David Herel profile picture

David Herel

Co-Founder @ BottleCap AI
dherel@bottlecapai.com

Hi, I’m David. I’m a co-founder at BottleCap AI, where we work on making large language models dramatically faster and cheaper to train and run.

I also co-founded Hive Spirit Technologies with Tomas Mikolov, building language models for algorithmic trading on stock markets. My research is in language models and LLM reasoning. I’m a co-inventor of Thinking Tokens, the idea behind the thinking mode in today’s chatbots, and have papers at ICLR, ECAI, Amazon Science and ALIFE. I continued this work as a visiting researcher at NUS and was a finalist in Amazon’s Alexa Prize.

Selected experience

  1. 2025–present Co-Founder, BottleCap AI. Founded with Tomas Mikolov and Jaroslav Beck. We raised a $7.5M seed round led by 20VC and are building efficiency-first language models.
  2. 2025 Visiting researcher, National University of Singapore. Worked with Prof. Michael Shieh on Thinking Tokens for LLM reasoning.
  3. 2024–present Co-Founder, Hive Spirit Technologies. Built language-model-driven algorithmic trading systems with Tomas Mikolov for stock markets.
  4. 2022–2025 PhD research, CTU/CIIRC. Research on language models under Tomas Mikolov. First-author papers at ICLR, ECAI and ALIFE.
  5. 2022–2023 Amazon Alexa Prize finalist, Alquist AI. Built open-domain conversational AI for the Socialbot Grand Challenge 5; the team reached the finals and placed 3rd.

Talks, interviews & press

YORD Beyond Reality Podcast thumbnail Can Europe Win the AI Race? YORD / Beyond Reality Podcast · 2026 Long-form conversation on efficient AI, model collapse, model bias, and Europe’s role in frontier AI. BottleCap AI launch press photo BottleCap AI launch Forbes · 2025 Coverage of BottleCap AI’s $7.5M seed round and mission to build more efficient language models. GoodMeetup talk thumbnail GoodMeetup #8: The Full Stack of AI GoodData · 2025 Panel on the modern AI stack with BottleCap AI, Qdrant, Macaly, and GoodData. Moderni Stack talk thumbnail Language models and artificial intelligence Moderní Stack · 2023 Talk on language models, intelligence, and the trajectory of AI systems. AI Association interview photo AI researcher David Herel Czech AI Association · 2023 Interview on language models, research, and how AI is changing software and work. AI explainer media thumbnail Artificial intelligence is already everyday technology ČTK / Forbes · 2024 Expert commentary on how language models are entering everyday technology.

Notable Publications & Releases

ThinkingCap launch graphic
ThinkingCap-Qwen3.6-27B BottleCap AI Open-weights model series, July 2026. Apache 2.0. Qwen3.6-27B with 2× fewer thinking tokens on average and up to 10× faster generation on individual examples, at the same accuracy. Half a million downloads on Hugging Face in its first day; over a million to date across official and community builds.
AI Scan: AI models have hidden preferences. See inside.
AI Scan BottleCap AI Web tool, April 2026. AI contains invisible modifications. AI Scan reveals them: opinion steering, topic avoidance and sensitivity filters, measured on 500,000+ test questions across GPT, Gemini, Grok, DeepSeek and open models.
Diagram from Collapse of self-trained language models
Collapse of self-trained language models David Herel and Tomas Mikolov ICLR 2024 Tiny Papers, Notable.
Alquist 5.0: Dialogue Trees Meet Generative Models. A Novel Approach for Enhancing SocialBot Conversations Ondrej Kobza, Jan Cuhel, Tommaso Gargiani, David Herel, Petr Marek, Jan Pichl, Jakub Konrad, Petr Lorenc and Jan Sedivy Amazon Science 2023.
On Difficulties of Attention Factorization through Shared Memory Uladzislau Yorsh, Martin Holeňa, Ondřej Bojar, David Herel ICLR 2023.
Preserving Semantics in Textual Adversarial Attacks David Herel and Hugo Cisneros and Tomas Mikolov ECAI 2023.
Thinking Tokens for Language Modeling David Herel and Tomas Mikolov AITP 2023.
Emergence of Novelty in Evolutionary Algorithms David Herel and Dominika Zogatova and Matej Kripner and Tomas Mikolov ALIFE 2022.