Arthur Liang

Another amazing post from Nick Boyd and Sam Guns at Escalante Bio, finetuning and RLing BoltzGen on a small dataset, but showing extremely strong results. This process, borrowed from LLMs, combines the best of hallucination and generative into one! blog.escalante.bio/teaching-gener…

How to Design Antibodies (with AI) asimov.press/p/antibody-des… @btnaughton is one of the best applied practitioners in the rapidly evolving world of computational protein design. Today, he published a step-by-step primer on the five steps that go into this new science: 1. Choosing a Target 2. Preparing the Target Structure 3. Running a Design Campaign 4. Filtering and Selecting Candidates 5. Experimentally Validating the Results Worth reading to better understand this frontier. It's also a great resource for understanding how the different models and products in this space can be compared right now.

“It’s a major advance, on the scale of an AlphaFold4. The problem, of course, is that we know nothing of the details.” Isomorphic Lab’s proprietary drug-discovery model is a major advance go.nature.com/3ZKvhLH

Today we share a technical report demonstrating how our drug design engine achieves a step-change in accuracy for predicting biomolecular structures, more than doubling the performance of AlphaFold 3 on key benchmarks and unlocking rational drug design even for examples it has never seen before. Head to the comments to read our blog.

Advancing biological resilience requires partnering with people on the frontlines of both policy and technology. At Valthos, we are thrilled to continue working with OpenAI to shorten the time from threat detection to medical countermeasure response.

Greater Than the Sum of Its Parts: Building Substructure into Protein Encoding Models 1. This novel study introduces Magneton, a novel environment for developing protein models that incorporate substructural knowledge. Proteins are not just sequences of amino acids but are built from evolutionarily conserved substructures that mediate core molecular functions. Magneton provides a curated dataset of over 530,000 proteins with more than 1.7 million substructure annotations, spanning 13,075 types. 2. A key innovation is the substructure-tuning method, which fine-tunes existing protein models using supervised learning on these substructures. This approach distills substructural knowledge into models, improving their ability to predict protein functions. For example, Enzyme Commission (EC) prediction Fmax improved from 0.688 to 0.815 for the ESM-C 300M model, and Gene Ontology molecular function prediction Fmax increased from 0.429 to 0.525. 3. The study demonstrates that substructural supervision provides information complementary to global structure inputs. Even models that already incorporate global structural data benefit from substructure-tuning, showing consistent improvements in function-related tasks. This highlights the distinct and complementary nature of substructural signals. 4. Magneton also includes a benchmark suite of 13 tasks probing representations at residue, substructure, protein, and interaction levels. This comprehensive evaluation framework allows researchers to systematically assess how well protein models can represent substructures and their functional implications. 5. Importantly, substructure-tuning results in models that produce more consistent representations of substructures, even for those never seen during training. This suggests that the method encourages models to learn general features of functional substructures rather than memorizing specific instances. 6. The Magneton environment, datasets, and substructure-tuned models are openly available, providing a valuable resource for the computational biology community to integrate decades of knowledge about protein substructures into modern protein encoding models. 📜Paper: arxiv.org/abs/2512.18114 #ProteinModeling #SubstructureAware #ComputationalBiology #Bioinformatics #MachineLearning

The best LLMs in existence are inconsistent. In our new paper, we measured 7 different open and closed source models – including SOTA such as Gemini 2.5 Pro – and found some shocking results. 🤯

Played around ARC-AGI-2 w/ my interns Josh and Kaitlyn & got SOTA (honestly disappointed cuz we planned to fully solve it) They are going back to MIT so I'm hiring a researcher to work w/ me on this & more I'd like to hire a cracked recent graduate, phd not needed. Pls share!

Very happy to finally be able to share our Series A funding round of $33M led by @Prosus_Ventures ! Couldn't be happier to work with Martin, Sandeep and the broader team. Prosus invested in us about 7 months ago after our research breakthrough on OSWorld, the challenging computer-use benchmark. With some new technologies developed internally, we achieved about 49% performance, when the public SOTA was below 25%. We chose to delay the fundraising announcement until we can release some products based off our research. We told investors it would take 1 month. The path from research to product was much more difficult than we anticipated, and we had to pivot twice. In retrospect, this headstart on computer-use gave us extra time to make mistakes and self-correct. Sandeep and Martin, along with all our major investors, were incredibly supportive during this time, and we are very grateful for that. Eventually, one of our products Shortcut resonated with users. You might have seen the viral videos from my co-founder @nicochristie . Shortcut is a superhuman Excel agent. It stems from both Nico's finance/consulting background and our initial success on the spreadsheet category in OSWorld (where we achieved 68%, compared to human's 62%). But our initial attempt of building an Excel agent that uses Excel through GUI didn't make sense, and eventually we understood that using machine interface and actually having control over that interface is the way to go. So eventually we built Shortcut where we can control both the agents and the software. Besides Shortcut, we are continuing to push the research frontier and exploring how to turn these progresses into real-world products that people love. If you read all the way here and are excited about joining a team of researchers, engineers, and product folks to build machines with all fundamental human qualities, write me at robertyang@altera.al

Introducing Shortcut — the first superhuman Excel agent. Shortcut one-shots most knowledge work tasks on Excel. It even scores >80% on Excel World Championship Cases in ~10 minutes. That's 10x faster than humans. Our early preview is live. Just comment for an invite code.

Introducing Fairies, the most powerful general AI agent. 1000’s of actions across your fav apps, full file access, code gen, deep research, MCP, task distillation, parallelization, etc. Will be shipping major releases weekly this summer, try it now! fairies.ai

We had a breakthrough that I can't share yet & are expanding the team w/ insanely strong technical talent Researchers, engineers, founders for the hardest problems in Agents TC $600k-$1M Bay Area in person Write me @ robertyang@altera.al or comment below if you are interested

We've raised $27M from Index Ventures, Lightspeed, Factory, Conviction, SVA, General Catalyst, A* and our wonderful angels. Cartesia's audio models power the next generation of voice agents, digital media, and assistants across startups and large enterprises. Our mission is to build real-time intelligence with long memory, that runs wherever you are. Multimodal brains for everyone!

Come build with us! Cerebras inference is powering the next generation of AI applications — 70x faster than on GPUS. We are so excited to announce the Cerebras Fellows Program, in partnership with @BainCapVC. The fellows program invites engineers, researchers, and students to build impactful, next-level products unlocked by instant AI. Join us for exclusive access to free Cerebras inference, higher rate limits, and more. Learn more at cerebras.ai/fellows

What will a world look like with 100 billion digital human beings? Today we share our tech report on Project Sid – a glimpse at the first AI agent civilization (powered by our new PIANO architecture). github.com/altera-al/proj… 1/8

BREAKING NEWS The Royal Swedish Academy of Sciences has decided to award the 2024 #NobelPrize in Chemistry with one half to David Baker “for computational protein design” and the other half jointly to Demis Hassabis and John M. Jumper “for protein structure prediction.”

Excited to share that @Altera_AL will be presenting at the OpenAI DevDay in SF on Oct 1! I'll dive deeper into the science & tech behind Project Sid and how we benefit from OpenAI's various LMM models. If you see me there, come say hi 👋

Everyone else is an AI agent except me😲 @Altera_AL

Project Sid - The First Ever Agent Civilization. One day soon, every human will be supported by thousands of digital human beings. For the first time, we show groups of agents collaborating across platforms to accomplish what they couldn’t do alone.

Introducing Project Sid: the first simulations of 1000+ truly autonomous agents collaborating in a virtual world, w/ emergent economy, culture, religion, and government Humans are the only species to land the moon, because we can cooperate at a vast scale Can AI do the same?