True Positive Inc.

"This book cleared up a lot of conceptual confusion for me about how Machine Learning actually works—it is a gem of clarity. The worked examples and notebook applications gave me a solid starting point for exploration. Even if you are not planning a career in machine learning applications, this is a solid foundation for thinking about the capabilities of these unique new tools.” —Vint Cerf, Internet pioneer and Turing Award recipient

"The book is a good start for anyone new to language modeling who aspires to improve on state of the art." —Tomáš Mikolov, the author of word2vec and FastText

Author of The Hundred-Page Language Models Book

“If you intend to use machine learning to solve business problems at scale, I'm delighted you got your hands on this book.”

—Cassie Kozyrkov, Chief Decision Scientist at Google

“Foundational work about the reality of building machine learning models in production. Comes at the right time when companies start to see through the AI hype, realizing it takes a conscious engineering effort and best practices to make machine learning work. Another great book from Andriy!”

—Karolis Urbonas, Head of Machine Learning and Science at Amazon

Released on Amazon, B&N, and Leanpub. Drafts are available on http://www.mlebook.com.

Author of the Machine Learning Engineering book

Peter Norvig, Research Director at Google, co-author of AIMA, the most popular AI textbook in the world: "Burkov has undertaken a very useful but impossibly hard task in reducing all of machine learning to 100 pages. He succeeds well in choosing the topics — both theory and practice — that will be useful to practitioners, and for the reader who understands that this is the first 100 (or actually 150) pages you will read, not the last, provides a solid introduction to the field."

Aurélien Géron, Senior AI Engineer, author of the bestseller Hands-On Machine Learning with Scikit-Learn and TensorFlow: "The breadth of topics the book covers is amazing for just 100 pages (plus few bonus pages!). Burkov doesn't hesitate to go into the math equations: that's one thing that short books usually drop. I really liked how the author explains the core concepts in just a few words. The book can be very useful for newcomers in the field, as well as for old-timers who can gain from such a broad view of the field."

CEO and Author of The Hundred-Page Machine Learning Book

MindsDB

I'm happy to share that I'm joining MindsDB as an Advisor.

MindsDB is focused on fascinating and, in my view, crucial aspect of applied AI: seamlessly integrating machine learning capabilities directly into the data layer.

They empower developers to leverage existing databases as predictive engines, using the familiar language of SQL. This dramatically lowers the barrier to building AI-powered applications, abstracting away much of the traditional MLOps complexity.

As you most likely know if you read this post, making advanced technology accessible and understandable has always been a core theme in my work. The approach Jorge and the team have taken – bringing AI to where the data lives and to the tools developers already master – resonates deeply with me.

I see significant potential in the product and the team's vision to democratize AI adoption within the broader software engineering community.

I will mainly contribute by:

- Helping in defining product features, making sure MindsDB continues to solve real-world developer problems with elegance and power.
- Hosting webinars designed to clearly demonstrate the simplicity and versatility of MindsDB for software engineers.

We'll explore practical applications and, of course, I'll be diving into how it all works "behind the scenes" to provide that deeper understanding.

I believe MindsDB is on a compelling trajectory, and I look forward to being part of their journey and engaging with the community.

Advisor

SwarmOne

SwarmOne is a machine learning platform that automates everything: from hardware choice to training distribution strategy. What's especially cool is that it's using machine learning to make decisions based on how various training projects performed in the past in terms of memory consumption, speed, and cost.

What the user needs to provide is only the model they want to fine-tune (all Hugging Face Hub models are supported), the training and validation datasets, and their preferences in terms of training cost vs. speed. SwarmOne's AI will do the rest.

SwarmOne has an API compatible with high-level APIs of Hugging Face, PyTorch, and TensorFlow, so all the user's training pipeline will remain functional with minimal changes as the two screenshots demonstrate.

In my strong opinion, making data scientists think about GPU models, their specs and costs, and training distribution strategies is a waste of their time. They should work on the business problem they are trying to solve, the dataset and the labeling architecture that would work the best, and the hyperparameters they want to fine-tune. All the rest must be decided automatically behind the scenes.

This is exactly what SwarmOne does. You can use it as a standalone service or you can install it in your corporate environment to keep your data secure and let SwarmOne optimize the use of your hardware.

I'm really excited to join SwarmOne at this precise moment, where there are a lot of GPUs, a lot of models, and a lot of users, but there's a huge disconnect between them.

Reach out to me if you would like to have a demo.

Akinn Group

Akinn is working on creating an AI-based investment banking platform. As it turns out, investment banking is a highly manual process. By leveraging modern advancements in machine learning and natural language processing, there's an opportunity to significantly scale the process, and I'm here to guide them and cut through the hype.

Advisor to the Board

TalentNeuron

Leading the team of Machine Learning developers digging through the gigabytes of talent marketplace data.

Tools and techniques used: machine learning (deep/shallow, LLMs, classification, generative AI, clustering, topic modeling), regular expressions, Python, MySQL, scikit-learn, PyTorch.

Machine Learning Lead

Gartner

Leading a team working on "hacking" text and other unstructured or semi-structured data to extract useful information. Problems successfully solved by me and my team include:

— conversational user interfaces (chatbots);
— parsing of free-form résumé to detect and extract candidate name, employers, skills, certifications, employment history;
— parsing of job descriptions to detect and extract job title, salary, employer name, assign a standard occupation code;
— automated noisy and multi-language data normalization, cleanup, deduplication;
— novel data detection;
— language detection / text segmentation by language;
— sequence labeling (queries, addresses, salaries);
— machine translation;
— Web-data analysis.

Tools and techniques used: machine learning (deep/shallow, classification, clustering, topic modeling, structured output prediction), regular expressions, Java, Python, MySQL, scikit-learn, Keras.

Director of Data Science — Machine Learning Team Leader

Wanted Technologies

Turning terabytes of job announcements from online job boards into structured knowledge: extraction and classification of skills, standardization of job titles, classification of job announcements by an occupational classification, extraction/cleanup of structured data (company names, proper names, salaries, skills, etc), extraction and standardization of postal addresses, fuzzy duplicate identification (announcements, skills).

Tools: linux shell scripting, MySQL, Sphinx, Python (scikit-learn, NLTK), Java.

Data Analyst

Fujitsu Canada

Computer scientist with the focus on applying machine learning and natural language processing in the context of big data; technology intelligence lead.

Technologies: R, Python, Hadoop, HBase.

Applied Scientist — Fujitsu Innovation Center

Working on research contracts for the Defence Research and Development Canada (DRDC).

Research projects mostly turned around the problems of fusion of sensor data and intelligent software assistants.

Technologies: Scilab.

Scientific Consultant

Université Laval

Research in multiagent systems resulted in multiple publications in top international conferences and journals (AAMAS, AAAI, Journal of Algorithms).

Student Researcher

Read my just released 📖 The Hundred-Page Language Models Book: https://thelmbook.com

Read my bestselling 📖 The Hundred-Page Machine Learning Book: https://themlbook.com

Read 📖 Machine Learning Engineering: https://mlebook.com

Subscribe to my weekly ✉️ Artificial Intelligence newsletter: https://www.linkedin.com/newsletters/artificial-intelligence-6598352935271358464/ and to my weekly ✉️ Data Science newsletter: https://www.linkedin.com/newsletters/7102511020270608384/

About me:

Ph.D. in Artificial Intelligence, passionate about data, fluent in English, French, and Russian. Solid scientific programming and team leadership skills, with over 20 years of experience working on various computing projects, including several of my own startups.

More than 15 years of hands-on experience in automated data analysis, machine learning, and natural language processing.

Trained a Transformer from scratch and fine-tuned pretrained transformers for various tasks. Built a robot that crawls the internet, finds websites with business-critical information, and retrieves updated information periodically. Developed an enterprise chatbot that doesn’t hallucinate.

Expert in Python and Java with several years of daily design and development experience in big data contexts.

Specialties: machine learning, natural language processing, conversational interfaces (chatbots), information retrieval.

Andriy Burkov

Please stop posting about TOON. Just because you never heard of CSV doesn't mean others haven't either.

Now that you've just learned about CSV, welcome to computing. The next step is this magical thing that allows you to write text and the computer does stuff, called Terminal.

PhD in AI, author of 📖 The Hundred-Page Language Models Book and 📖 The Hundred-Page Machine Learning Book

This paper shows a 27-million parameter model solving very difficult reasoning puzzles—like complex Sudoku and maze problems—using only 1,000 training examples, while current large language models get 0% accuracy on the same tasks.

Instead of having the model generate chains of intermediate reasoning steps as text tokens (like "first check row 1, then check column 2..."), this approach does all reasoning internally by running two coupled recurrent networks that repeatedly refine their hidden states until reaching a solution—one network updates its state every timestep while the other updates only every T timesteps and resets the first network each time it updates.

The training uses a one-step gradient that only backpropagates through the final states rather than through all timesteps, which keeps memory usage constant and makes training stable even when the model runs hundreds of iterative computation steps per problem.

Read it on ChapterPal by talking to it: https://lnkd.in/emT8P5yZ

Download a PDF: https://lnkd.in/etDRGA9B

This paper really is groundbreaking. It solves a long-standing embarrassment in machine learning: despite all the hype around deep learning, traditional tree-based methods (XGBoost, CatBoost, random forests, etc) have dominated tabular data—the most common data format in real-world applications—for two decades. Deep learning conquered images, text, and games, but spreadsheets remained stubbornly resistant.

This paper's (published in Nature by the way) main contribution is a foundation model that finally beats tree-based methods convincingly on small-to-medium datasets, and does so very fast. TabPFN in 2.8 seconds outperforms CatBoost tuned for 4 hours—a 5,000× speedup. That's not incremental; it's a different regime entirely.

The training approach is also fundamentally different. GPT trains on internet text; CLIP trains on image-caption pairs. TabPFN trains on entirely synthetic data—over 100 million artificial datasets generated from causal graphs.

TabPFN generates training data by randomly constructing directed acyclic graphs where each edge applies a random transformation (using neural networks, decision trees, discretization, or noise), then pushes random noise through the root nodes and lets it propagate through the graph—the intermediate values at various nodes become features, one becomes the target, and post-processing adds realistic messiness like missing values and outliers. By training on millions of these synthetic datasets with very different structures, the model learns general prediction strategies without ever seeing real data.

The inference mechanism is also unusual. Rather than finetuning or prompting, TabPFN performs both "training" and prediction in a single forward pass. You feed it your labeled training data and unlabeled test points together, and it outputs predictions immediately. There's no gradient descent at inference time—the model has learned how to learn from examples during pretraining.

The architecture respects tabular structure with two-way attention (across features within a row, then across samples within a column), unlike standard transformers that treat everything as a flat sequence.

So, the transformer has basically learned to do supervised learning.

Talk to the paper on ChapterPal: https://lnkd.in/e_rdrePu

Download the PDF: https://lnkd.in/ekrDRXZY

Frontier LLMs like GPT-o1 or DeepSeek-R1 have achieved much stronger skill by "reasoning" through problems step-by-step, but their internal design is either unknown to the public or known only on the surface.

If someone wants to actually train a skilled reasoning model, a lot of details would be missing and would have to be reinvented, resulting in wasted time and money.

This paper provides the first comprehensive roadmap that breaks these complex systems into understandable, modular components. The authors provide a unified framework showing how reasoning structures, search strategies, and training methods fit together—plus a working implementation called x1 that lets you experiment with these concepts directly.

Read interactively: https://lnkd.in/eQxCkXRd

Download a PDF: https://lnkd.in/eC_miwAH

In this remarkable paper, a team at DeepMind and Terence Tao addressed a fundamental challenge in mathematics using an AI system called AlphaEvolve. The challenge is the following:

Many practical math problems ask "what's the best possible arrangement?"—like packing the most circles in a square or finding the largest set of points with no isosceles triangles—and answering these requires actually finding specific examples from search spaces so large that systematically exploring them by hand is impractical.

AlphaEvolve doesn't have the LLM directly generate candidate solutions (which would require millions of expensive LLM calls to explore the search space)—instead, it evolves Python programs that search for solutions, so one expensive LLM call generates a search heuristic that then runs cheaply for millions of iterations exploring candidate configurations, and the population of these search programs evolves over time with better ones surviving, essentially doing meta-optimization where the thing being optimized is itself an optimizer.

Across 67 problems spanning combinatorics, geometry, and analysis, AlphaEvolve discovered new constructions improving decades-old results (hexagon packing, Heilbronn triangles, sum-product problems) while matching state-of-the-art in dozens of others.

The paper reports every problem they tried including failures where AlphaEvolve couldn't match known bounds, which makes this the most thorough empirical evaluation of AI for mathematical discovery.

Read and ask questions on ChapterPal: https://lnkd.in/ev4WH-bA

Download the PDF: https://lnkd.in/ecvNZaxy

How to lie with charts? Anthropic knows how.

Thanks to Nano Banana, we can see how Opus 4.5 is truly better compared to previous models.

On the left, the graph Anthropic shared. On the right, the graph they should have shared (produced by Nano Banana from the original graph) if they didn't want to deceive the users.

taplio