ML in PL

Tenstorrent builds AI hardware around an architecture designed to scale. Blackhole combines efficient data placement and flow with high-bandwidth on-chip memory, scaling from individual accelerators to Galaxy systems and multi-Galaxy superclusters. Using standard Ethernet rather than proprietary interconnects, the architecture scales under one software model while delivering near-linear performance gains as more compute is added. This lets Tenstorrent support everything from local AI development to high-throughput enterprise inference on the same architecture.

Tenstorrent is one of our Platinum Sponsors at ML in PL Conference 2026. If you're interested in fast inference and performance with no vendor lock-in, they're building something you'll want to hear about.

The transformer architecture has been the foundation of modern AI for nearly a decade. Pathway is working on what comes next.

Pathway is a frontier AI company building post-transformer model architecture. Their Dragon Hatchling (BDH) architecture keeps memory and reasoning inside evolving, fixed-size latent spaces rather than in a growing context window — an approach aimed at systems that can adapt to new environments during inference. Pathway led the development of BDH-CQ, with contributions from researchers at NYU and Bielik AI, demonstrating the efficiency of recurrent latent reasoning and establishing a new Pareto frontier on ARC-AGI-1. Their empirical work across ARC-AGI, Sudoku Extreme, and BABILong explores how this design supports constraint-based reasoning and the retention of relevant information over long contexts.

The team is led by Polish researchers: CEO Zuzanna Stamirowska, a complexity scientist; CTO Jan Chorowski, previously a researcher at Google Brain; and CSO Adrian Kosowski, a computer scientist who completed his PhD at 20. Their advisory board includes Łukasz Kaiser, a co-author of the paper that introduced the transformer architecture.

They're joining us as a Gold Sponsor at ML in PL Conference 2026. If you've been thinking about the limits of attention-based architectures, this is a conversation worth having.

With Polish roots, global ambition, and a rare mix of research depth and product mindset, ElevenLabs has become one of the companies putting Poland on the global AI map. We're proud to welcome them as a Strategic Sponsor of ML in PL Conference 2026.

They started with voice because that is where the gap between AI and human communication was most stark, and because it is a hard technical problem: beyond the words, it requires mastering tone, emotion, and the rhythm of speech, along with all the cultural nuance and back-and-forth of human conversation.

Their research team built the first AI voice model to achieve human-level expression across 70+ languages. The work now spans text-to-speech, speech-to-text, turn-taking, interruption handling, and music. For scale: their text-to-speech models generate 34 years of audio every day, across 90+ languages spoken natively by more than 5.5 billion people.

Two examples of what that looks like in practice, both close to home. Allegro, which serves over 20 million active buyers across Poland and Central Europe, is rolling out a voice agent on its delivery support hotlines, starting with Allegro Delivery and its network of more than 36,000 parcel lockers, Allegro One, and One Box. Instead of one assistant covering everything, the deployment uses a multi-agent architecture, with separate agents handling separate scenario types, which keeps each conversation narrow and accurate; when a call needs a person, the agent collects the tracking number and passes on a summary, so whoever takes over starts with context rather than from zero. In public services, their models power the voice agent for Centralna e-Rejestracja, supported by NFZ, which calls patients with reminders about upcoming appointments and helps them reschedule or cancel. With around 40 million appointments a year in Poland and a no-show rate of 10 to 20 percent, unlocking those slots could meaningfully improve access to care.

Advances in speech, emotion, and latency reach users immediately, and usage data feeds back into research to refine the models. That feedback loop is only possible when research and deployment sit under one roof.

Hubert Siuzdak from their research team is giving a talk on October 10. The rest of the team is at the booth in Zone A across all three days.

If you're deciding where to do research next, they're worth talking to while you're here.

We're excited to welcome Comarch, one of our Platinum sponsors, who will be taking the stage to talk about From Anomaly Detection to Autonomous Action: Bridging ML Analytics and Agentic AI in Sovereign Telecom Networks.

Modern telecommunications networks generate billions of events daily, requiring high-precision analytics and strict data sovereignty. Machine learning forms the foundation of network operations, providing real-time time-series forecasting, anomaly detection, and root cause analysis (RCA) across fault and performance management systems. Building on these core analytical capabilities, the next evolution lies in connecting ML-driven insights with multi-agent orchestration to achieve closed-loop automation.

In this session, we'll present the architecture of the Comarch OSS AI/ML Platform—a hybrid, sovereign ecosystem designed to bridge predictive ML with agentic AI. We'll discuss how production ML pipelines generate the critical operational triggers that feed directly into multi-agent workflows. Using standardized context sharing via the Model Context Protocol (MCP), AI agents interpret ML outputs, reason over network topology, and safely orchestrate remediation paths. Join us to explore how combining deterministic ML precision with agentic AI reasoning establishes a practical foundation for autonomous telecom operations.

The talk will be presented by Arkadiusz Zwierz, Head of the R&D Department at Comarch, alongside Sebastian Zaranek, Software Developer and Architect at Comarch.

Don't miss this deep dive into the future of autonomous telecom networks—see you there!

Being employee number one at Mistral is a good story on its own. What we find more interesting is what came before it: four years at FAIR spent studying how much of a large language model you can remove before it stops working.

Pierre Stock is VP Science, driving Mistral's model releases, roadmap strategy and ex*****on. He holds a Master's degree in Mathematics, Vision, Learning from Paris-Saclay and began his career at Facebook AI Research (FAIR) in 2017. At FAIR, he completed his PhD exploring efficiency and redundancy in LLMs, in collaboration with ENS de Lyon until 2021. He then continued his career at Meta before joining Mistral as the first employee.

Come hear him in Warsaw at the 10th edition of ML in PL Conference, October 8-10, 2026, at the Copernicus Science Center. Tickets and the full lineup are here: https://conference.mlinpl.org

We're excited to welcome JetBrains, one of our Platinum sponsors, who will be taking the stage to talk about So You Want to Train an LLM from Scratch? Lessons from Mellum 2.

Training an LLM from scratch requires a long chain of decisions before instruction tuning ever begins. Using Mellum 2, JetBrains' 12B-parameter Mixture-of-Experts model, as a case study, the speakers will walk through choices around tokenization, training data, and the pre-training recipe, as well as the experiments used to evaluate them. They'll also discuss the engineering trade-offs that arise when turning these choices into a full training run. Attendees will leave with a practical understanding of how pre-training decisions are made, tested, and iterated on in a real model development process.

The talk will be presented by Kseniia Lysaniuk, Machine Learning Engineer at JetBrains, whose work spans model evaluation and training data research, including contributions to Mellum 2.

Joining her is Petr Borovlev, Machine Learning Engineer at JetBrains, working on large language models with a focus on pre-training, architecture, and infrastructure. He contributed to Mellum 2 and its tokenizer, exploring how tokenization choices affect the efficiency and behavior of language models.

A behind-the-scenes look at what it really takes to build an LLM from the ground up, straight from the team that did it.

Regular tickets just closed. From 18:00 CEST today, late registration is the only way in: 1500 PLN, flat rate for everyone.

The 10th edition of ML in PL Conference runs 8–10 October 2026 at the Copernicus Science Centre in Warsaw, with Lenka Zdeborová, Johannes von Oswald, Alexia Jolicoeur-Martineau, Valentina Boeva, Emanuele Rodolà, Razvan Pascanu, Piotr Sankowski, Bartosz Piotrowski and Adam Paszke on the program so far.

Tutorials (11 October, separate ticket) also go on sale today - hands-on, room-capped, usually the first to sell out.

Seats at Copernicus are finite. No guarantees on how long they'll last.
More info in the links below.

Got plans this fall? If you’re a student, researcher, PhD candidate, or a bussiness professional, our friends at .org are hosting three great events in Chicago, Oxford, and Warsaw! 🇺🇸🇬🇧🇵🇱

Here’s where you can meet them:
☕ Mentoring Espresso Chicago (September 19, University of Chicago)
All about commercializing research and turning ideas into business. Listen to expert panels, take part in the small-group mentoring and network with Polish diaspora while sipping a cup of coffee.

☕ Mentoring Espresso Oxford (October 10, St. Vincent Club)
Focusing on career choices, navigating academic and industry moves, and making the most out of mentoring. And again – the best discussions happen over coffee!

🧠 To Poland For Science | Warsaw (October 19, Pałac Staszica)
A one-day conference on mental health and well-being in academia. Researchers and experts will cover brain health, neurobiology of well-being, and practical tools to build resilient research communities.

Spots for the events are limited, so grab yours before they fill up!
Details and registration links: https://poloniumfoundation.org

Warehouse robotics is one of those domains where the gap between research and deployment is immediately visible — literally, in the middle of a fulfillment center running around the clock.

Nomagic is a leading warehouse robotics company, applying breakthrough general-purpose Physical AI to optimize warehouse operations. The company's deployed robots learn from a massive set of real operational data, built over millions of tasks in 24/7 environments, that trains an adaptable Physical AI platform handling a variety of warehouse tasks. Nomagic's next-generation VLA (visual language action) models integrate automatically into the fleet of AI-powered robots, accelerating autonomy and improving efficiency, while setting the industry standard for the fastest deployment time. For more information visit http://nomagic.ai/.

They're joining us as a Silver Sponsor at ML in PL Conference 2026. If you work on physical AI or robotics, you'll have things to talk about.

Quantitative investing is, at its core, a data and modeling problem. QRT is a global multi strategy investment manager where we deploy a diverse range of investment strategies across geographies, asset classes and time frames.
Combining data, research, technology and trading expertise has shaped QRT’s collaborative mindset which enables us to solve the most complex challenges.
QRT’s culture of innovation continuously drives our ambition to deliver high-quality returns for our investors.

They're joining us as a Gold Sponsor at ML in PL Conference 2026. If quantitative research or ML in financial markets is anywhere near your interests, they're worth finding at the conference.