Breaking: The language in code can affect how LLMs reason π
Our research with Cohere Labs is live, 51 languages supported, + we're opening our language packs for community review.
Hi y'all! π
It's been a full few months, and we've been holding on to a lot of news waiting for the right moment to share it. That moment is here, so grab your chai, cafΓ©, kahve, or whatever you're sipping today, and let's catch up. β¨
in this issue π¬
- our research with Cohere Labs is live π§
- 51 languages and counting π£οΈ
- the language packs are opening up for review π«
- a full-circle moment with Grayhat + NUST πΈ
our research with Cohere Labs is live π§
Our crew, "Language Decoded", used Legesher language packs to transpile Python into δΈζ (Chinese), EspaΓ±ol (Spanish), + Ψ§Ψ±Ψ―Ω (Urdu). We then fine-tuned a small multilingual model (Cohere's Tiny Aya) on that native-language code datasets.
Writing code in these languages shifted the model's behavior in intriguing ways, often improving its natural language reasoning for that specific target language. The biggest learnings came from the model trained on Ψ§Ψ±Ψ―Ω (Urdu) code, the lowest-resourced language we tested, where the model often produced code leaked within its responses.
As AI tools become central to how the world builds software, the data we train them on shapes who gets to participate. Code in only one language teaches AI to think in only one way.
Huge thank you to Saad Bazaz and Grayhat, who co-captained this effort with me and the entire "Language Decoded" crew: Sohaib Bazaz, Rashik Shahjahan, Khojasteh Mirza, Sarah Jawaid, Rafay Mustafa + Tom Sherborne. π
π Read more on Cohere Lab's blog here
51 languages and counting π£οΈ
the language packs will be open soon to review π«
we don't like keeping secrets, but we've been planning a very special surprise for our community. We're releasing our 51 language packs for the community to review, together.
Each pack is to be viewed as "experimental", helping native speakers imagine a world where code was first built in their language. We don't believe that solely simple translations via ai model or transliteration of a loan word can capture the complexity and depth hidden in a language. Through this collaborative effort, we hope to provide the stage for each language community to define their own voice in code. These language packs will be open sourced to be enjoyed and built upon by the community.
(π€ This is where you come in!) We need native speakers to review each language's pack. We invite you to share with us which terminology feels natural, flag what doesn't, and help refine it for the next user. If your language isn't currently listed as one Legesher supports, please let us know so we can include it in a future release!
If you want to help your language get it right, come build the source of truth with us. π²Join the community on Slack
hacking full circle with Grayhat + NUST πΈ
This spring, we partnered with Grayhat to host and judge the Speed Programming Competition at COMPPEC, the flagship tech event of the National University of Sciences and Technology (NUST) in Pakistan. I (madi) joined virtually to meet the students, introduce Legesher and the story behind our software.
One thing that was sweet to note, was how this experience reminded me of my very first hackathon, which was the inception of Legesher. π₯Ή Years later, in a (virtual) room full of students across the world coding in their language.. definitely felt full-circle.
Thank you Grayhat, Saad Bazaz, + Talal Ahmed Khan for the collaboration, and to NUST + the COMPPEC team for the invitation to share Legesher. As always, thank you to the open source community whose brilliance brings all of this to life. π«Ά
π« that's all for now!
To building in the language you dream in,
Madi β¨ Founder of Legesher
