Kriz Tahimic
@kriztahimic.bsky.social
Undergraduate | Interested in AI & Mechanistic Interpretability
it just sunk in to me that a transformer's entire information flow is just a flat, fixed-width vector getting added to. superposition exploits high-dimensional geometry + interference tolerance to cram millions of features into thousands of dims. ChatGPT emerges from this somehow 🤯
December 8, 2025 at 2:21 PM
it just sunk in to me that a transformer's entire information flow is just a flat, fixed-width vector getting added to. superposition exploits high-dimensional geometry + interference tolerance to cram millions of features into thousands of dims. ChatGPT emerges from this somehow 🤯
It's hard to believe that the predicting machine can do math. LLMs do not do it through step-by-step logic but through more complex computation. We can definitely make this more efficient.
March 10, 2025 at 2:42 AM
It's hard to believe that the predicting machine can do math. LLMs do not do it through step-by-step logic but through more complex computation. We can definitely make this more efficient.
It's so fun to do research!
Especially when things are working out 😉😉😆
Especially when things are working out 😉😉😆
January 26, 2025 at 9:08 AM
It's so fun to do research!
Especially when things are working out 😉😉😆
Especially when things are working out 😉😉😆
I want to develop Mathematical Maturity 😅
December 19, 2024 at 4:51 AM
I want to develop Mathematical Maturity 😅
It’s a shock to me that we are not actually getting the y-coordinate when we do sin(θ). What we actually get is the ratio of rise over hypothenuse which just often the case is similar to the rise as we are commonly working with unit ratio. 😄
December 18, 2024 at 9:38 AM
It’s a shock to me that we are not actually getting the y-coordinate when we do sin(θ). What we actually get is the ratio of rise over hypothenuse which just often the case is similar to the rise as we are commonly working with unit ratio. 😄
Optics are important, but I'm happier that I decided to be less worried about it 😄
December 18, 2024 at 2:04 AM
Optics are important, but I'm happier that I decided to be less worried about it 😄
Experimenting with switching to Light mode throughout all the apps I’m using HAHA! Been nice so far. Felt more awake 😆
December 17, 2024 at 4:39 AM
Experimenting with switching to Light mode throughout all the apps I’m using HAHA! Been nice so far. Felt more awake 😆
Epiphany feels like unlocking a new real estate in my brain.
December 17, 2024 at 4:32 AM
Epiphany feels like unlocking a new real estate in my brain.
Some things can only be digested with pen & paper
December 16, 2024 at 5:09 AM
Some things can only be digested with pen & paper
I’m starting to accept the fact that I’m not a fast learner. But what I can take pride in is my perseverance. I may take longer time than others before I understand a topic but I’ll just defeat the problem with shear time investment and being strategic about what to work on 😄
December 16, 2024 at 3:10 AM
I’m starting to accept the fact that I’m not a fast learner. But what I can take pride in is my perseverance. I may take longer time than others before I understand a topic but I’ll just defeat the problem with shear time investment and being strategic about what to work on 😄
I'm increasingly fascinated by linear algebra, especially matrix multiplication. What's amazing is that "Linear Combinations" are really just an expanded view of matrix multiplication.
December 15, 2024 at 9:40 AM
I'm increasingly fascinated by linear algebra, especially matrix multiplication. What's amazing is that "Linear Combinations" are really just an expanded view of matrix multiplication.
Here I am in my 5th time reviewing @3blue1brown.bsky.social Essence of linear algebra series. Every time I rewatch it something new always clicks in my head. 😄
December 15, 2024 at 3:49 AM
Here I am in my 5th time reviewing @3blue1brown.bsky.social Essence of linear algebra series. Every time I rewatch it something new always clicks in my head. 😄
It just occurred to me that the word “integration” is synonymous with amalgamation, which makes more sense!
I now see why integration is a summation of all the infinitesimally small pieces.
Moreover, the symbol for integration is just a “long S,” which stands for “Sum!”
I now see why integration is a summation of all the infinitesimally small pieces.
Moreover, the symbol for integration is just a “long S,” which stands for “Sum!”
December 8, 2024 at 11:51 PM
It just occurred to me that the word “integration” is synonymous with amalgamation, which makes more sense!
I now see why integration is a summation of all the infinitesimally small pieces.
Moreover, the symbol for integration is just a “long S,” which stands for “Sum!”
I now see why integration is a summation of all the infinitesimally small pieces.
Moreover, the symbol for integration is just a “long S,” which stands for “Sum!”
I feel embarrassed but it just now makes more sense to me how probability works. That it is always in its normalized form so it can tell the relative likelihood of an event. Also, It’s just a smart way of using ratios 😄
December 8, 2024 at 5:57 AM
I feel embarrassed but it just now makes more sense to me how probability works. That it is always in its normalized form so it can tell the relative likelihood of an event. Also, It’s just a smart way of using ratios 😄
Reminder to self: Great things take time.
December 5, 2024 at 1:29 PM
Reminder to self: Great things take time.
I now worry less about the AI Takeover scenario and more about AI job displacement.
December 4, 2024 at 6:29 AM
I now worry less about the AI Takeover scenario and more about AI job displacement.
🧵I’m starting to lean my opinion that scaling will not take us to TAI or AI-reasoners. That we need n>2 breakthroughs or architectural improvements before TAI.
December 2, 2024 at 1:49 PM
🧵I’m starting to lean my opinion that scaling will not take us to TAI or AI-reasoners. That we need n>2 breakthroughs or architectural improvements before TAI.
My analogy for each part of the Adam optimizer: (a) Momentum is like the velocity of a car that speeds up the step in the average direction while (b) RMSprop is like the lowered feature of the vehicle so it can maneuver more swiftly during turns and not be off track.
November 28, 2024 at 7:01 AM
My analogy for each part of the Adam optimizer: (a) Momentum is like the velocity of a car that speeds up the step in the average direction while (b) RMSprop is like the lowered feature of the vehicle so it can maneuver more swiftly during turns and not be off track.
Reposted by Kriz Tahimic
Even as an interpretable ML researcher, I wasn't sure what to make of Mechanistic Interpretability, which seemed to come out of nowhere not too long ago.
But then I found the paper "Mechanistic?" by
@nsaphra.bsky.social and @sarah-nlp.bsky.social, which clarified things.
But then I found the paper "Mechanistic?" by
@nsaphra.bsky.social and @sarah-nlp.bsky.social, which clarified things.
November 20, 2024 at 8:00 AM
Even as an interpretable ML researcher, I wasn't sure what to make of Mechanistic Interpretability, which seemed to come out of nowhere not too long ago.
But then I found the paper "Mechanistic?" by
@nsaphra.bsky.social and @sarah-nlp.bsky.social, which clarified things.
But then I found the paper "Mechanistic?" by
@nsaphra.bsky.social and @sarah-nlp.bsky.social, which clarified things.