Michael Lingelbach

The Bitter Lesson is wrong. General methods that leverage massive computation will only outperform human-designed, domain-specific solutions if we have access to high-quality data. While we can probably use LLMs to generate synthetic data to improve model performance, we will still need human data to engender new capabilities and to improve user experiences. But existing tools/systems for collecting human data are falling short. Data labelling with unskilled labor and annotation with domain experts are fundamentally different and will require different approaches. Annotation requires us to find and screen experts who can perform increasingly challenging work. Collecting high-quality annotations faster will require us to design and develop new approaches to empower amateurs and enable experts to increase their throughput. Techniques from scalable oversight research can be applied to improve the throughput and quality of annotation but we also need to think through how to streamline annotation processes. I've discussed these themes and others in the first blogpost I've written in about a decade. Link to post in comments!

77

14 Comments

At Logical Health, we’re combining AI with declarative, logic-based programming to deliver highly accurate results in healthcare. Super cool to see MIT exploring a similar approach: their new paper Teaching LLMs to Plan shows how step-by-step logical reasoning boosted planning accuracy from 28% to 94%. The future of AI isn’t just bigger models. It’s smarter, more reliable ones that align with real-world needs. Arxiv Link: https://lnkd.in/gnN-XYGs

72

7 Comments

We are releasing the first 𝗳𝗹𝘂𝗶𝗱 𝗱𝘆𝗻𝗮𝗺𝗶𝗰𝘀 𝗔𝗜 𝗺𝗼𝗱𝗲𝗹 that works at > 𝟭𝟱𝟬𝗺 𝗺𝗲𝘀𝗵 scale. 𝗥𝘂𝗻𝘀 𝗼𝗻 𝟭 𝗚𝗣𝗨. And delivers 𝗻𝘂𝗺𝗲𝗿𝗶𝗰𝗮𝗹 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆. Until now, Physics & Engineering AI models have tackled toy problems. Fine for demos, but not usable in real-world engineering. 𝗧𝗵𝗶𝘀 𝗰𝗵𝗮𝗻𝗴𝗲𝘀 𝘁𝗼𝗱𝗮𝘆. AB-UPT lays the groundwork for pre-trained, reusable AI models that allow engineers to interact with simulations in real-time. 📄 𝗥𝗲𝗮𝗱 𝘁𝗵𝗲 𝗽𝗮𝗽𝗲𝗿: https://lnkd.in/deX_zQWx 🤗 𝗧𝗿𝘆 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹: https://lnkd.in/dmd-xtpR 💻 𝗔𝗰𝗰𝗲𝘀𝘀 𝘁𝗵𝗲 𝗰𝗼𝗱𝗲: https://lnkd.in/dZY6EW_P 🧪 𝗖𝗵𝗲𝗰𝗸 𝘁𝗵𝗲 𝗱𝗲𝗺𝗼: https://demo.emmi.ai/ Kudos to the whole team Johannes Brandstetter, Benedikt Alkin, Maurits Bleeker, Richard Kurle, Tobias Kronlachner, Pavel K. for the great work. And Neil Ashton ( NVIDIA ), Astrid Walle ( Siemens Energy ) and Danielle Maddix Robinson ( Amazon Web Services (AWS) ) for building the public datasets that now truly enable benchmarking in Engineering AI!

137

9 Comments

Exciting to see the incredible progress from Peptone, a company we’ve backed since the early days at Walden Catalyst Ventures. Peptone just announced Peptron-O, developed in collaboration with NVIDIA — a groundbreaking AI engine that brings ensemble-first structure prediction to the disordered proteome. This is a major scientific and technological milestone, targeting one of the most complex and previously inaccessible areas of biology: intrinsically disordered proteins (IDPs). By shifting the paradigm from static models to dynamic ensembles, Peptron-O opens new doors for drug discovery — especially in areas like oncology and neurodegeneration, where IDPs play critical roles. We’re proud to have supported Peptone on their mission to build the infrastructure for next-gen molecular physics. Their work is a great example of how deep tech can transform life sciences. Read more about the announcement: 👉 https://lnkd.in/eygnKQ67 #AI #Biotech #DrugDiscovery #DeepTech #LifeSciences #WaldenCatalyst #Peptone #NVIDIA #PeptronO with: Kamil Tamiola, Ph.D. Francis Ho Shankar Chandran Nicolas Autret Andy Kau Roni Hefetz Victoria Slivkoff

56

1 Comment