Marcos Esteve

Applied Scientist @ Thomson Reuters.

[Marcos Esteve @mecprojects]

@Footgolfexpert @CBSNews Same situation

[El Club del Arte 🎨📷📚🖼🕍🎼 @Arteymas_]

Zúrich acaba de inaugurar un túnel de 440 m iluminado con LED para bicicletas, con arte, señalización y estacionamiento para más de 1.200 vehículos, incluso bicicletas eléctricas y ciclomotores, todos a salvo de las inclemencias del tiempo 🚴‍♂️

[Juan Luis Domínguez @JuanLDominguez_]

I've always been amazed by the performance of ms-marco-MiniLM-L-6-v2, even years after its release!
I'm happy to announce MiniLM-L-6-rerank-m3, a small (22M) yet powerful cross-encoder!

Despite its size, it delivers impressive performance across benchmarks.

[Jo Kristian Bergum @jobergum]

This weekend marks a big milestone imho, credible voices promotes evals and baselines in retrieval.

This helps the larger community building LLM apps understand that retrieval is more than encoding text into a single vector representation. In the end, retrieval is search and the natural choice of technology is a search engine like Vespa where you can represent baselines and newer neural techniques.

Why is BM25 still a strong baseline? Because BM25 builds a statistical model of your data, and lets face it, the exact words that the user types are still vital for high-precision search.

Importantly, we are @vespaengine will not hide simpler baselines to sell storage and compute units without significant benefits for the application.

Image from https://t.co/2BYkhUCoVA

[Jo Kristian Bergum @jobergum]

Ask any search specialist about tokenization, especially in multilingual contexts. I’ve also been surprised by search experts thinking that vector search solves it😅

[Jo Kristian Bergum @jobergum]

Matryoshka Representation Learning is trending after @OpenAIDevs announced their variable embedding length!

Did you know that @vespaengine lets you peek into the original vector to perform any computations over k out of m dimensions? This example uses the first 256 out of 3072!

[Jo Kristian Bergum @jobergum]

Good morning

[Shreya Shankar @sh_reya]

📢 ML pipelines should not be Jupyter notebooks

[Marcos Esteve @mecprojects]

@Cristian_Zazo Enhorabuena! Donde te vas?

[Mounia Lalmas @mounialalmas]

My team is looking for research scientists interested in creating impact by developing & applying large foundational models that reshape the listening experience. Please talk to @pnbennett or myself if interested. Detail of what you would be doing below.
https://t.co/RHFYjUqg2u

[Marcos Esteve @mecprojects]

I just published my last weekend project: Building an image content search system using Vespa https://t.co/PObvfHGX1a #deepLearning #machineLearning #vectordatabase #vespa #imageRetrieval

[Marcos Esteve @mecprojects]

Another good-written post to understand how #ChatGPT and other #LLMs are trained https://t.co/u5qFTvDUFZ

[Marcos Esteve @mecprojects]

Very good video about #ChatGPT and #promptengineering https://t.co/VT9M3CHcrd

[Gabriel Peyré @gabrielpeyre]

Summing the log of the nearest neighbors distances: a simple estimator for the entropy from samples. https://t.co/G1RLNGGlqq

[Jo Kristian Bergum @jobergum]

Had to write longer about this.

https://t.co/2tRwGu3unS

[Marcos Esteve @mecprojects]

Even if #ChatGPT is a great advancement and game-changer in many fields, I strongly think that many companies use it without privacy in mind. Some data shouldn't go out from the company/device (especially with sensitive data)

[Marcos Esteve @mecprojects]

Also for some use cases having a very simple model with good labelled data should be better than big LLMs (and maybe cheaper). Prompt engineering can help to quick prototype or improve models by doing data augmentation but substituting DL engineers is kind of optimistic

[Marcos Esteve @mecprojects]

After almost 2 years of working in search and relevance, seeing how Twitter rank and present the results on the platform is exciting. Lots of Graph techniques, embeddings and simple/heavy rankers https://t.co/VGkHPfJCiI