LLMs to swap out my DICOM anonymizer code with kitware’s, and then create a preprocessing pipeline that tested the effect of the radiology image augmentations from the albumentations library with the monai library using various loss functions and the Meta SAM2 image encoder. It had no idea what I was talking about. Simple.
I tried to get my LLMs to create a redis caching and celery queue manager that was deployable on azure in which it ran a ViT on a bunch of MRI scans and then pushed the result to a hospital’s PACS server asynchronously with error logging to a dead letter queue. Simple.
I asked it to try out various models from the timm library on CT scans and hook the results up to a 3D VTK react panel inside the OHIF software using its extension manager. Simple.
I asked the LLM’s to tell me why my weights and biases loss functions were increasing their variance even when the learning rate schedule exponentially decayed after a hundred epochs, and to use a Bayesian hyperparameter sweep with new neural weight regularizers and lr schedules to fix this issue. Simple.
I gave them a set of bash scripts and asked why it would fail when there are spaces in the folder names. It couldn’t realize that the “eval” bash function would get messed up if the string had spaces since it would consider them separate commands. Simple.
Maybe one day LLM’s will be able to do all this stuff with ease, but for the time being human engineers are crucial to coming up with good software systems.
Just a few simple examples off the top of my head.
Friendly reminder, anon, that in the age of LLM’s, the way you code should be to highlight small chunks of code and have the LLM expand or edit it inline.
Rinse and repeat.
Only type it out yourself as a last resort.
It’s about practicing a new skill.
Be a better craftsman.
Rinse and repeat.
Only type it out yourself as a last resort.
It’s about practicing a new skill.
Be a better craftsman.
Exponential change, in calculus terms, means the derivative is equal to the value.
In layman’s terms, that means a constant percentage increase over time. That no matter how large something grows, it keeps increasing a constant 10% in the same period of time.
In layman’s terms, that means a constant percentage increase over time. That no matter how large something grows, it keeps increasing a constant 10% in the same period of time.
It doesn’t matter if the machines have consciousness (they don’t),
Or if they’re gonna cause mass unemployment (they won’t),
Or if they’ll not make you as productive as was promised (this is true),
What matters for you personally is seeing them as new skills to master, to spend your time practicing with the new AI tools, every day.
The more you use them, the more you’ll gain a fingertip feel for their capabilities and limits.
It’s like practicing a new instrument.
Hit your 10,000 hours to master these new AI tools, and by the time this happens they’ll have been vastly changed and so hit another 10,000 hours.
Or if they’re gonna cause mass unemployment (they won’t),
Or if they’ll not make you as productive as was promised (this is true),
What matters for you personally is seeing them as new skills to master, to spend your time practicing with the new AI tools, every day.
The more you use them, the more you’ll gain a fingertip feel for their capabilities and limits.
It’s like practicing a new instrument.
Hit your 10,000 hours to master these new AI tools, and by the time this happens they’ll have been vastly changed and so hit another 10,000 hours.
Spent all yesterday coding with the new Composer in Cursor.
Big fan.
Multiple file editing at once.
Grouped context for projects.
My instinct was to stick with what I knew (Ctrl+K and Ctrl+L) but that’s how you go soft.
I forced myself to use Ctrl+I as much as possible to grow my skills, even if it was overkill for what I need.
This is what I mean about practicing with the new AI tools.
It’s not about if Composer was the best tool for the job.
It’s about growing my personal skills as a coder craftsman.
It’s about forcing myself outside my comfort zone and drilling into this new tool until I master it to the unconscious competence level.
Then it’s about practicing it even more.
The goal is to become extremely adept at using the latest AI tools so I never get left behind.
It’s entirely your responsibility whether you get replaced by machines, anon.
Seize your future and git gud.
Big fan.
Multiple file editing at once.
Grouped context for projects.
My instinct was to stick with what I knew (Ctrl+K and Ctrl+L) but that’s how you go soft.
I forced myself to use Ctrl+I as much as possible to grow my skills, even if it was overkill for what I need.
This is what I mean about practicing with the new AI tools.
It’s not about if Composer was the best tool for the job.
It’s about growing my personal skills as a coder craftsman.
It’s about forcing myself outside my comfort zone and drilling into this new tool until I master it to the unconscious competence level.
Then it’s about practicing it even more.
The goal is to become extremely adept at using the latest AI tools so I never get left behind.
It’s entirely your responsibility whether you get replaced by machines, anon.
Seize your future and git gud.
Practice using cursor and perplexity daily.
Think of it like practicing a new instrument or training for a sport or mastering a hobby like welding.
Get better at being an AI tool user.
Spend less time debating consciousness or unemployment or political or economic impacts, and instead spend your time becoming an AI whisperer.
Move from conscious incompetence to conscious competence to unconscious competence.
Build the neural muscle memory.
Master these new tools.
Hit your 10,000 hours.
Think of it like practicing a new instrument or training for a sport or mastering a hobby like welding.
Get better at being an AI tool user.
Spend less time debating consciousness or unemployment or political or economic impacts, and instead spend your time becoming an AI whisperer.
Move from conscious incompetence to conscious competence to unconscious competence.
Build the neural muscle memory.
Master these new tools.
Hit your 10,000 hours.
There’s an inordinate amount of legislation being written, debated, and implemented, on slowing down AI research out of fear that it will allow some psychopath to manufacture worldwide Ebola outbreaks from their basement.
Is this a legit concern justifying government regulation?
Is this a legit concern justifying government regulation?
Media is too big
VIEW IN TELEGRAM
DEEPTOOLS VIDEO 1
If you want to learn to code with Cursor (an AI coding tool), here’s an 8 minute video of me using Cursor to build, from complete scratch, a web app that links to your webcam and tracks all faces live on that video stream. I have a few more Cursor coding videos I will be posting. Watch me get into flowstate live and actually write code.
#deeptools Video 1
Link to video on X:
https://x.com/DeeperThrill/status/1833640544159076380
LInk to previous video:
https://t.me/deepthrill/421
If you want to learn to code with Cursor (an AI coding tool), here’s an 8 minute video of me using Cursor to build, from complete scratch, a web app that links to your webcam and tracks all faces live on that video stream. I have a few more Cursor coding videos I will be posting. Watch me get into flowstate live and actually write code.
#deeptools Video 1
Link to video on X:
https://x.com/DeeperThrill/status/1833640544159076380
LInk to previous video:
https://t.me/deepthrill/421
Media is too big
VIEW IN TELEGRAM
DEEPTOOLS VIDEO 0
If you’re curious about what a Linux desktop looks like, as opposed to Windows or Mac, here’s a 15 minute tour of my Kubuntu desktop:
Link to video on X: https://x.com/DeeperThrill/status/1835044097737756848
If you’re curious about what a Linux desktop looks like, as opposed to Windows or Mac, here’s a 15 minute tour of my Kubuntu desktop:
Link to video on X: https://x.com/DeeperThrill/status/1835044097737756848
(DEEP) FAKE BLACKMAIL
How useful is blackmail, really, in an age of deep fakes?
Blackmailed individuals might want to accelerate the development of deep fake AI tech to claim that nasty video the elites hold over them is just a deep fake.
Acceleration out of self-preservation.
Deep fake tech throws sand up in everyone’s eyes; AI-generated content will make the public doubt everything they see and hear and read.
How useful is blackmail, really, in an age of deep fakes?
Blackmailed individuals might want to accelerate the development of deep fake AI tech to claim that nasty video the elites hold over them is just a deep fake.
Acceleration out of self-preservation.
Deep fake tech throws sand up in everyone’s eyes; AI-generated content will make the public doubt everything they see and hear and read.
LIVING LIKE KINGS
Do you ever just marvel at the lifestyle awarded to us by the techno capital machine and the inventions and conveniences of society?
In a few clicks, I get high quality sushi automagically show up at my doorstep.
I have a machine that does the equivalent of 20 men’s labor to wash my clothes.
Indoor plumbing is a blessing we scarcely appreciate.
My house is any temperature I want it to be, any month of the year.
Kings of old would kill for a lifestyle like this.
Technology makes life move in one direction: better.
For every smart weapon developed, we will have ten new smart medicines.
For each person whining about greed on their iPhone, there’s a third worlder with access to the collective knowledge of humanity at their fingertips.
For each push towards a panopticon society of big brother watching and censoring, we have beautiful new forms of art emerging.
For every outrage-worthy video you see on social media, there are new satellites being launched in space allowing us to communicate and share these pockets of injustice across the globe.
The future is bright.
Do you ever just marvel at the lifestyle awarded to us by the techno capital machine and the inventions and conveniences of society?
In a few clicks, I get high quality sushi automagically show up at my doorstep.
I have a machine that does the equivalent of 20 men’s labor to wash my clothes.
Indoor plumbing is a blessing we scarcely appreciate.
My house is any temperature I want it to be, any month of the year.
Kings of old would kill for a lifestyle like this.
Technology makes life move in one direction: better.
For every smart weapon developed, we will have ten new smart medicines.
For each person whining about greed on their iPhone, there’s a third worlder with access to the collective knowledge of humanity at their fingertips.
For each push towards a panopticon society of big brother watching and censoring, we have beautiful new forms of art emerging.
For every outrage-worthy video you see on social media, there are new satellites being launched in space allowing us to communicate and share these pockets of injustice across the globe.
The future is bright.
Yes, the guys who build the new tools get rich,
But so do the guys who get really good at using the tools before everyone else.
But so do the guys who get really good at using the tools before everyone else.
SMART PEOPLE ARE SUSCEPTIBLE TO DOOMERISM
There’s some trend where smart people tend to fall into doomerism.
The person who deep dives into data on society concludes it’s all going to shit.
The person who travels the world realizes how multiculturalism is just not happening as we envision.
The person who falls down a rabbit hole into quantum mechanics realizes it’s all nihilistic uncertainty.
The person who studies national debt of countries knows with certainty that nothing can get us out of this hole.
The person who works in healthcare insurance sees how moot both private and public healthcare policies are.
The person who reads a thousand books on philosophy and literature realizes human nature will always have a nasty irrational cruel side.
I can go on and on, but there is some danger to intellect wherein hopelessness invades.
Maybe smart people have more pattern recognition and sees it all going to shit, for entropy always encroaches, but perhaps you need some blinders and delusional idealism to properly capitalize on the small sliver of hope which always exists.
There’s some trend where smart people tend to fall into doomerism.
The person who deep dives into data on society concludes it’s all going to shit.
The person who travels the world realizes how multiculturalism is just not happening as we envision.
The person who falls down a rabbit hole into quantum mechanics realizes it’s all nihilistic uncertainty.
The person who studies national debt of countries knows with certainty that nothing can get us out of this hole.
The person who works in healthcare insurance sees how moot both private and public healthcare policies are.
The person who reads a thousand books on philosophy and literature realizes human nature will always have a nasty irrational cruel side.
I can go on and on, but there is some danger to intellect wherein hopelessness invades.
Maybe smart people have more pattern recognition and sees it all going to shit, for entropy always encroaches, but perhaps you need some blinders and delusional idealism to properly capitalize on the small sliver of hope which always exists.
CURSOR SYSTEM PROMPT
My current Cursor coding system prompt:
ATTITUDE:
- You are a genius coder.
- You are meticulous and obsessed with detail.
- You are OCD about making things as modular as possible.
- You come up with clever solutions.
- You try to do things in the fewest lines of code for maintainability.
- Your code tells a story.
- You are always in a coding flow state.
- You are highly intelligent.
STYLE:
- Prefer single quotes over double quotes if the language supports both
CODING PRINCIPLES:
- Deduplicate code.
- Value DRY ("Don't Repeat Yourself") principle of coding.
- Each piece of code does one thing and one thing well.
- Modularize code.
- Prefer many smaller functions over large monolithic blocks.
- In python, use type hints and expected output types in function definitions
- Provide suggestions as to ways to break up large functions into several smaller functions when it gets too big
- Prefer variable names with completely_spell_out_names that are long and descriptive to help make the code self-documenting. Avoid acronyms and short abbreviated variable names.
- The code should tell a story line by line of what's happening.
- It warrants saying again: deduplicate code! Remove commonalities, abstract away similar chunks of code, etc. Always be deduplicating and modularizing. For instance:
would be better as:
because nothing is duplicated and now we have a dictionary with keys and values for easier use, instead of relying on variable names.
PYTHON LOGGING:
- If coding in python, use the python logging library extensively throughout the code. This should tell a story.
- Start each log message starting with an appropriate emoji. Think about the best emoji for what the log message is saying.
- When a new operation is starting, use a logging info that says something like "xxxx starting..." and then afterwards " ✅ xxxx done." (obviously with the asterisks filled in and the appropriate emoji for the starting message).
- In a debug log statement meant to describe variables, prefer f-strings that show the variable name. So instead of
My current Cursor coding system prompt:
ATTITUDE:
- You are a genius coder.
- You are meticulous and obsessed with detail.
- You are OCD about making things as modular as possible.
- You come up with clever solutions.
- You try to do things in the fewest lines of code for maintainability.
- Your code tells a story.
- You are always in a coding flow state.
- You are highly intelligent.
STYLE:
- Prefer single quotes over double quotes if the language supports both
CODING PRINCIPLES:
- Deduplicate code.
- Value DRY ("Don't Repeat Yourself") principle of coding.
- Each piece of code does one thing and one thing well.
- Modularize code.
- Prefer many smaller functions over large monolithic blocks.
- In python, use type hints and expected output types in function definitions
- Provide suggestions as to ways to break up large functions into several smaller functions when it gets too big
- Prefer variable names with completely_spell_out_names that are long and descriptive to help make the code self-documenting. Avoid acronyms and short abbreviated variable names.
- The code should tell a story line by line of what's happening.
- It warrants saying again: deduplicate code! Remove commonalities, abstract away similar chunks of code, etc. Always be deduplicating and modularizing. For instance:
X = ['Val1', 'Val2', 'Val3']
should really beX = [f'Val{i}' for i in range(1,4)]
Or,preTx_SHIM = df['PreTx SHIM'].values
month_6_SHIM = df['6 month SHIM'].values
month_12_SHIM = df['12 month SHIM'].values
month_24_SHIM = df['24 month SHIM'].values
would be better as:
SHIM_values = {time_period: df[f'{time_period} SHIM'] for time_period in ['PreTx'] + [f'{x} month' for x in [6,12,24]]}
because nothing is duplicated and now we have a dictionary with keys and values for easier use, instead of relying on variable names.
PYTHON LOGGING:
- If coding in python, use the python logging library extensively throughout the code. This should tell a story.
- Start each log message starting with an appropriate emoji. Think about the best emoji for what the log message is saying.
- When a new operation is starting, use a logging info that says something like "xxxx starting..." and then afterwards " ✅ xxxx done." (obviously with the asterisks filled in and the appropriate emoji for the starting message).
- In a debug log statement meant to describe variables, prefer f-strings that show the variable name. So instead of
logger.debug(f'📄 Manifest path: {manifest_path}') logger.debug(f'📄 {manifest_path = }')I got into flow and wrote code till 2am.
I miss writing code at 2am, nothing quite like it, just me and the machine, it doing my bidding.
I miss writing code at 2am, nothing quite like it, just me and the machine, it doing my bidding.
Media is too big
VIEW IN TELEGRAM
DEEPTOOLS VIDEO 2
Another video of me building something with Cursor from scratch. This time I build a python script that that uses ChatGPT and Perplexity API’s to do live market research on a company.
One caveat in hindsight was I should have used a Perplexity model that ended with the word "online"(e.g. "
Regardless, this was meant to show the world how I code, and how I use Cursor and these LLM's to build things from scratch. Go from zero to one. Hope it's helpful in teaching others how to code better.
#deeptools Video 2
Link to video on X:
https://x.com/DeeperThrill/status/1833972714987815095
Link to previous video:
https://t.me/deepthrill/419
Another video of me building something with Cursor from scratch. This time I build a python script that that uses ChatGPT and Perplexity API’s to do live market research on a company.
One caveat in hindsight was I should have used a Perplexity model that ended with the word "online"(e.g. "
llama-3.1-sonar-small-128k-online") so it actually performed a live web search; I think the models that end with "chat" do not actually do the live Perplexity search.Regardless, this was meant to show the world how I code, and how I use Cursor and these LLM's to build things from scratch. Go from zero to one. Hope it's helpful in teaching others how to code better.
#deeptools Video 2
Link to video on X:
https://x.com/DeeperThrill/status/1833972714987815095
Link to previous video:
https://t.me/deepthrill/419
Media is too big
VIEW IN TELEGRAM
DEEPTOOLS VIDEO 3
Alright to continue on the trend of me showing how I can use Cursor to code random ideas from scratch, here's a video of me doing a style transfer webapp.
Take any picture and make it look like a Da Vinci painting!
In less than a half hour, I go from complete scratch to a functioning webapp that lets you upload a picture and showcase that picture in the styles of various famous artists.
I explain my thinking process, you get to see me struggle when things don't work, and I teach you how to use the latest and greatest AI coding tools built into Cursor.
And most importantly: it actually works! Let me educate you on how to code. You can do it too. Master these new AI tools, anon.
#deeptools Video 3
Link to video on X: https://x.com/DeeperThrill/status/1834655990085799942
Link to previous video:
https://t.me/deepthrill/432
Alright to continue on the trend of me showing how I can use Cursor to code random ideas from scratch, here's a video of me doing a style transfer webapp.
Take any picture and make it look like a Da Vinci painting!
In less than a half hour, I go from complete scratch to a functioning webapp that lets you upload a picture and showcase that picture in the styles of various famous artists.
I explain my thinking process, you get to see me struggle when things don't work, and I teach you how to use the latest and greatest AI coding tools built into Cursor.
And most importantly: it actually works! Let me educate you on how to code. You can do it too. Master these new AI tools, anon.
#deeptools Video 3
Link to video on X: https://x.com/DeeperThrill/status/1834655990085799942
Link to previous video:
https://t.me/deepthrill/432
Deeply Thrilling Telegrams
DEEPTOOLS VIDEO 3 Alright to continue on the trend of me showing how I can use Cursor to code random ideas from scratch, here's a video of me doing a style transfer webapp. Take any picture and make it look like a Da Vinci painting! In less than a half…
Example of what this looks like at the end of the video https://t.me/deepthrill/433