Explosion

Looking forward to my keynote at PyCon FR in Strasbourg in November 🐍💚 I'll talk about lessons learned from solving real-world information extraction problems in industry and a new approach and mindset for designing robust and modular #NLP pipelines in the age of #GenAI.

Wanted to highlight today that ZenML now integrates to 3 more data annotation tools: ➡ Argilla (Shout-out our good friends Daniel Vila Suero and Gabriel Martín Blázquez) ➡ Prodigy by Explosion (shout-out Ines Montani, last seen around these parts on the ZenML podcast) ➡ Pidgeon by Anastasis Germanidis - Really cool for notebook annotation workflows People forget that annotation is a really important part of the MLOps lifecycle and it's even more so with LLMs. Argilla's distilabel e.g. builds in a close feedback loop for the development around LLMs, as does Prodigy's spacy-llm package etc. With these integrations, we're not only enabling people to bring these annotators into their normal workflows but to connect to their LLM work to unlock all fine-tuning LLMs use cases. HUGE kudos to Alex S. who integrated ALL of these tools 🤯 Alex and Alexej Penner showed off one of the annotator integrations on the recent Computer Vision Webinar (Now uploaded on YouTube: Link in comments)

Thanks for a great PyCon Italia – so happy I finally made it over and it's easily in my top 3 of PyCons so far! Some highlights: 👩🏫 great talks and keynotes – including Jodie Burchell, PhD on LLMs and the illusion of humanity, and Vicki Boykis (finally met in person, yay!) on how to deal with distractions and complexity by staying close to the metal 🍍 yet another entertaining lightning talks session with the pizza wheel of fortune, a Lego building contest and other fun social activities 🤝 meeting so many cool developers (and spaCy users!) and discussing real-world NLP use cases 🔥 catching up with Samuel Colvin and looking forward to giving Pydantic Logfire a spin 💕 hanging out with new and old conference friends, including Jodie Burchell, PhD, Cheuk Ting Ho, Alexander C.S. Hendorf 👋, Martin Borus, Laís Carvalho and many more 🍦 sightseeing, shopping and lots of eating in Florence Here are the slides from my talk: https://lnkd.in/eyjRWTxZ

We're coming to India again! Very happy to be giving a keynote at DataHack Summit by Analytics Vidhya in Bengaluru in August 🚀 I'll be speaking about the new approach and mindset for designing robust and modular NLP pipelines in the age of Generative AI, and Matthew Honnibal's talk will go deeper into practical implementations. Hope to see you there – let me know if you're attending! You can also still submit a talk for the event 👇

One of the best talks I've seen. Technology, vision and inspiration. Thank you Ines Montani ! Pictured are two happy people together with the Queen! thanks PyCon Italia

It's easy to get fixated on full automation as a technologist, but if we're looking to improve or replace some business process with a new system, we should be asking "What's the real problem with the current workflow?" When I was a student, the problem I thought we'd be solving was cost. Isn't that what business is all about, reducing cost? But it's not. In most business business processes, cost is much less important than reliability. Imagine you have some machine that makes you $1000 a day. You decide to make another one, and you're out shopping for parts. For one component, there's two options: one for $0.05, and another for $1.00. Why would you get the one that costs 20x more? Well, the $0.05 component has a 1% failure rate. On 1% of days it'll break and your machine won't print any money. The $1.00 component has a 0.1% failure rate. So you can expect $365 of revenue disruption per year with the expensive component, and $3650 of revenue disruption per year from the cheap one. The difference in direct costs here is completely dwarfed. Often it's actually the reliability requirements that make business processes costly in the first place. If you have some work to be done at scale and you need it to be done right, you often can't get just one human to look at it. People make mistakes, especially if they're doing repetitive work in large quantities. So often businesses pay to have the work done in duplicate or triplicate. This is where it's easiest to deploy AI and get great value from it. AI systems definitely make mistakes, but the great thing about them is that they generally make different mistakes from humans. Instead of having two humans do the work, if you have one human and one automated system, you can get better reliability and lower cost. The AI system will pick up the careless errors the human makes, and you can review cases where the two disagree and even improve the model over time.

As a teenager I went to concerts and festivals all the time (name a 2000s indie band and I've probably seen them!) and even ran a mini online music magazine with interviews and reviews. Never thought this world would collide with my tech job but here we are 🎸 On June 1 I'm moderating the AI panel at immergutrocken e.V. Festival! https://lnkd.in/ehGq8J2A Together with Dr. Richard Sieg, Alexander C.S. Hendorf 👋, Dr. Maria Börner and David Bay, we'll discuss the latest developments in AI, hype vs. reality and what those new technologies mean for people, businesses, art, creativity and the music industry.

In Florence for PyCon Italia and enjoying my first time in Bella Italia ☀️ Staying at a cute hotel (with hotel cat!), already had pizza and made it to the conference early today to catch Jodie Burchell, PhD's keynote. My talk is tomorrow at 11am and looking forward to meeting everyone 👋 Live Stream: https://lnkd.in/e6MYr--g

🎤We are glad to welcome Ines Montani from Explosion as a speaker! Her talk “Taking LLMs out of the black box: A practical guide to human-in-the-loop distillation“ will be part of the in-person Data Science track on June 11. 👉 Find out more about the Budapest Data+ML Forum: https://lnkd.in/dvpGuuta #budapestml24 #speaker

My favourite NLP educator and calm coder 👋 Vincent Warmerdam and my favourite podcast host Michael Kennedy have teamed up for the latest Talk Python course 👨🏫✨ 2+ hours of hands-on content covering everything you need to know about NLP and spaCy, from preprocessing and training all the way to LLMs! https://lnkd.in/eD7mQfCP #nlp #llms