Joe Palermo
93 posts
Joe Palermo
@joepalerm0
Machine learning engineer from Toronto; Member of Technical Staff @OpenAI
Joined December 2013
- OpenAI is nothing without its people
- The overwhelming feeling is of love for each other, commitment to our mission and the absolutely indomitable will of the OpenAI team.
- we made a deepfake @dessa to raise awareness about the future of synthetic media and the internet thinks we're evil (spoiler: we're not — watch the doc)Don't trust your eyes. @DavidBarstow goes inside the race to create the perfect deepfake — an ultrarealistic video that could undermine your faith in reality. Watch – very closely – how they do it on #TheWeeklyNYT on @FXNetworks and @Hulu. nyti.ms/349vVUE
00:00 - Just released some work with @johnnyeet and @aloksingh. We convert the DeepMind Mathematics Dataset into a reinforcement learning environment by interpreting it as a program synthesis problem. arxiv.org/abs/2107.07373
- Really fun work from my colleague @pippinlee showing how well people could distinguish real speech from synthetic speech that we generated with deep learning at @dessaSpending the weekend in Montreal eating bagels and reading, so it's probably a good time to also publish results from our synthetic audio Turing test. Full summary: reading.supply/@pippin/how-ac…
- Replying to @recursecenter and @b0rkToday's Joy of Computing: Wolfram Beta, a computational knowledge engine like Wolfram Alpha, but worse! Powered by a tiny neural network, it only does addition and subtraction. Made by @mathemakitten, @j_w_palermo, and friends. joy.recurse.com/posts/563-wolf… wolframbeta.art
- Replying to @joepalerm0This work was in part inspired by @fchollet's explanation of the potential value of program synthesis (joepalermo.github.io/2021/01/10/tal…).
- Replying to @joepalerm0Much work remains to be done but we believe that program synthesis will become increasingly important in the quest to build AI capable of reasoning from limited data.
- Replying to @spoluGreat paper! I'm embarking on my own project and I'm trying to assess MetaMath vs. HOList. Since HOList is also context-free + amenable to search + has lots of training data, would you say that the main advantage of MetaMath is efficiency and simplicity of the implementation?
