Suggestions
Emmanuel Turlay
Airtrain AI – The key to better AI is better data | Founder/CEO (YC S22) | ex-Cruise, Instacart, CERN
Emmanuel Turlay is the Founder and CEO of Airtrain AI, a company that specializes in no-code data curation for Large Language Model (LLM) fine-tuning and evaluation.2 Here's a comprehensive overview of his background and career:
Education and Early Career
Emmanuel holds a PhD in Physics and began his career in academia, spending four years as a researcher at CERN (European Organization for Nuclear Research).1 During this time, he worked on the ATLAS experiment and gained expertise in programming languages such as C, C++, Python, and Bash.1
Tech Industry Experience
After transitioning from academia to the tech industry, Emmanuel has held several notable positions:
Instacart:
- Worked at Instacart for about 3.5 years, progressing from Senior Software Engineer to Technical Lead Manager1
- Built the order and payment infrastructure for the company2
Cruise:
- Spent nearly 4 years at Cruise, advancing from Senior Software Engineer to Senior Staff Software Engineer1
- Led the Machine Learning Platform team, focusing on orchestration and data processing1
Entrepreneurial Ventures
Sematic (YC S22):
- Founded Sematic in April 20221
- Sematic is an open-source Continuous Machine Learning platform built on learnings from the robotaxi industry1
Airtrain AI:
- Founded Airtrain AI in 20222
- The company is based in San Francisco and has 5 employees2
- Airtrain AI offers a no-code data platform for Large Language Models, helping businesses scale their AI prototypes into production-grade products2
Expertise and Focus
Emmanuel specializes in:
- Machine Learning platforms and pipelines
- AI and Large Language Models
- Software engineering, particularly in back-end development
Recent Developments
In late 2023, Emmanuel announced the launch of Airtrain.ai, a free AI-Assisted Batch Evaluation Tool for Large Language Models.3 This tool allows users to compare and evaluate different AI models across various properties such as groundedness, politeness, and creativity.3
Emmanuel is active in the AI community, frequently sharing insights and updates about Airtrain AI's progress on social media platforms like LinkedIn and Twitter.43
Highlights
"The Price-Intelligence Frontier has advanced by another order of magnitude in another quarter!" – @swyx
The price of AI inference is essentially going to zero. Models are getting smaller (GPT-4o mini, Gemma 2, Gemini 1.5 Flask) and cheaper. In a few months, inference cost for most use cases will be immaterial. It's a race to the bottom. I would be worried for any business built on metered billing.
Is this sustainable? Those models do need expensive hardware to train and run. Training and inference are indeed becoming more efficient, but it seems all these price cuts are fueled by two things, one sustainable, the other not:
- Sustainable: economy of scale. The more volume you have to support, the more discount you can get from GPU providers.
- Unsustainable: VC funding. I am not worried about Google slashing its inference cost, they can take it. But smaller providers such as Together AI or Lepton (just examples, not singling them out) are essentially burning VC money to stay relevant in the inference price wars.
There is a future in which all OSS models are available for a very low cost, fully hosted by cloud providers (à la AWS Bedrock), or even baked into rented VM chips, with a handful of proprietary providers (Open AI, Anthropic, Google) are fighting for proprietary ASI.
How is Gemma2 2B so good yet so small? Distillation. What is that?
With only 2B, it already beats Llama 2 70B (35x bigger) and GPT 3.5 (likely 100x bigger) on the LMSYS Chatbot Arena leaderboard.
How can such a small model perform as well? It was distilled from a larger (unnamed) LLM. What does that mean? The larger model (the teacher) was used to label certain topical datasets on which the student model was trained.
This ensures that the student model picks up high quality skills in specific domains. For example, Gemma 2 2B is very good at instruction following (IFEVal) and understanding (MMLU-Pro), but not as good at math.
A very powerful and cost-effective method to get small models (and therefore cheap and fast to run) to compete with larger ones on specific skill sets.
