Powered with a novel search algorithm, AlphaGeometry 2 can now solve 83% of all historical problems from the past 25 years - compared to the 53% rate by its predecessor.
It solved this year’s IMO Problem 4 within 19 seconds. 🚀
Here’s an illustration showing its solution ↓
Introducing AlphaGeometry: an AI system that solves Olympiad geometry problems at a level approaching a human gold-medalist. 📐
It was trained solely on synthetic data and marks a breakthrough for AI in mathematical reasoning. 🧵
We show large language models trained on massive text corpora (LM1b, CommonCrawl, Gutenberg) can be used for commonsense reasoning and obtain SOTA on Winograd Schema Challenge. Paper at , results reproducible at
As also observed by OpenAI's GPT-2, training data quality is important. We release the STORIES corpus introduced in our work . The corpus is a high quality subset of CommonCrawl with a total of ~7B words (~32GB) can be found here:
Wow! An old project of mine is now the 7th most popular Machine Learning project across all Github in 2018, alongside with Tensorflow and Scikit-learn? I really need to spend some time polishing it now...
From the programming languages you used most to the most popular data science packages, we’re digging into the data on Machine Learning from 2018. Find out what we discovered
Had the chance to sit next to Daniel
@xpearhead
in the early days of the project and tried out the interactive Meena. It has always been *this* surprising and funny :) BIG Congrats to the team with this publication. The possibilities to build up from here is endless.
Excited to share a new work by
#GoogleAI
resident
@thtrieu_
(with
@iamandrewdai
, me, & Quoc Le) on training very long RNNs (up to 16K long). See paper for extreme cases of zero or little backprop on RNNs ;)
"We see contributions to traditional conferences and publications in journals as an important part of our work, but also support efforts that go “beyond the research paper"".
I'm excited to finally share what I have been working on.
Today we are officially launching Cohere For AI
@forai_ml
a non-profit research lab that aims to reimagine how, where, and by whom research is done.
It's amazing how fast
#NLProc
is moving these days.
We have now reached super-human performance on SWAG, a commonsense task that will only be introduced at
@emnlp2018
in November!
We need even more challenging tasks!
BERT:
SWAG:
Gemini 1.5 Pro - A highly capable multimodal model with a 10M token context length
Today we are releasing the first demonstrations of the capabilities of the Gemini 1.5 series, with the Gemini 1.5 Pro model. One of the key differentiators of this model is its incredibly long
Making a synthetic dataset of mathematical proofs is hard! It's easy to make a whole lot of "1+1+1+...=491" style theorems.
I'm surprised this method of random construction and transformation finds so many classical geometric theorems.
Maybe because the domain is somewhat
Want to see space-time contract/dilation? This series on Special Relativity (SR) is beautiful. The author squashed space-time to 2D, explained the two postulates by geometric intuition and run a simulator on top of it. It is the 3blue1brown of SR!
Great paper exploring attention architecture for images! We encounter similar results in our latest work . Table 5 shows that keeping first layers conv, while using attention for last layers improve ResNet performance.
Further studies show that self-attention is the most useful in later layers while convolutions better capture lower-level features. Combining their will be an interesting research direction.
A new era of NLP has just begun a few days ago: large pretraining models (Transformer 24 layers, 1024 dim, 16 heads) + massive compute is all you need. BERT from
@GoogleAI
: SOTA results on everything . Results on SQuAD are just mind-blowing. Fun time ahead!
Autoformalization with LLMs in Lean!
@zhangir_azerbay
and Edward Ayers built a chat interface to formalize natural language mathematics in Lean:
Very impressive work!
New York Times article on AlphaGeometry
Geometry was my favorite subject in high school because solving them requires many step of reasoning and planning. Geometry problems however have been difficult for AI to solve. Our Nature paper shows my team’s progress in Geometry
This article speaks many, I believe, hidden truths about Quoc Le on GoogleBrain & seq2seq. Personally, I have enjoyed working with Quoc who cares less about credit assignment but rather teamwork and long-term vision :)
Excited to announce Piano Genie, an intelligent controller that allows anyone to improvise on the piano! This was my internship project at
@GoogleMagenta
with
@iansimon
and
@sedielem
. Blog post with more information:
This feels like a method that ought to be more generally applicable (as indeed the authors suggest). I have a few ideas of problems I'd be very interested for it to be tried out on and that seem to be of the right type.
i dont feel bad for ML PhDs who think "my job is ruined, i can't do anything outside a big lab" etc. i can think of >10 low hanging fruit things "someone should do" that im too busy for bc i have to "make revenue". academia is 100% unfettered from use value, yall have it so easy.
Introducing Sora, our text-to-video model.
Sora can create videos of up to 60 seconds featuring highly detailed scenes, complex camera motion, and multiple characters with vibrant emotions.
Prompt: “Beautiful, snowy
Our latest work on continuous control from pixels by
@debidatta
, Jonathan Tompson,
@coreylynch
, and I will be presented at the MLPC workshop at
#ICRA2018
this afternoon in room M4. Paper:
tried this for translation a while back but didnt try hard enough -- i think the transformer needs way more registry & will improve on all tasks. also related to
@ChrSzegedy
"the transformer is so small" at 52:15 .
What I mean when I say “registers”: additional learnable tokens (like the [CLS]), but these ones are not used at output. No additional info at input, not used at output: these tokens could seem useless!
one possible scenario: openai slack channel LTAO browsing X. because their method has nothing to do with q-learning nor rl. unlikely tho, because they are busy shipping q* into gpt5, not browsing X.