Robert Knight
@robertknight2.bsky.social
Frontend Developer. Machine learning in Rust. https://github.com/robertknight.
Google AI doesn't know my IP, but kindly hallucinates one.
August 28, 2025 at 5:26 PM
Google AI doesn't know my IP, but kindly hallucinates one.
Why yes, I am a senior Mustache engineer with over 10 years experience.
August 5, 2025 at 10:08 AM
Why yes, I am a senior Mustache engineer with over 10 years experience.
@searchliaison.bsky.social, WTF is going on with the first search result here.
June 27, 2025 at 4:14 PM
@searchliaison.bsky.social, WTF is going on with the first search result here.
Firefox's "AI sidebar" feature has me scratching my head a bit. It feels weird to integrate an assistant into your browser which doesn't even know what page you are on.
May 15, 2025 at 7:51 AM
Firefox's "AI sidebar" feature has me scratching my head a bit. It feels weird to integrate an assistant into your browser which doesn't even know what page you are on.
Hypothesis now supports annotating images in PDFs - web.hypothes.is/web-app-start/.
May 14, 2025 at 10:13 AM
Hypothesis now supports annotating images in PDFs - web.hypothes.is/web-app-start/.
On-device AI is great in theory, but for open-domain tasks constraining the model size does limit the usefulness a lot.
April 30, 2025 at 8:33 AM
On-device AI is great in theory, but for open-domain tasks constraining the model size does limit the usefulness a lot.
How many points is a reasoning model worth?
April 17, 2025 at 10:03 AM
How many points is a reasoning model worth?
Better late than never, Hypothesis now lets you ping people by @-mentioning them while annotating web pages or PDFs.
April 10, 2025 at 7:04 AM
Better late than never, Hypothesis now lets you ping people by @-mentioning them while annotating web pages or PDFs.
AI has taken my job as an introvert.
February 13, 2025 at 11:46 AM
AI has taken my job as an introvert.
When it comes to color choices for service status, Docker Hub follows its own rules.
February 6, 2025 at 9:43 AM
When it comes to color choices for service status, Docker Hub follows its own rules.
I have been working on quantized model support for the next RTen release. Here is Whisper running on a Raspberry Pi Zero 2 (4-core Arm Cortex A53, 512MB RAM).
January 23, 2025 at 9:58 PM
I have been working on quantized model support for the next RTen release. Here is Whisper running on a Raspberry Pi Zero 2 (4-core Arm Cortex A53, 512MB RAM).
The corresponding profile in Chrome. Still a noticeable delay, but 10x less bad. Also its profile points at a culprit I'd more intuitively expect to see.
January 17, 2025 at 9:45 AM
The corresponding profile in Chrome. Still a noticeable delay, but 10x less bad. Also its profile points at a culprit I'd more intuitively expect to see.
GitHub helpfully provide sourcemaps for the original TS code. The offending function containing the `appendChild` call is here.
January 17, 2025 at 9:38 AM
GitHub helpfully provide sourcemaps for the original TS code. The offending function containing the `appendChild` call is here.
Hmm, a single `appendChild` call taking a _really_ long time.
January 17, 2025 at 9:38 AM
Hmm, a single `appendChild` call taking a _really_ long time.
Poking into why creating an inline comment in a 500 line PR in GitHub results in a two-second browser hang for me in Safari 18.1. cc @github.com
January 17, 2025 at 9:38 AM
Poking into why creating an inline comment in a 500 line PR in GitHub results in a two-second browser hang for me in Safari 18.1. cc @github.com
When the reviewer goes the extra mile after "LGTM"
November 20, 2024 at 2:07 PM
When the reviewer goes the extra mile after "LGTM"
RTen currently has a naive implementation of the Transpose operator for permuting tensor dimensions. Due to the way CPU caches work, this turns out to be really inefficient in the common case where tensor dimensions are powers of 2 (or "round numbers", as programmers call them)
November 20, 2024 at 2:07 PM
RTen currently has a naive implementation of the Transpose operator for permuting tensor dimensions. Due to the way CPU caches work, this turns out to be really inefficient in the common case where tensor dimensions are powers of 2 (or "round numbers", as programmers call them)
GPT 4 does semantic HTML
November 20, 2024 at 3:00 PM
GPT 4 does semantic HTML
I heard all the cool kids are using ConvNeXt as their backbone model (per https://arxiv.org/abs/2310.19909v1) so I spent a bunch of time optimizing the base model in my ONNX engine. Results have improved from "rubbish" to "comparable to PyTorch eager".
November 20, 2024 at 3:23 PM
I heard all the cool kids are using ConvNeXt as their backbone model (per https://arxiv.org/abs/2310.19909v1) so I spent a bunch of time optimizing the base model in my ONNX engine. Results have improved from "rubbish" to "comparable to PyTorch eager".
Perf of my Rust ONNX runtime is getting to the point where it is worth comparing to other ML engines. There is a long way to go to be generally competitive, but matmul performance is OK.
Also tried a WASM Relaxed-SIMD kernel, but wasm-bindgen doesn't like the new opcodes :(
Also tried a WASM Relaxed-SIMD kernel, but wasm-bindgen doesn't like the new opcodes :(
November 20, 2024 at 3:54 PM
Perf of my Rust ONNX runtime is getting to the point where it is worth comparing to other ML engines. There is a long way to go to be generally competitive, but matmul performance is OK.
Also tried a WASM Relaxed-SIMD kernel, but wasm-bindgen doesn't like the new opcodes :(
Also tried a WASM Relaxed-SIMD kernel, but wasm-bindgen doesn't like the new opcodes :(
Finally got my new OCR engine for JS / WebAssembly / Rust working end-to-end in Node (right). There is a lot of work still to do on the recognition model and performance under WASM. Nevertheless there are images for which it performs a lot better than tesseract-wasm (left).
November 20, 2024 at 3:54 PM
Finally got my new OCR engine for JS / WebAssembly / Rust working end-to-end in Node (right). There is a lot of work still to do on the recognition model and performance under WASM. Nevertheless there are images for which it performs a lot better than tesseract-wasm (left).
We're rolling out a change in @hypothes_is to show annotations side-by-side with web page content, attempting to be somewhat intelligent in determining what content the sidebar should / should-not cover.
November 20, 2024 at 3:54 PM
We're rolling out a change in @hypothes_is to show annotations side-by-side with web page content, attempting to be somewhat intelligent in determining what content the sidebar should / should-not cover.
A few meters before the end of the London Marathon (2:37:58). I love how we're all getting ready to stop our watches exactly on the line.
November 20, 2024 at 4:26 PM
A few meters before the end of the London Marathon (2:37:58). I love how we're all getting ready to stop our watches exactly on the line.
My new Rust-based OCR engine is gradually getting closer to shipping a preview. First image shows detected text words from Tesseract, which struggles with shadows (see bottom) and curved lines (see right) in the photo. Second image shows text outlines from the new engine.
November 20, 2024 at 4:39 PM
My new Rust-based OCR engine is gradually getting closer to shipping a preview. First image shows detected text words from Tesseract, which struggles with shadows (see bottom) and curved lines (see right) in the photo. Second image shows text outlines from the new engine.
If you haven't implemented it with dot products, you're not trying hard enough.
November 21, 2024 at 4:01 PM
If you haven't implemented it with dot products, you're not trying hard enough.