Thrilled to release o1-mini, a model near and dear to my heart 💙. o1-mini is an efficient model in the o1 series that’s super performant in STEM reasoning, especially math and coding. I can’t wait to see what you all build with o1-mini!!
Excited to announce 2nd Stanford Graph Learning Workshop on Wed Sept 28th with leaders from academia and industry to showcase recent advances of Graph Representation Learning across a wide range of applications. Program & free registration:
📣It is time to rethink graph databases in the era of GNNs and neural reasoners. In the new work with
@michael_galkin
,
@michaelcochez
,
@zhu_zhaocheng
, and
@jure
we explore the concept of Neural Graph Databases (NGDBs)!
📜
🌐
🧵1/n
Introducing GPT-4o mini! It’s our most intelligent and affordable small model, available today in the API. GPT-4o mini is significantly smarter and cheaper than GPT-3.5 Turbo.
Excited to share our
#ICLR2020
paper and code for Query2box, a multi-hop reasoning framework on knowledge graphs. We design box embeddings to answer complex logical queries with conjunction and disjunction. Joint work with
@weihua916
and
@jure
.
Come join us for the tutorial @ Learning on Graph: Complex Reasoning over Relational Database, we will present SMORE + Scallop!
@StanfordAILab
@CIS_Penn
@GoogleAI
@LogConference
Time: 9-1030am PT, Dec. 10
Webpage:
(Attendees will get a virtual cookie👇)
No more waiting. o1's is officially on Chatbot Arena!
We tested o1-preview and mini with 6K+ community votes.
🥇o1-preview:
#1
across the board, especially in Math, Hard Prompts, and Coding. A huge leap in technical performance!
🥈o1-mini:
#1
in technical areas,
#2
overall.
We have reached an agreement in principle for Sam Altman to return to OpenAI as CEO with a new initial board of Bret Taylor (Chair), Larry Summers, and Adam D'Angelo.
We are collaborating to figure out the details. Thank you so much for your patience through this.
Come check our
#uai
paper on task inference for meta-reinforcement learning. We propose OCEAN, an online task inference framework that models tasks with global and local context variables. Joint work with
@AnimaAnandkumar
@animesh_garg
@yukez
@jure
. 👉
@legit_rumors
@OpenAIDevs
- OpenAI o1-mini is optimized for STEM applications at all stages of training & data. It has limitations of world knowledge. Check our research blog post for more details.
- We are working on adding more knowledge. Stay tuned for next version of o1-mini!
🎄It's 2023! In a new post, we provide an overview of Graph ML and its subfields (and hypothesize for '23), eg Generative Models, Physics, PDEs, Theory, KGs, Algorithmic Reasoning, and more!
With
@ren_hongyu
@zhu_zhaocheng
@chrsmrrs
and
@jo_brandstetter
Check out our new scalable KG embedding framework! It supports super efficient training of single/multi-hop algorithms on extremely large graphs (~86m nodes), including GQE, Query2box, BetaE, TransE, RotatE, DistMult, ComplEx and so on!🥳
Excited to share our collaboration with
@GoogleAI
: SMORE is a scalable knowledge graph completion and multi-hop reasoning system that scales to hundreds of millions of entities and relations.
@ren_hongyu
,
@hanjundai
, et al.
👉Check out Stanford Graph Learning workshop on Sep. 16! Super excited to present recent advances on Knowledge Graphs. Free registration at: , it will be live streamed!
Stanford is proud to bring together leaders from academia&industry to showcase advances in Graph Neural Networks. Program includes applications, frameworks and industry panels on challenges of graph-based machine learning models. Register at:
@dvyio
@OpenAIDevs
OpenAI o1-mini is optimized for STEM applications at all stages of training & data. It has limitations of world knowledge. Check our research blog post for more details.
🤓The project took us months of conceptualization and refinement, and we are glad to finally release it – look forward to hearing both from Graph ML and Database folks!
Check out the blogpost from
@michael_galkin
for more details :)
10/10, n=10
Check out Connection Subgraph Reasoner accepted at
#NeurIPS2022
, a novel pretraining subgraph-level objective for few-shot KG link prediction.
Paper:
Code:
Excited to share our
#NeurIPS2022
paper: Few-shot Relational Reasoning via Connection
Subgraph Pretraining! . We propose to pretrain on subgraph matching for few-shot relational reasoning tasks. 🧵
Joint work with
@ren_hongyu
and
@jure
!
@StanfordAILab
@InverseAGI
@OpenAIDevs
OpenAI o1-mini is optimized for STEM applications at all stages of training & data. It has limitations of world knowledge. Check our research blog post for more details.
How do we answer complex queries in an inductive setting? Check out this new
#NeurIPS2022
paper with a novel method and a set of benchmarks for inductive multi-hop reasoning. This guy makes cool gifs!!
📢Excited to give a talk about SMORE, the first scalable framework that supports link prediction and multi-hop reasoning over massive knowledge graphs!
Code:
Tuesday 10am ET Room 206
#KDD2022
We’re hosting an AMA for developers from 10–11 AM PT today. Reply to this thread with any questions and the OpenAI o1 team will answer as many as they can.
o1-mini is the most surprising research result i've seen in the past year
obviously i cannot spill the secret, but a small model getting >60% on AIME math competition is so good that it's hard to believe
congrats
@ren_hongyu
@shengjia_zhao
for the great work!
We’re rolling out a bunch of small updates to improve the ChatGPT experience. Shipping over the next week:
1. Prompt examples: A blank page can be intimidating. At the beginning of a new chat, you’ll now see examples to help you get started.
2. Suggested replies: Go deeper with
Why do we need NGDBs and what do graph DBs lack? The biggest motivation is incompleteness - symbolic SPARQL/Cypher-like engines can’t cope with incomplete graphs at scale. Neural graph reasoning, however, is already mature enough to work in large and noisy incomplete graphs.
2/n
Broadly, NGDBs are equipped to answer both “what is there?” and “what is missing?” queries whereas standard graph DBs are limited to traversal-only scenarios assuming the graph is complete.
4/n
✍️Check out our github where we collect all relevant references, feel free to send PRs - we welcome all contributions :)
We’ll be updating other materials on the project website .
9/n
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
@felixchin1
@OpenAIDevs
The weekly rate limit is 50 for o1-mini. We are working to increase those rates and enable ChatGPT to automatically choose the right model for a given prompt!
What are NGDBs? While their architecture might look similar to traditional DBs with Storage and Query Engine, the essential difference is in ditching symbolic edge traversal and answering queries in the latent space (including logical operators).
3/n
Thanks to GNNs, NGDBs can answer either the entire query graph at once or execute them sequentially. We don’t need any explicit indexes - the latent space formed by a neural encoder *is* the single uniform index.
5/n
Finally, we outline challenges and open problems for NGDBs. Lots of cool stuff to work on! (especially if you are in existential crisis with GPT4, in fact how to design LLM interface for NGDB and how to let NGDBs help compress and accelerate LLMs are very promising)
#GPT4
8/n
What’s the difference between NGDBs and vector DBs?
Vector DBs are fast and encoder-independent consuming embeddings from all kinds of models. Though they are limited to a few distance functions and lack query answering capabilities, they fit well into the NGDB blueprint!
6/n
In the NGDB framework, we create a taxonomy and survey 40+ neural graph reasoning models that can potentially serve as Neural Query Engines under 3 main categories: Graphs (theory and expressiveness), Modeling (graph learning), and Queries (what can we answer).
7/n
Super excited to share Open Graph Benchmark (OGB)! OGB provides large-scale, diverse graph datasets to catalyze graph ML research. The datasets are easily accessible via OGB Python package with unified evaluation protocols and public leaderboards.
Paper:
“how many r’s in strawberry?”
I had to ask this to demo our new model o1-preview 😎
LLMs process text at a subword level. A question that requires understanding the notion of both character and word confuses them.
OpenAI o1-preview "thinks harder" to avoid mistakes.
(2/x) The key idea is to factorize the conditional expectation formulation of attention in a structured way. Each token can attend to all the other tokens either by direct attention, or indirectly via a "proxy" that summarizes a local region.
Paper link:
@weihua916
@jure
We show that embedding queries as hyper-rectangles (box) is a natural way to handle relational projection and conjunction. We also analyze the hardness of KG embeddings handling disjunction and provide a clean solution using disjunctive normal form.
We embed queries as Beta distributions. For conjunctions, we take weighted product of the PDF of the Beta embeddings of input queries. For negation, we calculate the reciprocal of the parameters of the input so that the high density region will be become low and vice versa.
Check out query2box, joint work with
@weihua916
and
@jure
. Query2box performs multi-hop logical reasoning on knowledge graphs.
#ICLR2020
#iclr
Talk:
Website:
@weihua916
and I will present the work at 10 pm Tue. and 1pm Wed. PT.
The embedding also captures the uncertainty of a query, measured by the number of answers it has. We found a natural connection between the entropy of the Beta embedding and # answers, without any training needed to enforce this correlation.
(3/x) We propose six instantiations of Combiner idea with different factorization patterns. And we show we can take a prior sparse Transformer X and make it Combiner-X with full attention capacity while keeping the same asymptotic complexity!
The task is answering first-order logic queries (with existential, conjunction, disjunction and negation) on incomplete KGs. Key insight is to embed the queries and entities, and reason in the embedding space. Previous methods (GQE, Q2B..) cannot handle negation/set complement.
(4/x) We evaluate Combiner on autoregressive / bidirectional sequence modeling across images and texts. Combiner achieves much better performance across a wide range of tasks and is super scalable.
The code of graph information bottleneck (GIB) can be found at , we used the fantastic Pytorch Geometric library by
@rusty1s
! Feel free to check it out!
(5/n)
If all training images for a GAN/VAE/PixelCNN have 2 objects, will they only generate images with 2 objects? If trained on (🔵,💙,🔴), will they also generate ❤️? Find out in
@shengjia_zhao
's blog post on generalization and bias for generative models.
👉