Hadrien Mary

Yann LeCun recently made a striking statement: > “I’m not interested in LLMs anymore.” Coming from Meta’s Chief AI Scientist and one of the pioneers of modern deep learning, that’s a serious signal shift. LeCun argues that scaling language models — adding more data, compute, and parameters — has diminishing returns. Predicting the next token isn’t the same as understanding the world. His proposed direction for the next decade of AI focuses on systems that: 1. Understand the physical world through perception and interaction 2. Maintain consistent, persistent memory 3. Exhibit genuine reasoning and causal understanding 4. Can plan and execute complex, goal-directed actions He refers to this as Advanced Machine Intelligence (AMI) — AI that learns from observing and engaging with reality, not just from text scraped off the web. One of his comparisons drives the point home: A 4-year-old child learns more about physics in an afternoon of play than an LLM can infer from the entire internet. The child grasps why a ball rolls downhill. The model only knows the linguistic patterns that describe that motion. In essence, one learns through grounded experience. The other memorizes representations of experience. The next true leap in AI won’t come from larger models, but from architectures that integrate perception, memory, reasoning, and action — systems that learn from the world itself. What’s your take — are LLMs a stepping stone toward that future, or a technological cul-de-sac?

882

120 commentaires

🧬 Turn Claude into Your Personal AI Scientist Imagine having an AI research assistant that can: → Query 24 major scientific databases (PubMed, ChEMBL, AlphaFold, UniProt...) → Work with 40+ specialized packages (RDKit, BioPython, DeepChem, PyTorch Geometric...) → Connect to your lab infrastructure (Benchling, Opentrons, DNAnexus, OMERO...) → Apply 122 documented workflows across bioinformatics, cheminformatics, ML, and materials science This is the equivalent of thousands of 'tools'! That’s exactly what Claude Scientific Skills delivers. This open-source collection from the K-Dense team transforms Claude Code into a powerful scientific computing environment. Whether you’re analyzing genomic data, designing molecules, processing mass spectrometry results, or running multi-omics analyses. Getting started is ridiculously simple: /plugin marketplace add K-Dense-AI/claude-scientific-skills Then browse and install the packages you need. No complex setup. No endless configuration files. Just instant access to cutting-edge scientific tools. The best part? It includes "Scientific Thinking" skills, from exploratory data analysis and hypothesis generation to peer review and scientific writing. It’s not just about running code; it’s about doing better science. 🔬 Bioinformatics researchers: Analyze RNA-seq, run GWAS, query protein structures 💊 Drug discovery teams: Screen compounds, predict ADMET, perform molecular docking 🧪 Lab scientists: Automate protocols, manage ELN data, process experimental results 📊 Data analysts: Run statistical tests, create publication-quality figures, integrate multi-omics data 👉 Check it out: https://lnkd.in/eE9rzy8K What scientific workflow would you automate first? Drop your thoughts in the comments! 👇 #ScientificComputing #AI #Bioinformatics #DrugDiscovery #MachineLearning #OpenSource #ResearchTools #LabAutomation

178

7 commentaires

what you think a productive bioinformatician’s day looks like: • focus time 9a–5p what it actually looks like: • cluster job fail • quick Slack from biology team: “can you re-run?” • daily standup • cross-functional sync • status update on pipeline migration • data wrangling for a meeting you’re not invited to • meeting to discuss the metadata formatting issue • meeting to plan a future meeting about pipeline handoff • final review of deck for science update • an ask for a “quick analysis” • focus time 4:30p–5p delivering impact at scale

219

10 commentaires

CodonTransformer: multispecies codon optimizer with context-aware neural network (Fallahpour et al., 2025, Nat Commun) proposes an encoder BigBird Transformer that generates host-specific coding sequences given protein input. The researchers train on ~1M gene–protein pairs across 164 organisms and encodes organism identity + amino-acid + codon in the STREAM shared token scheme such that the model captures global and species-specific codon rules. In inferencing, the masked-LM decode "reverse-translates" the protein into the DNA conditioning on the target host. The model creates sequence with natural-like codon patterns and with fewer negative cis-regulatory elements, and higher codon similarity index (CSI) than native genomic baselines for the majority of the fine-tuned species indicating better host preferences matching. The authors believe that the bi-directional non-autoregressive encoder prevents the choice-locking at an earlier time step and facilitates region editing (optimizing e.g., a subdomain region). The authors also present an example of the use of the model in the form of synonymous variant score where the token probabilities map onto the experimental fitness and ribosome-stall readouts indicating usefulness in applications beyond design. The release includes the reproducibility friendly open base & fine-tuned model, the Python package, the Colab interface, the Zenodo datasets which facilitate the reproducibility & fine-tuning for specific host or protein family. Practically it offers the scalable organism-aware CDS optimizer for the heterologous expression in the fields of biotechnology, vaccines & therapeutics. https://lnkd.in/ed_W7CgE

13

🎆 Happy New Year from the ViennaRNA team. We’re kicking off 2026 with a new release: 𝗩𝗶𝗲𝗻𝗻𝗮𝗥𝗡𝗔 𝗣𝗮𝗰𝗸𝗮𝗴𝗲 𝘃𝟮.𝟳.𝟮 𝗶𝘀 𝗼𝘂𝘁! Download here: https://lnkd.in/giwnX_N Source download: https://lnkd.in/dBGsDe3C This update focuses on a solid mix of important bug fixes, benchmarking support, and API improvements across programs and the core library. 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀 (𝗣𝗿𝗼𝗴𝗿𝗮𝗺𝘀) - RNAeval: fixed structure output for >2 interacting RNAs - RNAcofold: fixed wrong application of soft constraints to single sequences - RNAfold / RNAalifold: new --benchmark option for easier performance tracking 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀 (𝗟𝗶𝗯𝗿𝗮𝗿𝘆 / 𝗔𝗣𝗜) - Fixed soft-constraint handling for single sequences in vrna_mfe_dimer() - Restored backward compatibility after early removal of vrna_constraints_add_SHAPE() (and related functions) - Smarter parameter reuse for successive vrna_params() / vrna_exp_params() calls - Added negative log probability (information content) as a generic probing-data strategy - Added a new function to chain probing-data strategies - Structure benchmark numbers switched to double precision If your workflows rely on 𝘀𝗼𝗳𝘁 𝗰𝗼𝗻𝘀𝘁𝗿𝗮𝗶𝗻𝘁𝘀, 𝗽𝗿𝗼𝗯𝗶𝗻𝗴-𝗶𝗻𝗳𝗼𝗿𝗺𝗲𝗱 𝗳𝗼𝗹𝗱𝗶𝗻𝗴, or you’re collecting 𝗿𝘂𝗻𝘁𝗶𝗺𝗲 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸𝘀, this release is particularly relevant. #ViennaRNA #RNA #Bioinformatics #ComputationalBiology #RNAstructure #RNAfolding #SoftwareRelease #OpenSource #SHAPE #DMS #Reproducibility

113

While armchair bioinformaticians are talking about how writing code is no longer a career moat, most of us are caught up in unglamorous problems that we wish AI could help with. Things like: • Metadata drift • Experimental context that never made it into the data • Deciding when a result is “technically correct but biologically wrong” • Aligning expectations across teams • Deciding what not to do Writing code used to be a bigger challenge in bioinformatics, sure. But it has always been just one of many challenges in a complex, and frankly beautiful, craft.

153

11 commentaires

Outliers and missing data in #omics datasets can complicate #biocomputational analysis – but they can also reveal unexpected insights. In this article, Sapient data scientists Shreya Gaddam and Cameron Lamoureux, PhD explore practical approaches to identify and address missingness and outliers in large-scale biological data, and how thoughtful analytical strategies can improve the reliability of downstream results. If you're working with high-dimensional datasets like those generated by discovery #proteomics and #metabolomics studies, this is surely worth a read. 🔗 https://lnkd.in/g6tSCVZ4 #datascience #mindthegaps #biologicaldata

12

My new blog post about why AI in spatial biology need new technologies capable of high-quality data and new technologies capable of delivering them at scale: https://lnkd.in/eSVqe6Rm ... Here is a cool preview of the future of spatial biology from the Genentech group: https://lnkd.in/eYVGybfF

44

1 commentaire