Installed a bird camera in my balcony and captured this red bulbul in a frame.
November 18, 2025 at 7:49 AM
Installed a bird camera in my balcony and captured this red bulbul in a frame.
Ok, whoever filled this, you have my respect!
made me chuckle (:
made me chuckle (:
November 18, 2025 at 4:30 AM
Ok, whoever filled this, you have my respect!
made me chuckle (:
made me chuckle (:
An intriguing #book.
In it, the author - a philosopher - argues that finetuning of our physical constants is an evidence of a purpose in our universe, which includes emergence of life and realization of value.
In it, the author - a philosopher - argues that finetuning of our physical constants is an evidence of a purpose in our universe, which includes emergence of life and realization of value.
November 17, 2025 at 3:23 AM
An intriguing #book.
In it, the author - a philosopher - argues that finetuning of our physical constants is an evidence of a purpose in our universe, which includes emergence of life and realization of value.
In it, the author - a philosopher - argues that finetuning of our physical constants is an evidence of a purpose in our universe, which includes emergence of life and realization of value.
No, Calvin, you can be a kid forever :)
Just keep playing.
Just keep playing.
November 14, 2025 at 2:54 AM
No, Calvin, you can be a kid forever :)
Just keep playing.
Just keep playing.
It's crazy that ChatGPT writes like @sama – often in all small case.
Goes on to show how models inherit biases and style of their creators.
Goes on to show how models inherit biases and style of their creators.
November 13, 2025 at 2:30 PM
It's crazy that ChatGPT writes like @sama – often in all small case.
Goes on to show how models inherit biases and style of their creators.
Goes on to show how models inherit biases and style of their creators.
This is a beautiful article by Vitalik!
Strongly recommended.
Strongly recommended.
November 11, 2025 at 1:58 AM
This is a beautiful article by Vitalik!
Strongly recommended.
Strongly recommended.
It has been roughly 6 months of pursuing research at @lossfunk!
Check out the kinds of questions we're exploring 👇
Check out the kinds of questions we're exploring 👇
November 4, 2025 at 1:05 PM
It has been roughly 6 months of pursuing research at @lossfunk!
Check out the kinds of questions we're exploring 👇
Check out the kinds of questions we're exploring 👇
Before creating superintelligence, Meta should teach its dumb AI that it cannot set reminders and shouldn’t promise so.
Honestly, the inability to do something as simple as this on WhatsApp is a lost opportunity.
Honestly, the inability to do something as simple as this on WhatsApp is a lost opportunity.
November 3, 2025 at 1:34 PM
Before creating superintelligence, Meta should teach its dumb AI that it cannot set reminders and shouldn’t promise so.
Honestly, the inability to do something as simple as this on WhatsApp is a lost opportunity.
Honestly, the inability to do something as simple as this on WhatsApp is a lost opportunity.
The colors we can name in our language is influenced by what ripe fruits and edible objects can we find in our ancestral environment.
🤯
🤯
November 3, 2025 at 10:07 AM
The colors we can name in our language is influenced by what ripe fruits and edible objects can we find in our ancestral environment.
🤯
🤯
Truer that it should be!
November 2, 2025 at 1:40 PM
Truer that it should be!
Why this happens?
Because modern base models have ingested all the discussion on chat templates on the internet and hence will know how to close chat tags etc.
Because modern base models have ingested all the discussion on chat templates on the internet and hence will know how to close chat tags etc.
November 2, 2025 at 3:49 AM
Why this happens?
Because modern base models have ingested all the discussion on chat templates on the internet and hence will know how to close chat tags etc.
Because modern base models have ingested all the discussion on chat templates on the internet and hence will know how to close chat tags etc.
OK, i love base models!
Not many may know but base models can follow instructions. No separate instruction tuning needed!
See here Qwen-8B-Base. All you do is to use a chat template as prompt for completion and your base model will happily give a response in the *same* format.
Not many may know but base models can follow instructions. No separate instruction tuning needed!
See here Qwen-8B-Base. All you do is to use a chat template as prompt for completion and your base model will happily give a response in the *same* format.
November 2, 2025 at 3:49 AM
OK, i love base models!
Not many may know but base models can follow instructions. No separate instruction tuning needed!
See here Qwen-8B-Base. All you do is to use a chat template as prompt for completion and your base model will happily give a response in the *same* format.
Not many may know but base models can follow instructions. No separate instruction tuning needed!
See here Qwen-8B-Base. All you do is to use a chat template as prompt for completion and your base model will happily give a response in the *same* format.
Can’t fucking believe it - I finished my first 10k run! 🎉🏃
Slept at 9pm, woke up at 4:30am, ran 1.5 hours nonstop - you can just do things!
Slept at 9pm, woke up at 4:30am, ran 1.5 hours nonstop - you can just do things!
October 31, 2025 at 3:08 AM
Can’t fucking believe it - I finished my first 10k run! 🎉🏃
Slept at 9pm, woke up at 4:30am, ran 1.5 hours nonstop - you can just do things!
Slept at 9pm, woke up at 4:30am, ran 1.5 hours nonstop - you can just do things!
I'm in love with perplexity (not the app, but the measure of surprise)
Had a hypothesis that perplexity for human-written text is higher and qualitatively different than AI-generated text.
Did some quick experiments to verify that and it's true for different text sizes and domains!
Had a hypothesis that perplexity for human-written text is higher and qualitatively different than AI-generated text.
Did some quick experiments to verify that and it's true for different text sizes and domains!
October 30, 2025 at 12:30 PM
I'm in love with perplexity (not the app, but the measure of surprise)
Had a hypothesis that perplexity for human-written text is higher and qualitatively different than AI-generated text.
Did some quick experiments to verify that and it's true for different text sizes and domains!
Had a hypothesis that perplexity for human-written text is higher and qualitatively different than AI-generated text.
Did some quick experiments to verify that and it's true for different text sizes and domains!
Today I learned: forced yawning is a thing, and may help fall asleep / relax by lowering cortisol.
(If that’s the case, it may help during anxiety too)
(If that’s the case, it may help during anxiety too)
October 29, 2025 at 2:30 AM
Today I learned: forced yawning is a thing, and may help fall asleep / relax by lowering cortisol.
(If that’s the case, it may help during anxiety too)
(If that’s the case, it may help during anxiety too)
This was surprising to me!
New @lossfunk exploration on how the language in which you ask question to LLM impacts its "reasoning" performance.
Full article below 👇
New @lossfunk exploration on how the language in which you ask question to LLM impacts its "reasoning" performance.
Full article below 👇
October 28, 2025 at 11:14 AM
This was surprising to me!
New @lossfunk exploration on how the language in which you ask question to LLM impacts its "reasoning" performance.
Full article below 👇
New @lossfunk exploration on how the language in which you ask question to LLM impacts its "reasoning" performance.
Full article below 👇
A fantastic demo of how manufactured goods get cheaper (in real terms) over time.
October 28, 2025 at 2:44 AM
A fantastic demo of how manufactured goods get cheaper (in real terms) over time.
My favorite thing these days is to deep dive into a topic, make lots of handwritten notes and then ask ChatGPT for its thoughts.
I often ask it to:
- steelman my case, and its criticism
- share what papers and books I should read
- suggest more ideas and experiments
I often ask it to:
- steelman my case, and its criticism
- share what papers and books I should read
- suggest more ideas and experiments
October 27, 2025 at 10:14 AM
My favorite thing these days is to deep dive into a topic, make lots of handwritten notes and then ask ChatGPT for its thoughts.
I often ask it to:
- steelman my case, and its criticism
- share what papers and books I should read
- suggest more ideas and experiments
I often ask it to:
- steelman my case, and its criticism
- share what papers and books I should read
- suggest more ideas and experiments
What an amazing, amazing #book!
Written by three psychiatrists, it blends science and poetry to describe how hopelessly we humans (and all mammals) are dependent on love.
Written by three psychiatrists, it blends science and poetry to describe how hopelessly we humans (and all mammals) are dependent on love.
October 26, 2025 at 6:48 AM
What an amazing, amazing #book!
Written by three psychiatrists, it blends science and poetry to describe how hopelessly we humans (and all mammals) are dependent on love.
Written by three psychiatrists, it blends science and poetry to describe how hopelessly we humans (and all mammals) are dependent on love.
We're (perhaps) the first and only AI lab in 🇮🇳 India working on accelerating science and research using AI.
Super happy to see the paper written by @lossfunk's AI scientist get accepted at the world's first conference with LLMs as primary authors and reviewers!
Super happy to see the paper written by @lossfunk's AI scientist get accepted at the world's first conference with LLMs as primary authors and reviewers!
October 25, 2025 at 11:27 AM
We're (perhaps) the first and only AI lab in 🇮🇳 India working on accelerating science and research using AI.
Super happy to see the paper written by @lossfunk's AI scientist get accepted at the world's first conference with LLMs as primary authors and reviewers!
Super happy to see the paper written by @lossfunk's AI scientist get accepted at the world's first conference with LLMs as primary authors and reviewers!
LLMs are a blessing for book lovers. They're able to fish out specific books on any topic imaginable - all of these would have been impossible to discover without them.
My favorite thing to do is to end my chats with: "based on what we talked, what books should I read?"
My favorite thing to do is to end my chats with: "based on what we talked, what books should I read?"
October 24, 2025 at 1:30 PM
LLMs are a blessing for book lovers. They're able to fish out specific books on any topic imaginable - all of these would have been impossible to discover without them.
My favorite thing to do is to end my chats with: "based on what we talked, what books should I read?"
My favorite thing to do is to end my chats with: "based on what we talked, what books should I read?"
That moment when your LLM says something that soothes the soul.
October 23, 2025 at 9:21 AM
That moment when your LLM says something that soothes the soul.
It's not commonly known but BMI cutoffs for obesity for indians is different vs west.
This is because caucasian populations generally have more lean mass while asian populations have higher visceral fat.
The correct test for obesity is via fat %, and it should be <25% for men.
This is because caucasian populations generally have more lean mass while asian populations have higher visceral fat.
The correct test for obesity is via fat %, and it should be <25% for men.
October 23, 2025 at 3:30 AM
It's not commonly known but BMI cutoffs for obesity for indians is different vs west.
This is because caucasian populations generally have more lean mass while asian populations have higher visceral fat.
The correct test for obesity is via fat %, and it should be <25% for men.
This is because caucasian populations generally have more lean mass while asian populations have higher visceral fat.
The correct test for obesity is via fat %, and it should be <25% for men.
It's so crazy that *different* networks with different modalities end up having similar representations!
The fact that "red ball" and an image of a red ball share embedding space happens because they tend to co-occur, and that is enough for them to converge.
The fact that "red ball" and an image of a red ball share embedding space happens because they tend to co-occur, and that is enough for them to converge.
October 22, 2025 at 4:00 PM
It's so crazy that *different* networks with different modalities end up having similar representations!
The fact that "red ball" and an image of a red ball share embedding space happens because they tend to co-occur, and that is enough for them to converge.
The fact that "red ball" and an image of a red ball share embedding space happens because they tend to co-occur, and that is enough for them to converge.