Yoann Oza

Just a curious guy, passionate about AI 🤖✨

Breaking things — fixing them — and learning a little more every time.

[Yoann Oza @yoannoza]

@FishAudio Fish

[Fish Audio @FishAudio]

Today we’ve raised $52M Seed and we are announcing the public launch of S2.1 Pro.

>It can clone a voice from 5 seconds of audio
>2x faster than Cartesia & 1/6th the cost of Eleven Labs
>most expressive model with word level control over emotion, intonation, pacing etc

We support frontier AI companies including HeyGen, LiveKit, Retell, Sanas, and OpenArt all run our model in production.

If you're a business and we can't cut your voice AI costs by 50%, we'll give you 1 year of Fish Audio for free.
Book a demo: https://t.co/vHkyZf9JoG

To celebrate our first birthday, we'll give you 1 month of S2.1 Pro for free. Like, retweet, and comment “Fish” to get it.

[AprenX @AprenXSchool]

𝐋'𝐈𝐀 𝐞𝐬𝐭 𝐥𝐞 𝐧𝐨𝐮𝐯𝐞𝐚𝐮 𝐬𝐭𝐚𝐧𝐝𝐚𝐫𝐝 𝐝𝐞 𝐩𝐫𝐨𝐝𝐮𝐜𝐭𝐢𝐯𝐢𝐭é.

RH, chefs de projet, commerciaux ou entrepreneurs : l’IA permet d'éliminer les tâches répétitives qui s'accumulent dans vos journées.

Envie de passer au concret ?
https://t.co/Up4HOiw9IV

[AprenX @AprenXSchool]

𝐀𝐧𝐭𝐡𝐫𝐨𝐩𝐢𝐜 𝐦𝐨𝐧𝐭𝐫𝐞 𝐜𝐨𝐦𝐦𝐞𝐧𝐭 𝐂𝐥𝐚𝐮𝐝𝐞 𝐬’𝐢𝐧𝐭è𝐠𝐫𝐞 𝐚𝐮𝐱 é𝐪𝐮𝐢𝐩𝐞𝐬 𝐣𝐮𝐫𝐢𝐝𝐢𝐪𝐮𝐞𝐬 𝐚𝐯𝐞𝐜 "𝐂𝐥𝐚𝐮𝐝𝐞 𝐟𝐨𝐫 𝐋𝐞𝐠𝐚𝐥 𝐭𝐞𝐚𝐦".

#AI #LegalTech #ClaudeAI #Anthropic #Law #Automation #LLM #Innovation #AprenX

[Kevin Degila👨🏾‍💻 @kevindegila]

J'ai fait cette vidéo pour le moi d'il y a quelques années qui avait du mal à comprendre le mécanisme d'Attention dans les Transformers.

Pas de formules intimidantes. Pas de jargon. Juste la logique derrière avec de l'intuition.

Enjoy

https://t.co/42m0ryEqzG

[Karan🧋 @kmeanskaran]

Starting this week, I will begin my LLM infra and inference engineering project.

- Data preparation
- Fine-tuning LLMs with reasoning models
- Quantization
- Inference layer for LLMs
- vLLM serving
- Multi-GPU training
- GPU deployment
- Exposing APIs with auth, just like Claude and GPT
- Chain of thought
- Observability for token usage and the system

I will definitely be vibe coding, but the most important part is designing the infra rather than coding, solving edge cases, and building a complete system.

I did the same with my MLOps project this year, where I posted daily learnings and insights.

This list will be updated as I explore more.

Let's learn inference engineering together🫡

[Yoann Oza @yoannoza]

@paulabartabajo_ 1

[Yoann Oza @yoannoza]

Hey future Yoann,

I just want to remind you that you dream about being this guy who knows, like nobody else, how to push any AI model to prod.

This AI Ops Engineer.
And a sports analytics lover too.

Don't forget it next time
See you later

[Yoann Oza @yoannoza]

Hey future Yoann,

I just want to remind you that you dream about being this guy who knows, like nobody else, how to push any AI model to prod.

This AI Ops Engineer.
And a sports analytics lover too.

Don't forget it next time
See you later

[Raj Shamani @rajshamani]

Distraction is not the problem. Distraction is what happens when the work in front of you stops feeling connected to anything that matters. Fix the meaning, and the focus follows.

[Yoann Oza @yoannoza]

@no_stp_on_snek Amazing work

And why does just choose a draft and target models from the same family isn't the best option ?? Or where is the optimisation opportunity based on this ??

[Tom Turney @no_stp_on_snek]

been digging into dflash (block diffusion for speculative decoding). the headline is ~6x speedup without changing outputs

what’s actually interesting isn’t just “parallel tokens”

it’s removing the sequential drafting bottleneck entirely. instead of predicting tokens one by one, the draft model generates a whole block in a single pass

but the real trick is how it stays accurate

the draft isn’t guessing from scratch. it’s conditioned on internal features from the target model itself. basically borrowing the model’s own representation of what comes next

so instead of a weak drafter trying to imitate the target, it’s more like a fast adapter plugged directly into it

spent the last day trying simpler approaches on qwen-style hybrids. small draft heads, independent parallel prediction, etc

they all run into the same problem:
acceptance rates stay too low, and you lose everything to replay

once you move toward feature-conditioned + joint prediction, things start to click. that seems to be the real unlock here

also been cross-referencing some turboquant+ ideas while building:
some layers carry way more signal than others
attention is effectively sparse at long context
simpler structures often outperform over-engineered ones

working on a variant that leans into those constraints. still early, but this is the first time the path to high acceptance actually looks realistic. TBD.

[Tamebi AI @Tamebiai]

No more VRAM calculations.

Discover tamebi, the CLI that scans your hardware and lists the LLMs compatible with your machine (precision, KV cache, VRAM).

The AI infrastructure for Africa starts here.
pip install tamebi
Docs: https://t.co/whUQsbbdaB

[Bin Liu @liu8in]

https://t.co/eFzPqAOeiB

[Kevin Degila👨🏾‍💻 @kevindegila]

@AnthropicAI Please disable your "Gift Claude" feature immediately. Several people I know have discovered that someone accessed their accounts overnight and gifted themselves a Pro or Max subscription. This needs to be fixed ASAP.

cc @bcherny

https://t.co/JDdOEA7bdX

[Liran Ringel @liranringel]

Introducing DDTree: accelerates speculative decoding by drafting a tree with one block diffusion pass, then verifying multiple likely continuations together.

Paper: https://t.co/cgYBw70O5i
Project page: https://t.co/ygFukxrZLB
Code: https://t.co/2z7U00NsuH

[AprenX @AprenXSchool]

Clap de fin pour la Cohorte 2 AprenX !

On a appris à faire de Claude un vrai assistant : automatisation, analyse Excel express et rédaction.

Prochaine étape le 4 mai : Cohorte 3 (focus écosystème Claude & Cowork). Liste d'attente ouverte (prix réduit)

https://t.co/l5tvy9GuPD

[Tamebi AI @Tamebiai]

Announcing the Tamebi AI Lab: Benin Republic’s 1st team hosting & serving the best open-source AI models for Africa!

Sending data to foreign servers isn't an option. We benchmark & serve models locally at a lower cost.

Read the full announcement here: https://t.co/rsUDe16YUT

[Amin Karbasi @aminkarbasi]

I left @GoogleResearch almost two years ago, so it makes me genuinely happy to see our work on polar quantization (my last project), which eventually led to extreme compression, being recognized there. It is a nice reminder that good fundamental work tends to find its place with time.

[Prince Canuma @Prince_Canuma]

Just implemented Google’s TurboQuant in MLX and the results are wild!

Needle-in-a-haystack using Qwen3.5-35B-A3B across 8.5K, 32.7K, and 64.2K context lengths:

→ 6/6 exact match at every quant level
→ TurboQuant 2.5-bit: 4.9x smaller KV cache
→ TurboQuant 3.5-bit: 3.8x smaller KV cache

The best part: Zero accuracy loss compared to full KV cache.