Toptal

Machine Learning engineer / Data Scientist

Braintrust

Machine Learning Engineer | Founding Talent Member

EasyHealth

Machine Learning Engineer

Cogsy

Time series ML engineer

Speedy Mentors

Mentor

Fifteenkay

Machine Learning Consultant

Speak

Data Engineer

Lola Market

- Developed, deployed, and maintained a Machine Learning model to improve the efficiency of the shoppers' fleet.
- Bootstrapped the first data-warehouse and reporting layer in the company (Amazon Redshift + Amazon DMS + Tableau).

Iridium Advisors

- Statistical analysis of financial markets using explainable Machine Learning techniques.
- Benchmarked a range of ML solutions to enhance model accuracy and explainability.
- Wrote a Python package to ensure the in-house reproducibility of each step of the analysis (data processing, data validation, data visualization, model construction, model validation, and model explanation).

cyngn

- Developed and maintained the analytics dashboards for Cyngn's engineering team and Cyngn's clients.
- Developed components of the internal ETL tool in Python and SQL.
- Assisted the back-end engineer to integrate both frontend and backend of the analytics stack inside Amazon Redshift.

Data scientist

Author of the Hands-on Reinforcement Learning course 🚀

Data scientist and ML engineer with over 8 years of experience. I have a passion for building ML-based solutions, from development to deployment. I love transforming an idea into a model, and a model into an API/product.

I have worked on different problems: financial derivative pricing, digital marketing analytics, deep learning for art generation, or demand prediction for on-line shopping. I have a background in pure mathematics and strong coding skills in Python.

RL is back, baby! Here's a course to get you started
  
  
𝗖𝗼𝗻𝘁𝗲𝘅𝘁

Reinforcement Learning (RL) is one of the key ingredients to post-train Language Models for long, complex tasks, hard-for-humans tasks, especially with verifiable rewards. For example, maths or coding, which is what DeepSeekMath excels at. 

RL is no new kid in the block. RL was born in the 1980s, and over time merged with other very important branches in the history of AI, like Deep Learning. This is what we nowadays call Deep RL.

Google DeepMind and OpenAI did heavy research in Deep RL in the context of game environments like Atari games and board games, creating mega-popular models like AlphaGo and its next evolution AlphaZero, who mastered the most popular heavy-reasoning board games (chess, shogi, go) by playing against itself.


𝗪𝗵𝘆 𝗶𝘀 𝘁𝗵𝗶𝘀 𝗿𝗲𝗹𝗲𝘃𝗮𝗻𝘁 𝗶𝗻 𝘁𝗼𝗱𝗮𝘆'𝘀 𝗔𝗜 𝘄𝗼𝗿𝗹𝗱?

Language Models generate responses token by token, receiving a reward at the end of their text completion. If you frame this as a RL (why not?) you can start using all the powerful algorithms that DeepMind, OpenAI and other researchers developed 10 years ago.

The fundamentals of RL have not changed an inch. What has changed are the RL algorithms used for LMs (like GRPO) which are slight modifications of existing algos developed almost 10 years ago (like PPO).

If you master the fundamentals, the path ahead is clear.


𝗪𝗵𝗮𝘁'𝘀 𝗻𝗲𝘅𝘁?

I am preparing a hands-on tutorial on GRPO for post-training a VLM to do browser automation in collaboration with Hugging Face.

It will be fun, but I want you to be prepared. Hence, I am sharing with you a previous hands-on course I created, that teaches the fundamentals of RL.

This is the first open-source I created years ago.
My older son  :-)

Enjoy ↓
https://lnkd.in/dwKdDUH5

Pau Labarta Bajo

Google DeepMind

OpenAI

Hugging Face

Building and teaching AI that works ​> Maths Olympian​> Father of 1.. sorry 2 kids

Oscar S.

AI is not a bubble.
$500B investments in data centers for centralized AI is THE bubble 🫧
  
  
Why?
Because of 3 reasons:

𝗥𝗲𝗮𝘀𝗼𝗻 1️⃣ -> Demand for AI != Frontier models

88% of ChatGPT queries are solvable by small local models (less than 20B parameters).

> Write an essay
> Summarize this email
> Extract meeting notes
> ...

Most AI demand can be accurately covered with smaller specialized models. No need to use frontier model for most day-to-day tasks.

And the thing is, this percentage will just keep on growing because of reasons 2 and 3.


𝗥𝗲𝗮𝘀𝗼𝗻 2️⃣ -> Smaller models are getting better year after year

Local LMs (≤20B active parameters) such as

> LFM2 by Liquid AI
> Gemma by Google
> Qwen3 4B-14B by Alibaba Cloud
> Granite 4.0 1B-7B by IBM

are improving quickly, taking on an expanding range of tasks once limited to frontier‑scale models.


𝗥𝗲𝗮𝘀𝗼𝗻 3️⃣ -> Local accelerators are getting better year after year

> Smartphone NPUs can run 7B model parameters at over 8 tokens/second

> Laptop accelerators by AMD , Apple or NVIDIA with unified memory above 128GB can run models approaching 200B parameters

> Desktop GPUs enable 200B parameter models.

----
These 3 reasons are clearly exposed and beautifully explained by Hazy Research at Stanford University.

Here is the link to the blog post. Highly recommended read :-)
↳

hazyresearch.stanford.edu

Intelligence Per Watt: A Study of Local Intelligence Efficiency

Liquid AI

Google

Alibaba Cloud

Apple

NVIDIA

Stanford University

Andrew Batutin

Tip for AI engineers 💡

If you want to reduce your cloud bills by 80% you can replace your god-like model with a router + SLM where

- 80% of the tasks are handled by a small language model
- 20% of the tasks are sent to your large/powerful/expensive model.

You are welcome

Are you interested in building small-models that outperform GPT-4/5 for your specific use case? Continue reading...
  
  
At Liquid AI we built a web UI that lets you iteratively improve an SLM in minutes and we’re running a 36-hour sprint to collect real use-cases.

You can either
> come in person to our SF office
> do it remotely.

You get: 
> a model customized to your workflow 
> direct support from our team 
> access to other builders + food 
> we’ll feature the best tuned models

If you're interested, send me a private message and I’ll send the link + get you onboarded.

Bright Nyadzro

The future of agentic AI is hybrid and smaller models are the ones calling the shots ↓
  
  
A couple of weeks ago NVIDIA released Orchestrator-8B

Orchestrator-8B is a state-of-the-art 8B parameter orchestration model designed to solve complex, multi-turn agentic tasks by coordinating a diverse set of expert models and tools.

In other words, this model helps you optimally combine:

> Basic tools, like internet search or a sandboxed code interpreter
> Specialized LLMs, like Qwen2.5-Math-7B
> Frontier LLMs like GPT-5, Claude Opus 4.1 and Gemini

to solve hard problems, while optimizing for speed, accuracy, AND cost.


𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁𝘀? 🤯

On the world's hardest benchmarks:
> 37.1% on Humanity's Last Exam (vs GPT-5's 23.4%)
> 76.3% on FRAMES (vs GPT-5's 66.3%)
> 80.2% on τ²-Bench (beating all SOTA models)

At a fraction of the cost:
> $9.2 average cost (vs $30.2 for GPT-5, $52.5 for Claude Opus 4.1)
> 8.2 latency score (vs 19.8 for GPT-5, 25.6 for Claude Opus 4.1)


𝗪𝗵𝘆 𝗱𝗼𝗲𝘀 𝘁𝗵𝗶𝘀 𝘄𝗼𝗿𝗸?

The counterintuitive magic: small models excel at orchestration precisely because they're unburdened by excessive knowledge and trained to capture the essence of problem-solving.

They act as intelligent coordinators by calling the right expert at the right time, rather than trying to do everything themselves.

This is trained using ToolOrchestra, a method combining:

-> Synthetic data generation
-> Multi-objective reinforcement learning
-> Cost-aware optimization


𝗧𝗵𝗲 𝗯𝗶𝗴𝗴𝗲𝗿 𝗽𝗶𝗰𝘁𝘂𝗿𝗲

This represents a shift from monolithic LLMs to compound AI systems, where specialized models work together under smart coordination.
The result? Better performance, lower costs, and the scalability needed for real-world agentic AI.

----
Follow me for more high-signal (and less BS) on your feed.

Small Language Models aren't just for edge devices. They're silently cutting cloud costs by 80%+ for production workloads. Here's the math 🧵👇
  
    
Most people think: Small models = edge deployment only.

But here's what they miss: Cloud inference costs are DOMINATED by compute time, not storage.

A 1B param model can be 10-20x cheaper to run than a 70B model on the same hardware.


𝗘𝘅𝗮𝗺𝗽𝗹𝗲

You need to process 10M API calls/day for classification or simple reasoning tasks.

70B model: ~$X per 1K tokens
3B model: ~$X/10 per 1K tokens

That's potentially $XXXk saved per month. 💰

The secret sauce? Task-specific optimization.
You don't need GPT-5 level intelligence for 80% of the problems:

> Content moderation
> Simple Q&A
> Data extraction
> Classification tasks
> Routing queries

A fine-tuned 3B model often outperforms a generic 70B here.


𝗣𝗿𝗼 𝘁𝗶𝗽: 𝗨𝘀𝗲 𝗮 𝗿𝗼𝘂𝘁𝗲𝗿 𝗽𝗮𝘁𝘁𝗲𝗿𝗻

→ Small model handles 80% of simple queries
→ Large model only for complex reasoning
This gives you the best of both worlds while slashing costs.


TL;DR: Stop thinking of small models as "just for mobile."

They're your secret weapon for profitable AI products in the cloud.

Start with the smallest model that can do the job. Scale intelligence only when needed.

You are welcome.

Are you someone who checks loss curves before checking your phone in the morning?
  
  
Then Liquid AI Labs is your place.

We're hiring ↓

Open article: Liquid AI Jobs by jobs.ashbyhq.com

If you spent the week on Saturn and just got back to Earth, this is the technical report by Liquid AI that explains how efficient AI models are built.

You are welcome.

Open article: LFM2 Technical Report by arxiv.org

If you spent the week on Mars and just got back to Earth, this is the blog post you should start reading by the Hugging Face team ↓↓↓

Open article: Evaluation Guidebook - a Hugging Face Space by OpenEvals by huggingface.co

Evaluation Guidebook - a Hugging Face Space by OpenEvals

I fine-tuned a Visual LM for a complex image classification task.
Here's what I learned ↓
  
  
𝗪𝗵𝗮𝘁 𝘄𝗮𝘀 𝘁𝗵𝗲 𝗰𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲?
Two weeks ago I fine-tuned LFM2-VL-450M by Liquid AI to perfectly identify images of cats vs dogs. That was my first victory, but it felt too easy.

Hence, on Friday last week I decided to solve a problem closer to the real-world.

I chose the Stanford Cars dataset, which contains roughly 7,000 images of cars with their exact car maker, model and year.

Let me show you what I did


𝗦𝘁𝗲𝗽 𝟭 -> 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗲 𝘁𝗵𝗲 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗮𝗻𝗱 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻 𝗱𝗮𝘁𝗮𝘀𝗲𝘁𝘀
To simplify things a bit, I stripped only the car maker, so the problem to solve was a 49-class classification problem from images.

I split the data into disjoint training and evaluation datasets to perform fine-tuning without data contamination.


𝗦𝘁𝗲𝗽 𝟮 -> 𝗘𝘀𝘁𝗮𝗯𝗹𝗶𝘀𝗵𝗶𝗻𝗴 𝗮 𝗯𝗮𝘀𝗲𝗹𝗶𝗻𝗲 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆
I tested the accuracy of 3 Visual Language models for this task, without any fine-tuning. 

> LFM2-VL-450M -> 60%
> LFM2-VL-1.6B  -> 72%
> LFM2-VL-3B -> 80%

The latest 3B parameter model released by Liquid AI is a small beast, that already sets a high-bar on this difficult task.


𝗦𝘁𝗲𝗽 𝟯 -> 𝗙𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴 𝘁𝗼 𝘀𝗾𝘂𝗲𝗲𝘇𝗲 𝗮𝗹𝗹 𝘁𝗵𝗲 𝗷𝘂𝗶𝗰𝗲
Using LoRA for efficient fine-tuning, I trained across 500 steps and tracked intermediate checkpoints.

Observations:

> The train loss curve for each model saturates at very different levels (see image attached to this post). Both the 450M and 1.6B models struggle to fit on the training data, while the 3B seems to manage it well. Based on this, I picked the 3B for further analysis.

> The eval loss curve of the 3B model is decreasing up to 500 steps, meaning not overfitting on the training data. Good.

> I saved checkpoints every 100 steps so I could run evaluations on intermediate checkpoints to better understand the model’s performance at different stages of the training process.

𝗥𝗲𝘀𝘂𝗹𝘁𝘀
- After 100 steps -> 𝟱𝟮% accuracy, worse than baseline.
This is because the LoRA parameters are still to noisy and damage the model performance.

- After 200 steps -> 𝟴𝟰% accuracy, surpassing the original.
The model is already "in the money" and performs better than the baseline.

- At 500 steps -> 𝟴𝟲% accuracy, approaching 90%.

I did not have time to fine-tune further, but my intuition is that as I cover a whole epoch on the training dataset, I will start to overfit, and the evaluation accuracy will not get close to what I want (aka 99.9% accuracy)

𝗦𝗼, 𝘄𝗵𝗮𝘁'𝘀 𝗻𝗲𝘅𝘁?
I will continue the training process, to confirm my intuition and (if I have time) I will try to do some data engineering to get more samples per class and improve the model’s accuracy.

Let’s keep on VLM-fine tuning, and follow me to get more hands-on content like this.

All code is open-source on GitHub ⭐ You can find the link in the comments ⬇️⬇️⬇️

Open Science is what this planet needs 🌎🌍🌏
Thank you Liquid AI for doing this ↓

Today we introduce Liquid Labs, our advanced research unit, with the goal of understanding and building efficient and adaptive intelligence systems.

Liquid Labs consolidates our existing research efforts at Liquid across architecture of foundation models, multimodality, training, data, and inference. 

The lab also will be home to new frontier research work across the broad range of foundation model build-up stack.

Read the full announcement: https://lnkd.in/ezChkkGD

We are hiring: https://lnkd.in/eYKRC5sH

Also find us at NeurIPS 2025 exhibition hall! 🚀

The most underrated paper of the year
arxiv.org/pdf/2511.21689

Open article: ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration by arxiv.org

ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration

Mathematical certainty, not institutional trust.

I work at the intersection of AI and cryptography.

Language barriers affect almost everything in international business.. but they are about to disappear. 

In 2026 there will be zero excuses why every single doc in your org is not instantly accessible by everyone in your org (regardless of language)!

Here's one example from my past. A problem killing many Japan startups quietly:  In Japan, only ~15% of people read English proficiently. 
Globally? About 20%.

Yet 95% of cutting-edge AI research, technical documentation, and developer resources are published in English first.

This creates an invisible ceiling. 

Japanese startups can't access the knowledge they need if they don't also have English proficient talent.. Found the best dev? Probably not going to have the best english. Getting the best dev who is also fluent!? Super hard. 

I know I have run startups in Japan. It's a vicious cycle. 

But what if language barriers just... disappeared?

I've been experimenting with Liquid AI 's LFM2-ColBERT-350M — a late interaction retriever that does something super cool:

Upload docs in English → Query in Japanese (or any of 8+ languages) → Get accurate results. 

Any knowledge base can instantly be accessible to your JP team — at basically no cost!

Your Japanese engineers can now search English technical docs in their native language. Your French team can query your English knowledge base. 

The research gap closes. But there's a trust problem. Does every company need to in-house index all potentially valuable docs into their own vector DB? Or can we trust one another and share?

RAG systems are only as good as the data you feed them. How do you prove your knowledge base isn't poisoned by a competitor, outdated info, or manipulated content?

This is where zkML comes in.

Zero-knowledge machine learning lets you cryptographically prove: 
✓ Only verified sources entered the database 
✓ No poisoned data contaminated your embeddings
✓ The retrieval model ran correctly without tampering 
✓ All without revealing the actual data

Imagine: A Japanese developer queries your system in Japanese, retrieves English documentation, and can cryptographically verify that every piece of information came from legitimate, unaltered sources.

The implications:
For regulated industries (healthcare, finance, legal) → Provable compliance across languages.

For open-source communities → Verifiable knowledge bases anyone can trust. 

For global teams → One source of truth, accessible in any language, with mathematical guarantees.

We're not just breaking down language barriers. We're building trustless, multilingual knowledge systems that scale globally.

The fabled Tower of Babel scattered us with confusion...
Modern tech is bringing us back together w/ cryptographic receipts.

Check out JOLT Atlas! This plus LFM2-ColBERT-350M is used in our PoC to do  RAG that can be used across orgs without trust!

https://lnkd.in/gticfE8e

#zkML #AI #RAG #MultilingualAI #TechInnovation #StartupChallenges #ZeroKnowledge #LiquidAI

Open article: GitHub - ICME-Lab/jolt-atlas: Fast zkVM born at a16z Crypto substantially adapted by ICME Labs (NovaNet) for verifiable machine learning. ⚡ by github.com

Fast zkVM born at a16z Crypto substantially adapted by ICME Labs (NovaNet) for verifiable machine learning. ⚡ - ICME-Lab/jolt-atlas

GitHub - ICME-Lab/jolt-atlas: Fast zkVM born at a16z Crypto substantially adapted by ICME Labs (NovaNet) for verifiable machine learning. ⚡

taplio

Pau Labarta Bajo's Linkedin Analytics

Pau Labarta Bajo

Performance ScorePau Labarta Bajo's LinkedIn Performance

What is Pau talking about?

Pau Labarta Bajo's Best Posts (last 30 days)

Ready to grow your LinkedIn presence?

taplio

.css-6ytdur{background-image:linear-gradient(284deg, rgba(39, 43, 72, 1) 0%, rgba(33, 47, 183, 1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Pau Labarta Bajo's Linkedin Analytics

Pau Labarta Bajo

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Pau Labarta Bajo's LinkedIn Performance

What is Pau talking about?

Pau Labarta Bajo's Best Posts .css-1emoo84{font-size:var(--chakra-fontSizes-md);font-weight:400;color:var(--chakra-colors-gray-600);}(last 30 days)

Ready to grow your LinkedIn presence?

Pau Labarta Bajo's Linkedin Analytics

Performance ScorePau Labarta Bajo's LinkedIn Performance

Pau Labarta Bajo's Best Posts (last 30 days)