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 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.

LanceSoft, Inc.

Senior Information Technology Recruiter

ThreatOptix

IT Recruitment and Business Development 

InterQuest Group

Senior Technology Recruiter

Creative Data Resources

Recruited, sourced and managed the pipeline for hundreds of technology candidates over the last 8 years. Full life cycle recruiter with a background in sales. On boarded clients while also dealing directly with HR at fortune 500 firms.

taplio

The Battle

Andriy Burkov

Daniel Rodriguez

Followers

Posts

Engagements

Likes

Performance ScoreAndriy Burkov's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Andriy Burkov

Daniel Rodriguez

Followers

Posts

Engagements

Likes

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);}Andriy Burkov's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreAndriy Burkov's LinkedIn Performance