Odun.

Research Scientist @ TARESCO : @taresco_hq
LLMs @ Amazon

Blog: this twitter thing again ..........

[Odun. @ToluClassics]

Just published a new blogpost on the @taresco_hq blog on different encoding schemes and how they impact tokenization for African Languages. This is the first in a series of articles about encoding and tokenization.

Hope you enjoy reading it!

https://t.co/BJlwb1rvrr

[Odun. @ToluClassics]

?

[KlawWrld @kawhi_ring]

Hot take: the Nuggets will go 16-0 this playoffs

Come back when I’m right

[Damian Lillard @Dame_Lillard]

OMG @CJMcCollum lol

[Jimmy Lin @lintool]

BrowseComp-Plus (just accepted at #ACL2026 BTW 🎉 congrats @zijian42chen @xueguang_ma @ShengyaoZhuang et al.) can be described as a projection of "riddle-style" questions from BrowseComp on the live web to a custom-built static corpus. https://t.co/bjFeN8LuQg

[Awni Hannun @awnihannun]

Great blogpost from @pcuenq on making a new skill + test harness to automate porting new models from Transformers to mlx-lm

[RRinehart @RRinehart2012]

Hall of fame tweet

[Odun. @ToluClassics]

Haha, welcome..

[Yacine Mahdid @yacinelearning]

I've been studying this paradigm for the past few weeks guys and I get this feeling that this is it

[Odun. @ToluClassics]

https://t.co/ufOoXFCRIS

[The African Research Collective @taresco_hq]

We curated a list of high-quality multilingual LLM training datasets containing at least one African language to save you hours of searching! It includes pre-training, post-training data, reasoning, math, and more – all with sources, split sizes and language breakdown

[David Ifeoluwa Adelani 🇳🇬 @davlanade]

Excited to share that two papers led by Kosei Uemura are accepted at #EACL2026! 🎉

MERLIN: Multi-Stage Curriculum Alignment for Multilingual Encoder-LLM Integration in Cross-Lingual Reasoning https://t.co/Tfbayyn8Da

AfriMTEB and AfriE5: https://t.co/LF65rtCt3l

[Wálé Akínfadérìn @WaleAkinfaderin]

Heading to AWS re:Invent in Las Vegas next week! ☁️

I’ll be leading and supporting a few sessions on optimizing open weight models, automated reasoning, building trustworthy AI apps on Bedrock, and Amazon Nova.

Looking forward to catching up. Reach out if you're attending.

[Oluwaseun Ajayi @Oluwatuntise]

@taresco_hq But we didn't stop at numbers.

We show actual model outputs for Machine Translation and Mathematical Reasoning tasks, so you can see exactly where N-ATLaS succeeds and where it struggles, in each language.

4/

[Odun. @ToluClassics]

@Oluwatuntise @taresco_hq Nice Work!

[Oluwaseun Ajayi @Oluwatuntise]

How Much Did N-ATLaS-LLM Move the Needle?

The African Research Collective @taresco_hq evaluated Nigeria's multilingual LLM (N-ATLaS) using AfroBench across Yoruba, Igbo, Hausa & English.

The results? More interesting than you'd expect

Full report: https://t.co/fgCayaVxgd
1/

[GLADIA Research Lab @GladiaLab]

LLMs are injective and invertible.

In our new paper, we show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space.

(1/6)

[Odun. @ToluClassics]

@fjzzq2002 This is lovely work! Imagine a model dropping a picture prod db table because it is missing some fields 💀

[Odun. @ToluClassics]

@trailblazers we need more Hansen photos

[Odun. @ToluClassics]

Pray?

[Hater Report @HaterReport]

If Jabari Smith and Ausar Thompson start fighting, what does Amen Thompson do 😭

[Odun. @ToluClassics]

@davlanade @eaclmeeting @MasakhaneNLP @Africa_NLP Excited!

[Odun. @ToluClassics]

@thyrbysn 👀

[Andrej Karpathy @karpathy]

How to become expert at thing:
1 iteratively take on concrete projects and accomplish them depth wise, learning “on demand” (ie don’t learn bottom up breadth wise)
2 teach/summarize everything you learn in your own words
3 only compare yourself to younger you, never to others