100%, it's very interesting to me when a lot value is created but the awareness/esteem of it, the story of it, or the people involved is next to zero. It's not fully true but it's sufficiently true.
RESEARCH
-
Meta Releases OMC25 Dataset with 25 Million Molecular Crystal Structures
By
–
We’re also releasing the Open Molecular Crystals (OMC25) dataset, comprised of 25 million structures, that was created to enable the FastCSP workflow. Read the paper: https://
ai.meta.com/research/publi
cations/open-molecular-crystals-2025-omc25-dataset-and-models/
… Find the dataset: https://
huggingface.co/facebook/OMC25 We’re excited to share these advances, -

Meta FAIR Announces FastCSP for Accelerating Organic Crystal Structure Discovery
By
–
The Meta FAIR Chemistry team continues to make meaningful strides. Today we’re announcing FastCSP, a workflow that generates stable crystal structures for organic molecules. This accelerates material discovery efforts and cuts down the time to design molecular crystals from
-
Iterating and Evaluating System Prompts for AI Models
By
–
Content strategy suggestion for @AnthropicAI or @OpenAI or any other AI Lab shipping systems with epic system prompts… I'd love to see a detailed write-up of how you iterate on, test and eval those prompts Would be a huge help for your customers who are building on your models
-
LLM capabilities surge dramatically in past twelve months
By
–
A year ago, the vast majority of LLMs we had weren't even capable of telling us how many Rs STRAWBERRY has The leap in capabilities we've experienced since September 2024 has been enormous, and it seems we're going to close out these 12 months with a golden flourish.
-

Faculty Engagement Session at DLI2025 Unites African AI Academics
By
–
🎓 We’re excited to announce the Faculty Engagement Session at #DLI2025 ! This special session will bring together African academics to explore the challenges and opportunities in advancing AI research and education on the continent. It will feature a panel discussion, mentorship circles, and policy forums — all aimed at building stronger, more connected AI faculty networks across Africa. The panel will feature @ulrichpaquet (@AIMSacza ), @vodumuyiwa ( @UnilagNigeria, @nithub_lag ), Esther Gathenya Khakata and Njeri Ngaruiya-Ng'ang'a (@StrathU ), Paul Amayo (@UCT_news ), and will be moderated by Shruti Sheth (@Google ). We’ll close with an open networking session to connect, collaborate, and inspire each other. 🌐 Virtual registration is still open — join us online! 👉 Register: buy.stripe.com/eVqfZh4BIeaNe… 🔗 More info: deeplearningindaba.com/2025/… #Indaba2025 #Urunana
→ View original post on X — @shakir_za, 2025-08-05 09:20 UTC
-
Estimating LLM Sizes: GPT-3 vs GPT-4 Parameters
By
–
Not completely made up actually – best guess more like. If you look at the size of GPT-3 and GPT-4, they reflect the relative size for each model (175b vs rumoured 1.8t for GPT-4). Then for the others I reduced the size for the models based on their price – so yes we don't know,
-
Elon Musk’s hiring challenges affect AI research talent acquisition
By
–
Elon’s ability to hire the best researchers just cratered. x.com/elonmusk/statu…
-
World Models: Action Conditioning and Representation Learning
By
–
– conditioning on actions is what makes it a world model
– planning action sequences is what makes it useful.
– predicting in representation space trained in a completely self-supervised, task-independent manner is what makes it complicated.
– training the encoder and predictor