@mockapapella.bsky.social
Pinned
Focal loss is a method of calculating loss for a neural network where less common and difficult classification examples are balanced against more common and easier classification tasks.
Just realized something. You know how new programming languages need to come with basically everything (package manager, linter, LSP, etc.) to gain adoption nowadays?
I think LLMs might be included soon. Like a small, local LLM trained only on that language.
I think LLMs might be included soon. Like a small, local LLM trained only on that language.
April 10, 2025 at 5:01 PM
Just realized something. You know how new programming languages need to come with basically everything (package manager, linter, LSP, etc.) to gain adoption nowadays?
I think LLMs might be included soon. Like a small, local LLM trained only on that language.
I think LLMs might be included soon. Like a small, local LLM trained only on that language.
RAG is dead, long live RAG
An article in which I point out the biggest flaw with how RAG systems are implemented today, and how you can fix them.
An article in which I point out the biggest flaw with how RAG systems are implemented today, and how you can fix them.
RAG is dead, long live RAG
With the release of Llama 4 and its 10 million token context window, once again the usefulness of RAG as an architecture pattern is being brought up.
open.substack.com
April 8, 2025 at 3:41 PM
RAG is dead, long live RAG
An article in which I point out the biggest flaw with how RAG systems are implemented today, and how you can fix them.
An article in which I point out the biggest flaw with how RAG systems are implemented today, and how you can fix them.
New article: SWE-bench does not determine agency
SWE-bench does not determine agency
3 days ago, I came across this post on Twitter.
open.substack.com
April 6, 2025 at 11:08 PM
New article: SWE-bench does not determine agency
At this stage in the video creation process. I'm so close to finishing the outline for the animations. I'm eager to start actually building them.
April 2, 2025 at 1:55 AM
At this stage in the video creation process. I'm so close to finishing the outline for the animations. I'm eager to start actually building them.
Get so good that when people ask who you are you can tell them to just ask their favorite AI model
April 1, 2025 at 8:42 PM
Get so good that when people ask who you are you can tell them to just ask their favorite AI model
4o is really consistent at generating stock icon images
I take this, throw it into my PNG -> SVG converter, and I've got icons I can animate in Manim
I take this, throw it into my PNG -> SVG converter, and I've got icons I can animate in Manim
April 1, 2025 at 5:08 AM
4o is really consistent at generating stock icon images
I take this, throw it into my PNG -> SVG converter, and I've got icons I can animate in Manim
I take this, throw it into my PNG -> SVG converter, and I've got icons I can animate in Manim
📢 New repo!
I put together a minimal demo showing CUDA failing in child processes - a bug I ran into ages ago. Shows why PyTorch/CUDA breaks when working with forked processes.
Code's up now if you wanna play with it. I'll break it down in one of my future YouTube videos.
I put together a minimal demo showing CUDA failing in child processes - a bug I ran into ages ago. Shows why PyTorch/CUDA breaks when working with forked processes.
Code's up now if you wanna play with it. I'll break it down in one of my future YouTube videos.
March 30, 2025 at 7:59 PM
📢 New repo!
I put together a minimal demo showing CUDA failing in child processes - a bug I ran into ages ago. Shows why PyTorch/CUDA breaks when working with forked processes.
Code's up now if you wanna play with it. I'll break it down in one of my future YouTube videos.
I put together a minimal demo showing CUDA failing in child processes - a bug I ran into ages ago. Shows why PyTorch/CUDA breaks when working with forked processes.
Code's up now if you wanna play with it. I'll break it down in one of my future YouTube videos.
My personal litmus test for image models is if they can generate realistic circuit board traces. It's one of the first things I try on every image model.
So far every one I've tried has failed. They always come out smudgy.
So far every one I've tried has failed. They always come out smudgy.
March 30, 2025 at 5:12 AM
My personal litmus test for image models is if they can generate realistic circuit board traces. It's one of the first things I try on every image model.
So far every one I've tried has failed. They always come out smudgy.
So far every one I've tried has failed. They always come out smudgy.
I may need to split my video (working title: "How does ChatGPT handle 1B messages daily?") into sections. It's a semi-technical overview of LLM deployment from first principles, but I'm ~20min in and JUST hit the cloud. I've got 20-30 more minutes worth of content!
March 30, 2025 at 4:51 AM
I may need to split my video (working title: "How does ChatGPT handle 1B messages daily?") into sections. It's a semi-technical overview of LLM deployment from first principles, but I'm ~20min in and JUST hit the cloud. I've got 20-30 more minutes worth of content!
New Claude UI redesign lets you download artifacts as markdown or PDF
March 30, 2025 at 3:29 AM
New Claude UI redesign lets you download artifacts as markdown or PDF
I set up a script to automatically load certain programs on startup and place them on different desktops. This feels great. I no longer need to have the mental overhead of remembering where I last was in my projects and can just power on my computer and jump right back into the flow.
March 29, 2025 at 3:25 PM
I set up a script to automatically load certain programs on startup and place them on different desktops. This feels great. I no longer need to have the mental overhead of remembering where I last was in my projects and can just power on my computer and jump right back into the flow.
New Anthropic UI feels really crisp
March 28, 2025 at 2:49 AM
New Anthropic UI feels really crisp
I take a boost or block approach to trolls and haters.
If your negativity boosts my content, I'll let it slide. Once it doesn't though, and you start polluting my other posts, I'm blocking you.
Does not apply to criticisms done in good faith. I like healthy discussions, just not one sided bashing.
If your negativity boosts my content, I'll let it slide. Once it doesn't though, and you start polluting my other posts, I'm blocking you.
Does not apply to criticisms done in good faith. I like healthy discussions, just not one sided bashing.
March 27, 2025 at 5:19 PM
I take a boost or block approach to trolls and haters.
If your negativity boosts my content, I'll let it slide. Once it doesn't though, and you start polluting my other posts, I'm blocking you.
Does not apply to criticisms done in good faith. I like healthy discussions, just not one sided bashing.
If your negativity boosts my content, I'll let it slide. Once it doesn't though, and you start polluting my other posts, I'm blocking you.
Does not apply to criticisms done in good faith. I like healthy discussions, just not one sided bashing.
OK Ghiblifying everything was fun but how about a slightly different change of pace.
A take on an old classic
A take on an old classic
March 27, 2025 at 3:47 PM
OK Ghiblifying everything was fun but how about a slightly different change of pace.
A take on an old classic
A take on an old classic
Anime profile pics are no longer a strong signal of a high quality account
March 27, 2025 at 4:29 AM
Anime profile pics are no longer a strong signal of a high quality account
Buying blinds on Ali Baba and every time I'm on this site I just like to explore around.
Decided to look up various food products and found this gem.
Decided to look up various food products and found this gem.
March 27, 2025 at 3:23 AM
Buying blinds on Ali Baba and every time I'm on this site I just like to explore around.
Decided to look up various food products and found this gem.
Decided to look up various food products and found this gem.
Gemini 2.5 Pro Dropped
- 1M input context
- 65K output context
- Higher (per minute) rate limits
Try it here: buff.ly/uFVJEr7
- 1M input context
- 65K output context
- Higher (per minute) rate limits
Try it here: buff.ly/uFVJEr7
March 25, 2025 at 5:37 PM
Gemini 2.5 Pro Dropped
- 1M input context
- 65K output context
- Higher (per minute) rate limits
Try it here: buff.ly/uFVJEr7
- 1M input context
- 65K output context
- Higher (per minute) rate limits
Try it here: buff.ly/uFVJEr7
Asked it to add a very small feature in a small code base. It proceeded to doom loop, reading and rereading the same files to "get more context". Planning is not its strong suite, but when I reset the chat and just asked it point blank to implement the feature it did.
Time to try out the newest DeepSeek V3 update
March 25, 2025 at 3:22 PM
Asked it to add a very small feature in a small code base. It proceeded to doom loop, reading and rereading the same files to "get more context". Planning is not its strong suite, but when I reset the chat and just asked it point blank to implement the feature it did.
Time to try out the newest DeepSeek V3 update
March 25, 2025 at 2:04 PM
Time to try out the newest DeepSeek V3 update
Been quiet over the last week. Been switching up my focus. I still plan on going deep with learning AI, but there is an obvious gap in the ecosystem right now around how to build high quality systems.
I’m going to make videos to fix this simply because it needs to be done.
I’m going to make videos to fix this simply because it needs to be done.
March 25, 2025 at 4:17 AM
Been quiet over the last week. Been switching up my focus. I still plan on going deep with learning AI, but there is an obvious gap in the ecosystem right now around how to build high quality systems.
I’m going to make videos to fix this simply because it needs to be done.
I’m going to make videos to fix this simply because it needs to be done.
o1-pro API pricing makes GPT 4.5's pricing look like Sonnet 3.7's
March 20, 2025 at 2:04 AM
o1-pro API pricing makes GPT 4.5's pricing look like Sonnet 3.7's
I just gave Vercel another shot using Claude Code this time around. Never deployed a JS app to prod before.
God damn that's a smooth experience. Is this what I've been missing? This is how I'm deploying all of my side projects from now on.
God damn that's a smooth experience. Is this what I've been missing? This is how I'm deploying all of my side projects from now on.
March 19, 2025 at 8:51 PM
I just gave Vercel another shot using Claude Code this time around. Never deployed a JS app to prod before.
God damn that's a smooth experience. Is this what I've been missing? This is how I'm deploying all of my side projects from now on.
God damn that's a smooth experience. Is this what I've been missing? This is how I'm deploying all of my side projects from now on.
Experimenting with a Gemini Flash -> PNG to SVG converter -> Manim pipeline
Asked Gemini to create a stick figure, then to create a modified pose for that stick figure, converted them to SVG using a tool that Claude built, and placed them into a Manim animation to interpolate between the two poses
Asked Gemini to create a stick figure, then to create a modified pose for that stick figure, converted them to SVG using a tool that Claude built, and placed them into a Manim animation to interpolate between the two poses
March 19, 2025 at 7:33 PM
Experimenting with a Gemini Flash -> PNG to SVG converter -> Manim pipeline
Asked Gemini to create a stick figure, then to create a modified pose for that stick figure, converted them to SVG using a tool that Claude built, and placed them into a Manim animation to interpolate between the two poses
Asked Gemini to create a stick figure, then to create a modified pose for that stick figure, converted them to SVG using a tool that Claude built, and placed them into a Manim animation to interpolate between the two poses
Information asymmetry at the extremes makes sane ideas look extremely stupid to the average person.
March 19, 2025 at 3:55 PM
Information asymmetry at the extremes makes sane ideas look extremely stupid to the average person.
Neat trick I’ve been trying the last couple of days that seems to be working well.
As long as I drink a large amount of water in the evening (been doing 64oz), I can have a few cups of coffee late in the afternoon and still get a good night’s rest.
As long as I drink a large amount of water in the evening (been doing 64oz), I can have a few cups of coffee late in the afternoon and still get a good night’s rest.
March 19, 2025 at 12:08 PM
Neat trick I’ve been trying the last couple of days that seems to be working well.
As long as I drink a large amount of water in the evening (been doing 64oz), I can have a few cups of coffee late in the afternoon and still get a good night’s rest.
As long as I drink a large amount of water in the evening (been doing 64oz), I can have a few cups of coffee late in the afternoon and still get a good night’s rest.