Cristian David Rodriguez

Awesome talk today by Mike Gartrell from Rhizome Labs in our Lunch & Learn Europe at Datadog about how to make AI more robust and reliable. At Datadog, our learning-focused culture and growth mindset mean countless opportunities for skill development, professional growth, and personal interests. Our "Lunch & Learn Talks" are internal weekly sessions led by engineers, designers, product managers, and analysts, covering topics like existing and new systems, technical challenges, and ongoing projects.

33

1 commentaire

Meta just released new multilingual MetaCLIP2 models supporting 300+ models 🌎 you can use these models to do image ⇆ text search in databases in any language! 🔍 find in comments the collection of models and a notebook on how to do so 🫡

742

7 commentaires

Exclusive: Datadog alumni and serial entrepreneurs Gabriel-James Safar and Sébastien Deprez have raised a $10m seed round led by General Catalyst for Tsuga, a Paris-based startup building observability for the AI era. Safar and Deprez previously cofounded Madumbo, which was acquired by Datadog in 2018. The pair left Datadog in 2023, both to take a break. “We wanted to think about the meaning of life,” Safar tells me. “In the end, we didn’t find it, so we created another company.” Read more: https://lnkd.in/e-YvuxaD

182

12 commentaires

Yann LeCun Bets Against the Hype: AI Needs Brains, Not Just Words When the industry obsesses over scaling LLMs, Yann LeCun’s new venture is a much-needed reality check: true intelligence requires understanding the world — not just language. AMI Labs’ focus on open, European-led development could shift how we approach AI long-term, especially in robotics and planning. This isn’t just technical — it’s a philosophical reset.

1

How much can a freelance Data Scientist / AI Engineer in France make in one year? - Main gig: € 144 k (80%) - Side projects: € 32k (17%) - Online tutoring: € 3.7 k (2%) - Medium articles: € 0.5 k (0%) Total: € 180k (net, after taxes) A few takeaways: - If you are an employee, consider becoming a freelance, you can easily x2 your salary. - If you’re in the US, this might sound low. In France, it puts you well above the income threshold for the top 1%. - Side projects alone can get you into the top 10%. That’s not “extra money”, that’s meaningful financial hedge. - If you're hiring in the US, consider hiring abroad: you can get good, cheap talent in EU (and in LATAM, of course). - There’s an obvious 80/20 split here. But this isn’t a “focus on the 80% and ignore the rest” story. The 20% can keep you alive if the main gig disappears. - Writing on Medium doesn’t pay. It’s useful for visibility, not income — which is why I’m moving my articles to my own blog: https://lnkd.in/dn8VJyYD * This is a hypothetical person, obviously. If it were me (it’s not), everything would be properly declared to the tax authorities 😇

45

10 commentaires

Before they were co-founders, they were on opposite sides of a campus network investigation. Olivier Pomel and Alexis Lê-Quôc trace their origin story — from college to IBM, the dot-com bust, and eventually, Datadog — on This Month in Datadog: https://bit.ly/4sckqpw

151

Here is the side story of MOTPE back in 2019 from my perspective... When Optuna was first released in 2018, one of my supervisors (Yoshiaki Bando) told me that a black-box optimization framework was released by Preferred Networks, Inc. Since the company is very famous in Japan, I took a look at the code, understood how the TPE algorithm works, and isolated the implementation. At the same time, Yoshihiko Ozaki, my colleague at National Institute of Advanced Industrial Science and Technology (AIST), was collaborating with Yuta Suzuki and he asked Yoshihiko whether we can extend Optuna to multi-objective optimization. After I heard about the demand, I read the expected hypervolume improvement (EHVI) paper and showed that TPE could also be extended to the same formulation. Yoshihiko and Yuki Tanigaki, a multi-objective optimization researcher who received his PhD at the Hisao Ishibuchi's lab, cocktailed the formulation into multi-objective optimization carefully and extended my isolated implementation. Importantly, keeping the diversity among observations was the hardest part for MOTPE because the TPE algorithm itself tends to search locally and collect similar observations. Yoshihiko and Yuki successfully broke the pathological behavior, leading to its acceptance at GECCO 2020 and the Optuna merge later on. I joined the Optuna Team in 2023, and I significantly sped up the MOTPE implementation; see the article below: - https://lnkd.in/gQVd_7_t Indeed, it was not possible to promote MOTPE to the multi-objective default sampler without the speedup, and I am proud of the improvement I made. When I first heard about Optuna from Yoshiaki, I did not expect it to be so popular. When I first heard about the multi-objective optimization demand from Yuta and Yoshihiko, I did not expect multi-objective TPE to become so influential. But the numbers are real: Optuna is now downloaded 16M times monthly and multi-objective optimization is inarguably one of the most important features in Optuna. The papers themselves are now cited 600 times in total on Google Scholar, including for some important real-world applications. The back story of MOTPE is full of small miracles and it may not have existed without every single piece coming together, but I really appreciate seeing that this method is finally the default algorithm in Optuna.

6