Santiago
@svpino.com
I help companies build Machine Learning • I run http://ml.school. • Posts about what I learn along the way.
Gemini's documentation says that "responses for a given prompt are mostly deterministic, but a small amount of variation is still possible."
I'm not sure about you, but "mostly deterministic" is not the same as "deterministic."
9/9
I'm not sure about you, but "mostly deterministic" is not the same as "deterministic."
9/9
September 30, 2025 at 1:04 PM
Gemini's documentation says that "responses for a given prompt are mostly deterministic, but a small amount of variation is still possible."
I'm not sure about you, but "mostly deterministic" is not the same as "deterministic."
9/9
I'm not sure about you, but "mostly deterministic" is not the same as "deterministic."
9/9
So, in practice, the LLM you are calling is non-deterministic, even at temperature = 0.
OpenAI's documentation tells us that we can expect "(mostly) deterministic outputs across API calls."
8/9
OpenAI's documentation tells us that we can expect "(mostly) deterministic outputs across API calls."
8/9
September 30, 2025 at 1:04 PM
So, in practice, the LLM you are calling is non-deterministic, even at temperature = 0.
OpenAI's documentation tells us that we can expect "(mostly) deterministic outputs across API calls."
8/9
OpenAI's documentation tells us that we can expect "(mostly) deterministic outputs across API calls."
8/9
This guy freaked out when he tried a model much dumber than GPT 3.5. He ended up losing his job.
He was an engineer and thought the model was sentient.
Imagine what will happen when 99.99% of the world population uses one of these models for the first time.
He was an engineer and thought the model was sentient.
Imagine what will happen when 99.99% of the world population uses one of these models for the first time.
January 6, 2025 at 6:23 PM
This guy freaked out when he tried a model much dumber than GPT 3.5. He ended up losing his job.
He was an engineer and thought the model was sentient.
Imagine what will happen when 99.99% of the world population uses one of these models for the first time.
He was an engineer and thought the model was sentient.
Imagine what will happen when 99.99% of the world population uses one of these models for the first time.
This is a 10x move here. One of my favorite ways to leverage models.
December 29, 2024 at 4:14 PM
This is a 10x move here. One of my favorite ways to leverage models.
Midwits with a chatbot will never build better software than you.
These are two of my all-time favorite Software Engineering books:
1. The Pragmatic Programmer
2. Clean Code
Read them in 2025. They will pay off for years to come.
These are two of my all-time favorite Software Engineering books:
1. The Pragmatic Programmer
2. Clean Code
Read them in 2025. They will pay off for years to come.
December 25, 2024 at 5:10 PM
Midwits with a chatbot will never build better software than you.
These are two of my all-time favorite Software Engineering books:
1. The Pragmatic Programmer
2. Clean Code
Read them in 2025. They will pay off for years to come.
These are two of my all-time favorite Software Engineering books:
1. The Pragmatic Programmer
2. Clean Code
Read them in 2025. They will pay off for years to come.
This MIT class is still the best way to learn Linear Algebra.
It's free.
Gilbert Strang is one of those generational professors. The type of person that will leave a positive mark on you.
ocw.mit.edu/courses/18-...
It's free.
Gilbert Strang is one of those generational professors. The type of person that will leave a positive mark on you.
ocw.mit.edu/courses/18-...
December 22, 2024 at 2:24 PM
This MIT class is still the best way to learn Linear Algebra.
It's free.
Gilbert Strang is one of those generational professors. The type of person that will leave a positive mark on you.
ocw.mit.edu/courses/18-...
It's free.
Gilbert Strang is one of those generational professors. The type of person that will leave a positive mark on you.
ocw.mit.edu/courses/18-...
“This is not merely incremental improvement, but a genuine breakthrough”
François Chollet about the just released OpenAI o3 model.
François Chollet about the just released OpenAI o3 model.
December 20, 2024 at 7:27 PM
“This is not merely incremental improvement, but a genuine breakthrough”
François Chollet about the just released OpenAI o3 model.
François Chollet about the just released OpenAI o3 model.
You can opt out of Visual Studio Code's Copilot from using your data to train their models.
Works for free and pro users.
This option is enabled by default.
(You'll find this in your Copilot settings.)
Works for free and pro users.
This option is enabled by default.
(You'll find this in your Copilot settings.)
December 19, 2024 at 8:12 PM
You can opt out of Visual Studio Code's Copilot from using your data to train their models.
Works for free and pro users.
This option is enabled by default.
(You'll find this in your Copilot settings.)
Works for free and pro users.
This option is enabled by default.
(You'll find this in your Copilot settings.)
Boomer developers are not going to make it.
December 18, 2024 at 9:30 PM
Boomer developers are not going to make it.
You can integrate uv with your shell to enable autocompletion on the terminal!
The good stuff keeps getting better!
Take 5 minutes and look into uv. I'm willing to bet you'll like it a lot.
The good stuff keeps getting better!
Take 5 minutes and look into uv. I'm willing to bet you'll like it a lot.
December 16, 2024 at 7:05 PM
You can integrate uv with your shell to enable autocompletion on the terminal!
The good stuff keeps getting better!
Take 5 minutes and look into uv. I'm willing to bet you'll like it a lot.
The good stuff keeps getting better!
Take 5 minutes and look into uv. I'm willing to bet you'll like it a lot.
Game-changing upgrade:
Start using an automatic formatting tool that's opinionated.
I haven't found one tool for Python that's better than black. It's fast, and I love the way it formats my code.
Start using an automatic formatting tool that's opinionated.
I haven't found one tool for Python that's better than black. It's fast, and I love the way it formats my code.
December 13, 2024 at 9:15 PM
Game-changing upgrade:
Start using an automatic formatting tool that's opinionated.
I haven't found one tool for Python that's better than black. It's fast, and I love the way it formats my code.
Start using an automatic formatting tool that's opinionated.
I haven't found one tool for Python that's better than black. It's fast, and I love the way it formats my code.
Most companies are still experimenting with AI.
We are still far, far away from finding product-market-fit for all of the recent AI progress we've made.
Talk to people outside your online bubble:
99% of AI applications still use boosted trees and convolutional neural networks.
We are still far, far away from finding product-market-fit for all of the recent AI progress we've made.
Talk to people outside your online bubble:
99% of AI applications still use boosted trees and convolutional neural networks.
December 11, 2024 at 8:22 PM
Most companies are still experimenting with AI.
We are still far, far away from finding product-market-fit for all of the recent AI progress we've made.
Talk to people outside your online bubble:
99% of AI applications still use boosted trees and convolutional neural networks.
We are still far, far away from finding product-market-fit for all of the recent AI progress we've made.
Talk to people outside your online bubble:
99% of AI applications still use boosted trees and convolutional neural networks.
I feel seen.
(h/t to @forrestbrazeal)
(h/t to @forrestbrazeal)
December 11, 2024 at 12:00 AM
I feel seen.
(h/t to @forrestbrazeal)
(h/t to @forrestbrazeal)
List comprehensions are one of the most elegant features of Python.
I understand they look weird for people new to the language, but they are Pythonic and everyone uses them.
Attached you'll see two examples. The second one is the way to go.
I understand they look weird for people new to the language, but they are Pythonic and everyone uses them.
Attached you'll see two examples. The second one is the way to go.
December 10, 2024 at 6:00 PM
List comprehensions are one of the most elegant features of Python.
I understand they look weird for people new to the language, but they are Pythonic and everyone uses them.
Attached you'll see two examples. The second one is the way to go.
I understand they look weird for people new to the language, but they are Pythonic and everyone uses them.
Attached you'll see two examples. The second one is the way to go.
I used to think that writing comments on your code showed a lack of ability.
Your code must be self-explanatory. Comments are a crutch in many ways. They are hard to maintain, go out of sync easily, and often offer redundant information.
1/3
Your code must be self-explanatory. Comments are a crutch in many ways. They are hard to maintain, go out of sync easily, and often offer redundant information.
1/3
December 8, 2024 at 2:11 PM
I used to think that writing comments on your code showed a lack of ability.
Your code must be self-explanatory. Comments are a crutch in many ways. They are hard to maintain, go out of sync easily, and often offer redundant information.
1/3
Your code must be self-explanatory. Comments are a crutch in many ways. They are hard to maintain, go out of sync easily, and often offer redundant information.
1/3
This is how Large Language Models work.
December 5, 2024 at 6:09 PM
This is how Large Language Models work.
Using a calculator to review my son’s math is so slow.
I can now ask ChatGPT to review the answers.
It took me 10 minutes before. It now takes me 10 seconds.
I can now ask ChatGPT to review the answers.
It took me 10 minutes before. It now takes me 10 seconds.
December 3, 2024 at 11:59 PM
Using a calculator to review my son’s math is so slow.
I can now ask ChatGPT to review the answers.
It took me 10 minutes before. It now takes me 10 seconds.
I can now ask ChatGPT to review the answers.
It took me 10 minutes before. It now takes me 10 seconds.
We need a new word to represent "next level of hype."
December 2, 2024 at 6:33 PM
We need a new word to represent "next level of hype."
What’s happening? Who is David Mayer and why ChatGPT blows up when you ask?
December 1, 2024 at 6:00 PM
What’s happening? Who is David Mayer and why ChatGPT blows up when you ask?
November 30, 2024 at 8:26 PM
This is probably one of the little Python tricks I use the most:
`dir` returns everything that's available from an object/library.
Super convenient when you don't know what exactly you can do with the result of a function, or don't remember the name of a specific property.
`dir` returns everything that's available from an object/library.
Super convenient when you don't know what exactly you can do with the result of a function, or don't remember the name of a specific property.
November 29, 2024 at 7:00 PM
This is probably one of the little Python tricks I use the most:
`dir` returns everything that's available from an object/library.
Super convenient when you don't know what exactly you can do with the result of a function, or don't remember the name of a specific property.
`dir` returns everything that's available from an object/library.
Super convenient when you don't know what exactly you can do with the result of a function, or don't remember the name of a specific property.
A 10x improvement to any customer-facing application:
Start using semantic search in addition to your regular full-text searches.
This is a PostgreSQL query using OpenAI's embeddings to find relevant articles.
Results are magic.
Start using semantic search in addition to your regular full-text searches.
This is a PostgreSQL query using OpenAI's embeddings to find relevant articles.
Results are magic.
November 28, 2024 at 7:00 PM
A 10x improvement to any customer-facing application:
Start using semantic search in addition to your regular full-text searches.
This is a PostgreSQL query using OpenAI's embeddings to find relevant articles.
Results are magic.
Start using semantic search in addition to your regular full-text searches.
This is a PostgreSQL query using OpenAI's embeddings to find relevant articles.
Results are magic.
Chrome doesn't like your site:
November 27, 2024 at 3:54 PM
Chrome doesn't like your site:
If we change the activation function from "Linear" to "ReLU", the network will find a solution!
Same network structure. Different activation function.
ReLU introduces the non-linearity we need for the network to become much more powerful.
6/7
Same network structure. Different activation function.
ReLU introduces the non-linearity we need for the network to become much more powerful.
6/7
November 27, 2024 at 2:00 PM
If we change the activation function from "Linear" to "ReLU", the network will find a solution!
Same network structure. Different activation function.
ReLU introduces the non-linearity we need for the network to become much more powerful.
6/7
Same network structure. Different activation function.
ReLU introduces the non-linearity we need for the network to become much more powerful.
6/7
The same network, however, can't solve the second problem.
Look at the output here: it's all messed up.
We can't solve this problem with a single line, so we need to configure this network to find non-linear solutions.
5/7
Look at the output here: it's all messed up.
We can't solve this problem with a single line, so we need to configure this network to find non-linear solutions.
5/7
November 27, 2024 at 2:00 PM
The same network, however, can't solve the second problem.
Look at the output here: it's all messed up.
We can't solve this problem with a single line, so we need to configure this network to find non-linear solutions.
5/7
Look at the output here: it's all messed up.
We can't solve this problem with a single line, so we need to configure this network to find non-linear solutions.
5/7