Demis Hassabis

The authors present LERD, a novel Bayesian neural dynamical framework designed to improve diagnosis and monitoring of neurodegenerative diseases like Alzheimer’s using multichannel EEG data. Traditional machine learning approaches in this domain often treat classifiers as “black boxes” and fail to explicitly model the underlying neural dynamics generating the observed signals. In contrast, LERD infers latent neural events and their relational structure directly from EEG without requiring external event or interaction labels, combining a continuous-time event inference module with a stochastic event generation process. An electrophysiology-inspired dynamical prior guides the learning process, promoting physiologically meaningful representations. The paper also contributes a theoretical analysis that provides a tractable training bound and stability guarantees for the inferred latent relational dynamics. Extensive experiments on synthetic data and two real Alzheimer’s EEG cohorts show LERD consistently outperforms strong baseline methods, while yielding latent summaries that align with known physiological characteristics of the disease. These results suggest that modeling latent event interactions and dynamics offers a more interpretable and accurate approach for EEG-based neurodegenerative classification than traditional black-box classifiers. https://lnkd.in/g-nH6tyT

2

Dynamic scene reconstruction from single-camera videos remains challenging for neural radiance field (NeRF) models, especially when generating novel views far outside the original camera’s viewpoint. Traditional dynamic NeRF approaches often fail under large viewpoint changes, producing unstable or unrealistic renderings that do not generalize well beyond the constrained input motion. To address this limitation, the authors introduce ExpanDyNeRF, a novel monocular NeRF framework that combines Gaussian splatting priors with a pseudo-ground-truth generation strategy to support reliable view synthesis under extreme rotations and large angular deviations. ExpanDyNeRF jointly optimizes density and color features to improve scene representation quality from difficult perspectives that would otherwise degrade reconstruction fidelity. In addition to the model, the paper presents SynDM, the Synthetic Dynamic Multiview dataset—the first synthetic multiview dataset for dynamic scenes with explicit side-view supervision, created using a customized rendering pipeline based on a videogame engine. Experiments on both SynDM and real-world data demonstrate that ExpanDyNeRF consistently outperforms existing dynamic NeRF baselines in rendering fidelity and stability under large viewpoint shifts, validating its effectiveness for broadening view synthesis in dynamic environments. The supplementary materials provide additional quantitative and qualitative evaluations that further support these findings. https://lnkd.in/gTe29awy

1

If you're running on Modal and would like to quickly try out profiling tools like Torch Profile and understand your AI system's performance - all within your VSCode / Cursor environment, check out the tutorial we've put together on using the nCompass VSCode/Cursor extension with remote GPUs. Comments have a link to a detailed tutorial on how to get started with our tooling on Modal. #MLSys #Profiling #AIPerformance #Modal #nCompass

24

1 Comment

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

Building on LLMs? The hardest problem isn't getting "good" output. It's getting reproducible output every single time. LLMs are stochastic by nature. But production systems need guarantees. The solution is Stochastic Determinism: the engineering goal of constraining a probabilistic system until its variance is intentional, reproducible and acceptable. While building LlamaBrain (an open-source LLM governance framework), my architecture looked flawless on paper. I did some stress-testing on it with a real-world app and hit 96.7% consistency. Sounds high, right? It isn't. A 3.3% failure rate means your constraints break roughly 1 in 30 calls. At scale, that’s a systematic failure mode. Getting from 96.7% to 100% took ten days and required building a Black Box Audit Recorder to log, replay, and structurally compare every interaction. The lesson? Determinism isn't something you just design into an AI system. It's a property you must instrument, measure, and prove. Anything less is just a toy framework. #LLM #SoftwareEngineering #AI #OpenSource #LlamaBrain

5

2 Comments

After a very public disagreement and split from Meta, Yann LeCun (who believes LLMs are a dead end in the search for AGI) has now set up a new company to build World Models. And he just raised over USD 1 Billion, one of the largest "seed" rounds in history! What are World Models and how will he build them? That's the AI concept explained in today's Deep Tech Stars Newsletter. Plus, Microsoft launches Copilot Cowork, which is basically Claude Code for all enterprise work. And some really awesome remote AI jobs. https://lnkd.in/ghnq5xez

17

ISTS Protocol: A Technical Deep Dive into Isotropic Spatiotemporal Tensor-Spline Compression Following my recent announcement, I am sharing an explanatory video that demonstrates the core mechanics of the ISTS Protocol in action. The ISTS algorithm introduces a dual-path architecture that fundamentally changes how we handle data density: Text-Based Atomization: The video highlights the real-time breakdown of raw text into structural components, which are then deduplicated via a Global Ledger. This ensures that redundant patterns occupy zero additional space beyond their initial coefficient. Vectorized Visual Storage: Watch the transition from raw image data to high-efficiency WebP Base64 strings, followed by secondary LZW bit-packing. Fibonacci Sphere Distribution: See how we use a 3D ASCII-based plotting system to visualize data clusters, providing a geometric representation of compressed information. By combining tensor-spline logic with bit-level compression, ISTS achieves significant storage optimization for local IndexedDB environments without sacrificing retrieval speed. Watch the full walkthrough below to see how these layers integrate. Technical Documentation: https://lnkd.in/ghPqEd4j

1

We cut the hallucinations of an LLM in half using a new technique: RLFR (Reinforcement Learning from Feature Rewards). RLFR reduces hallucinations in Gemma 12B by 58% (when run with our probing harness), for ~90x cheaper per intervention than the LLM-as-judge alternative. The technique uses lightweight probes on a model's internal representations as reward signals for reinforcement learning. This is all while avoiding: - off-target effects (no degraded performance on standard benchmarks) - degrading/reward-hacking the probes (the probes still work as monitors at test time!) This is just our first public demonstration of our work building towards "intentional design": the ability to intelligently guide gradient descent, enabling a new paradigm of far more precise, robust, and effective training. Kudos to Aaditya Prasad, Connor Watts, Jack Merullo, Dhruvil Gala, Owen Lewis, Tom McGrath, and Ekdeep Singh for this groundbreaking new work!

236

9 Comments

Interesting perspectives on scaling, RL and what the current AI approaches struggle with. Back to the age of research (or AI winter 😅) https://lnkd.in/dZvZh2Ta #AI #scaling

13

4 Comments

I've become interested in mechanistic interpretability recently and wanted to experiment with some of the latest innovations in the space. In a limited fashion, we can look inside a LLM and see what's going on. Using Sparse Autoencoders (SAEs) from Gemma Scope, this tool lets you peek under the hood of Google's Gemma 2 2B model and actually steer its behavior in real time. You can: - Browse and search 16K+ interpretable features by description - Visualize which features activate on any input text - Amplify or suppress specific features during generation and see how outputs change - Decompose predictions into per-feature logit contributions Built with TransformerLens, SAELens, and Gradio. Try it on Huggingface Spaces: https://lnkd.in/e88v8wym Code here: https://lnkd.in/e7dSS9WA #MechanisticInterpretability #AISafety #MachineLearning #NLP #OpenSource

6