David Mezzetti (@DavidMezzetti) header
David Mezzetti (@DavidMezzetti) profile picture

Founder @neumll | Creator of TxtAI

Washington DC Metro 🇺🇸
Joined February 2013
Photo by David Mezzetti (@DavidMezzetti)Photo by David Mezzetti (@DavidMezzetti)Photo by David Mezzetti (@DavidMezzetti)Photo by David Mezzetti (@DavidMezzetti)Photo by David Mezzetti (@DavidMezzetti)Photo by David Mezzetti (@DavidMezzetti)Photo by David Mezzetti (@DavidMezzetti)Photo by David Mezzetti (@DavidMezzetti)Photo by David Mezzetti (@DavidMezzetti)Photo by David Mezzetti (@DavidMezzetti)
🎉 100 Contributors to txtai! I'm honored that the project has so many people around the world using it and contributing. Over 50 of those contributors started in the last month+. Very exciting!
1
65
@lateinteraction You might find this interesting...
11
David Mezzetti retweeted
NeuML is excited to introduce Atto, a new family of compact encoder models! This is one of our largest model releases to date, with more than 30 new models ranging from just 840 parameters to 28 million parameters. One of the most interesting findings is how well extremely small models can perform, particularly when leveraging late-interaction techniques. Even an 840-parameter model can deliver surprisingly strong results. Atto also explores how late-interaction principles can be incorporated directly into model training, bringing some of those benefits to dense embedding models. The goal is to explore what's possible when we rethink how encoder models are designed, trained, and optimized. We're excited to share this work! Read more: huggingface.co/blog/NeuML/in…
1
2
3
116
The current drafts for AI Regulation talk about frontier models. What happens when someone (or AI) creates a new architecture that shatters our understanding of scaling laws of models? Let's say things can be done in hundreds or thousands of parameters? Then what?
1
76
David Mezzetti retweeted
New contributors continue to come in on almost daily basis for txtai. The project had little over 20 contributors only a few months ago. Why the influx now? The biggest factor is the push towards Local AI and Open Models. We're firmed committed to this approach. Those who are like-minded in the passionate pursuit of AI that's open for all, join us!
1
2
347
Innovation is never over. If you missed out on one wave, there is another ahead. Look forward not backwards. With AI, we're much closer to the beginning than the end.
1
44
It's pretty obvious that making smaller and more efficient AI models is what 2027 will be all about.
3
172
I'm not an AI doomsdayer. I'm not too concerned about AI replacing people on a mass scale. Although I do run into people from time to time whom perhaps I wouldn't mind being replaced by AI 😀
1
67
Between PIP certificates, PHP and LiteLLM (again 😡), it's a busy day in the security department. Luckily, none of this effects txtai directly. Just make sure you read the following articles. discuss.python.org/t/informa… apnews.com/article/php-websi… docs.litellm.ai/blog/securit…
1
155
If someone told me that coding agents create each PR in anticipation of a follow on PR to drive up token counts, it wouldn't surprise me.
3
83
Late interaction and MaxSim don't take a lot of parameters
1
67
I've never been shy to share that NeuML is just 1 person. Now with the power of open source, it has a 35-60+ person engineering team behind it. Whether that's people, people with Claude Code Agents or Agents, it doesn't matter to me. txtai keeps improving "for free".
1
65
Top 4 currently trending models on the Hugging Face Hub. None generate text.
1
62
Out of all courses I took in College, Signals and Systems has the principles that I've leaned the most on in machine learning. Fourier Transforms and Convolution just seems to be something I often come back to when solving hard problems.
3
152