Javier Tordable

LLMs will not lead to generalized AGI. Yann Lecun said as much in a podcast with langchain and I couldn't agree more. LLMs are able to build an 'understanding' of a complex system whose rules are not clear. Like language. Based on the training data they can build complex functions. Functions that would be otherwise be too difficult to right down. ChatGPT has figured out rules of human language in a way that we can not explicitly right now. We just don't know what they are. But ChatGPT, somewhere in it's parameters, has figured it out. But the thing is that they have been built on a lot of data. A lot of data! Almost everything that has ever been published on the web. That is millions of documents conceivably covering all facets of writing in English. I mean, can you really say something that has not said somewhere on Reddit? The same approach doesn't extend to other problems just cause we do not have that size and diversity of data. We often hear that data in genomics is huge. Yes it is. But it's mostly cause the raw data that is captured is too noisy for it to be useful. That is why genomics data requires that amount of data processing before it can be used. It's a long winded say to that genomics data is just not as rich per unit of storage. Information content per mb is not that high. So the data might look large on the disk, but it does not have that much information. Generative models can only do so much without data from a similar space. The space of all possible molecules in all physical configrautions is just way too vast. The data available is minuscule compared to the possibilities. Secondly, just having a lot of data will not lead to generalized intelligence. That is a few fundamental breakthroughs away. No amount of compute or data thrown at DNNs are going to solve it. We need fundamental breakthroughs in our understanding of real intelligence - something that the field is still many years if not decades away from. What does this mean for leaders in life-sciences research? - Apply LLM like principles to spaces where enough data exists. - Don't bet on problems that need fundamental discoveries. Leave that to AI research labs in academia. - Maintain your data well and continue to invest in data management. Data will continue to be the moat. - Know which problems are 0 to 1 and which are 1 to n. This requires an understanding of the fundamentals of AI and the maths behind it. The road to successful AI adoption is long and not an event. It is a set of small steps that overtime will create value. There are no magic bullets out there. Link in comments to a review of new models by OpenAI for deep-research in comments. Nidhi and team have done an excellent job of summarizing we where are and what remains.

24

3 Comments

"The companies that build the data layer will matter more than most of the models sitting on top of it."

5

Comparing models.. Finding that a small 1B #llm hallucinates less than a 4B for semantic to structural matching. Small llms for targeted tasks for the win. The new Meta llama3.2 1B is better than expected and cheap on token cost.

3

World models ≠ LLMs ≠ Foundation models People use these three terms interchangeably. They shouldn't. The conflation is causing real confusion—and misplaced expectations about what AI can and can't do. Foundation models are large models trained on broad data, then fine-tuned for specific tasks. BERT was an early example. GPT-4 is a recent one. So are protein language models trained on sequence databases. LLMs are foundation models trained on text. A subset. All LLMs are foundation models. Not all foundation models are LLMs. World models are something different entirely. A world model is an internal representation that captures causal structure and supports counterfactual reasoning. It's a functional property, not an architecture. A system has a world model if it can answer: "What would happen if I did X, even though I've never seen anyone do X?" That last part is crucial. World models can extrapolate beyond their training data because they encode structure, not just statistics. They support intervention, not just prediction. The question "do LLMs have world models?" conflates architecture with function. Can LLMs generate plausible text about cause and effect? Yes. Can they reliably predict the consequences of interventions they haven't seen? No. Last year, Harvard and MIT researchers trained transformers on 10 million simulated solar systems. The models made accurate trajectory predictions—but had learned completely wrong force laws. Case-specific heuristics that don't generalise. Accurate interpolation, no understanding. Ptolemy's epicycles also fit planetary motion beautifully. They just happened to be completely wrong. Kepler's laws fit the data and reflected underlying structure. That's the difference between curve-fitting and science. Yann LeCun has been warning for years that LLMs lack causal world models, so genuine reasoning cannot emerge from scaling them. If we think our LLM has a world model and trust its predictions in novel situations we'll be disappointed. If we think scaling will produce causal understanding, we're making a category error. Statistical patterns are not causal structure. For AI in science this matters especially. Scientific questions are often counterfactual: What happens if I knock out this gene? Change this reaction rate? Give this drug to this patient? A model trained on observational data learns what has happened. Not what would happen under conditions outside the training set. That requires structure—a model that extrapolates because it has learned how the system works, not merely what the system has done. None of this means these tools are useless. Foundation models are powerful. LLMs have transformed how we interact with textual and sequential information. Specialised world models like DreamerV3 show impressive cross-domain transfer. But using the wrong tool—or expecting capabilities it doesn't have—wastes resources and erodes trust. Three terms. Three different meanings. Worth keeping straight.

63

5 Comments

A dirty secret about proteomics mass spec data is that we have no idea what most of the signals in the data are and we pretty much ignore them unless they can be confidently identified by the processing software. This means even if an unidentified analyte shows extremely high correlation with the biological process we are studying, we are completely blind to it and miss out on those potentially valuable biomarkers or targets. This newly published approach from Golden Haystack flips the paradigm of "Identify then Quantify" to "Quantify then Identify", unlocking 10x more features in the data. Researchers have sought to do this since data independent acquisition (DIA) was introduced, but this type of spectrum-centric analysis is computationally intensive. That's why Gautam Saxena was the perfect person to tackle the problem. His background in computer engineering enabled him to develop a more efficient algorithm that not only enables this type of processing, but achieves faster processing times than current peptide-centric approaches. I can't wait to see the impact this new tool has on the field. It's time to rethink how you process your proteomics data...

84

3 Comments