Lex
@notesbylex.com
Senior MLE at @Canva. Full-stack developer.
Kaggle Notebook Master and collector of competition silver medals. https://www.kaggle.com/lextoumbourou
Talks about Software dev, ML, generative art, note-taking, dog photos. Mostly cross-posting from Mastodon.
Kaggle Notebook Master and collector of competition silver medals. https://www.kaggle.com/lextoumbourou
Talks about Software dev, ML, generative art, note-taking, dog photos. Mostly cross-posting from Mastodon.
This new image editing model from Black Forest Labs called **FLUX.1 Kontext** is really good. I ran some experiments on photos of Doggo, and couldn't believe how well it could maintain character consistent across multiple turns of editing.
https://notesbylex.com/absurdly-good-doggo-consistency-wit…
https://notesbylex.com/absurdly-good-doggo-consistency-wit…
June 1, 2025 at 11:25 PM
This new image editing model from Black Forest Labs called **FLUX.1 Kontext** is really good. I ran some experiments on photos of Doggo, and couldn't believe how well it could maintain character consistent across multiple turns of editing.
https://notesbylex.com/absurdly-good-doggo-consistency-wit…
https://notesbylex.com/absurdly-good-doggo-consistency-wit…
A hilariously simple repro of OpenAI's test-time scaling paradigm called "Budget Scaling": end the thinking when your token budget is met, or append "Wait" to the model's generation to keep thinking, allowing the model to fix incorrect reasoning steps.
arxiv.org/abs/2501.19393
arxiv.org/abs/2501.19393
February 3, 2025 at 6:14 AM
A hilariously simple repro of OpenAI's test-time scaling paradigm called "Budget Scaling": end the thinking when your token budget is met, or append "Wait" to the model's generation to keep thinking, allowing the model to fix incorrect reasoning steps.
arxiv.org/abs/2501.19393
arxiv.org/abs/2501.19393
A method for evaluating data for preference optimisation.
Rejecting Instruction Preferences (RIP) can filter prompts from existing training sets or make high-quality synthetic datasets. They see large performance gains across various benchmarks compared to unfiltered data.
arxiv.org/abs/2501.18578
Rejecting Instruction Preferences (RIP) can filter prompts from existing training sets or make high-quality synthetic datasets. They see large performance gains across various benchmarks compared to unfiltered data.
arxiv.org/abs/2501.18578
January 31, 2025 at 5:22 AM
A method for evaluating data for preference optimisation.
Rejecting Instruction Preferences (RIP) can filter prompts from existing training sets or make high-quality synthetic datasets. They see large performance gains across various benchmarks compared to unfiltered data.
arxiv.org/abs/2501.18578
Rejecting Instruction Preferences (RIP) can filter prompts from existing training sets or make high-quality synthetic datasets. They see large performance gains across various benchmarks compared to unfiltered data.
arxiv.org/abs/2501.18578
ChatGPT seems to be down.
January 23, 2025 at 11:54 AM
ChatGPT seems to be down.
I finally got around to updating my Obsidian plugin: obsidian-title-as-link-text.
1. Now support Wikilinks with a display name: `[[my-page|A Cool Page]]`
2. Also added support for aliases.
See the release notes: github.com/lextoumbouro...
1. Now support Wikilinks with a display name: `[[my-page|A Cool Page]]`
2. Also added support for aliases.
See the release notes: github.com/lextoumbouro...
January 18, 2025 at 5:50 AM
I finally got around to updating my Obsidian plugin: obsidian-title-as-link-text.
1. Now support Wikilinks with a display name: `[[my-page|A Cool Page]]`
2. Also added support for aliases.
See the release notes: github.com/lextoumbouro...
1. Now support Wikilinks with a display name: `[[my-page|A Cool Page]]`
2. Also added support for aliases.
See the release notes: github.com/lextoumbouro...
Hello world! First post on Bluesky.
I'm a senior MLE at Canva, working on LLMs and agents.
Sometimes post ML paper summaries and notes about software engineering, although I'm very inconsistent.
Here's a photo of me and my dog, Doggo.
I'm a senior MLE at Canva, working on LLMs and agents.
Sometimes post ML paper summaries and notes about software engineering, although I'm very inconsistent.
Here's a photo of me and my dog, Doggo.
November 16, 2024 at 11:15 PM
Hello world! First post on Bluesky.
I'm a senior MLE at Canva, working on LLMs and agents.
Sometimes post ML paper summaries and notes about software engineering, although I'm very inconsistent.
Here's a photo of me and my dog, Doggo.
I'm a senior MLE at Canva, working on LLMs and agents.
Sometimes post ML paper summaries and notes about software engineering, although I'm very inconsistent.
Here's a photo of me and my dog, Doggo.