Hacker News: jadidbourbakihttps://news.ycombinator.com/user?id=jadidbourbakiHacker News RSShttps://hnrss.org/hnrss v2.1.1Tue, 29 Sep 2026 04:32:31 +0000<![CDATA[New comment by jadidbourbaki in "Faster prompt lookup drafting in llama.cpp"]]>As an update to this, Lemire absolutely crushed it! I benchmarked his change today. Here is my PR comment:

https://github.com/jadidbourbaki/llama.cpp/pull/12#issuecomm...

]]>Mon, 28 Sep 2026 03:13:57 +0000https://news.ycombinator.com/item?id=49873164jadidbourbakihttps://news.ycombinator.com/item?id=49873164https://news.ycombinator.com/item?id=49873164<![CDATA[New comment by jadidbourbaki in "Faster prompt lookup drafting in llama.cpp"]]>Btw, if anyone has experience with the open source community in general and llama.cpp in specific, I would greatly appreciate some advice. I’m facing a bit of an interpersonal issue that I really hope is resolved without any ill will. Here is the context:

https://www.reddit.com/r/LocalLLaMA/comments/1wr5ylm/comment...

Any advice for what I can do? Due to this, I cannot create a PR or issue in the llama.cpp repository. However, I am worried about bothering the maintainers on other channels in case it aggravates them further. Thank you for your help!

]]>Sun, 27 Sep 2026 03:29:40 +0000https://news.ycombinator.com/item?id=49863097jadidbourbakihttps://news.ycombinator.com/item?id=49863097https://news.ycombinator.com/item?id=49863097<![CDATA[New comment by jadidbourbaki in "Faster prompt lookup drafting in llama.cpp"]]>Fun update to this: Daniel Lemire added another optimization to make this even faster. https://github.com/jadidbourbaki/llama.cpp/pull/12

I’ll benchmark his change and add it to the article, crediting him for this improvement.

]]>Sun, 27 Sep 2026 00:20:34 +0000https://news.ycombinator.com/item?id=49861956jadidbourbakihttps://news.ycombinator.com/item?id=49861956https://news.ycombinator.com/item?id=49861956<![CDATA[World Models from Scratch Part 1: Tokenizing Super Mario Land]]>Article URL: https://www.youtube.com/watch?v=KyqrUA25424

Comments URL: https://news.ycombinator.com/item?id=49604664

Points: 2

# Comments: 0

]]>Tue, 08 Sep 2026 01:21:43 +0000https://www.youtube.com/watch?v=KyqrUA25424jadidbourbakihttps://news.ycombinator.com/item?id=49604664https://news.ycombinator.com/item?id=49604664<![CDATA[Show HN: Time Wizard]]>How scalably can we cheaply fine-tune small models for well-defined tasks? Frontier models are bad at clock reading. I fine-tuned a tiny 450M model from Liquid AI to match GPT-5.6 Sol and beat Fable 5.1 on this task. Here’s my fine-tuned open model called Time Wizard.

Comments URL: https://news.ycombinator.com/item?id=49570829

Points: 3

# Comments: 0

]]>Fri, 04 Sep 2026 22:10:39 +0000https://huggingface.co/jadidbourbaki/time-wizardjadidbourbakihttps://news.ycombinator.com/item?id=49570829https://news.ycombinator.com/item?id=49570829<![CDATA[Context Compaction Theory]]>Article URL: https://arxiv.org/abs/2608.01326

Comments URL: https://news.ycombinator.com/item?id=49164228

Points: 1

# Comments: 0

]]>Tue, 04 Aug 2026 03:58:14 +0000https://arxiv.org/abs/2608.01326jadidbourbakihttps://news.ycombinator.com/item?id=49164228https://news.ycombinator.com/item?id=49164228