AI scientist and engineer with 20 years of academic and industrial experience, with hands-on knowledge of the latest AI techniques, and building great AI teams.
Repeat after me:
1. Current Auto-Regressive LLMs are *very* useful as writing aids (yes, even for medical reports).
2. They are not reliable as factual information sources.
3. Writing assistance is like driving assistance: your hands must remain on the keyboard/wheel at all times
Awesome, I love simple and innovative ideas like the new data2vec model från Meta AI, with a very simple method that works across modalities such as images, text, and audio. Occam's razor in action! https://t.co/Zf2v4ca61k
Interesting article talking about the paper "Spinning Language Models for Propaganda-As-A-Service" that describes how an "attacker" can train a model to generate propaganda text. Nice that they also talk about how to detect this type of attack.
https://t.co/Ci0b0GDDZa
The general trend of going multitask, multimodal and multilingual continues to progress. With the latest model from Facebook, a multilingual machine translation model OUTPERFORMS previous bilingual approaches for the first time. https://t.co/t4cfHgb7oK
Fascinating work that can control the opinion/messaging of text automatically generated from language models such as GPT-3 without ANY fine-tuning at all. Scary if abused to generate text with potentially subjective and propaganda messaging. https://t.co/wTblGB8mqs
Check out this 2020 @TEDTalks video from @ylecun (@nyuniversity@facebookai) in which he discusses "self-supervised machine learning" and how he's trying to build machines that "learn with common sense". Wave 🌊 of the future 🤖 stuff... https://t.co/bz3QMB2Tx7
LEAF, a neural net that seems to make it possible to remove static Mel-spectrograms with an automatically trainable representation. Sounds awesome, we have a number of audio projects in Peltarion and I will certainly start experimenting with LEAF directly. https://t.co/84Y8PEKGoz
Thank you, @BenTheElder, for your great article about "Open Source Virtual Background." This is something that I have been missing for a long time, so I just had to repackage that to make it easy for everyone to use. See https://t.co/zJrKTLMFfi
Highly recommend the https://t.co/Ax9ONcIv3O weekly summary to keep up with the field of ML. They're using AI to identify the top publications and breakthroughs, it's a game-changer.
Impressive research from DeepMind, which is also very interesting from a philosophical perspective. Demonstrates that rule-based symbolic systems are not needed for reasoning, and Transformer-based approaches can efficiently learn how to reason. https://t.co/eEcZgVMdfE
In this first episode of Artificial Intelligence After Work (AIAW Podcast), we had the pleasure to meet Louise Callenberg, Section Director, Digital Renewal and Collaboration, The Swedish Association of Local Authorities and Regions (SKR).👇🏻
https://t.co/SrTOPClLtn