arXiv:2603.12296v1 Announce Type: new Abstract: Deep learning has achieved transformative performance across diverse domains, largely driven by the large-scale, high-quality training data. In contrast, the development of brain-computer interfaces (BCIs) is fundamentally constrained by the limited, heterogeneous, and privacy-sensitive neural recordings. Generating synthetic yet physiologically plausible brain signals has therefore emerged as a compelling way to mitigate data scarcity and enhance model capacity. This survey provides a comprehensive review of brain signal generation for BCIs, covering methodological taxonomies, […]

Ver mais

Is language modeling fundamentally token-level or sequence-level? There is evidence for both: pretraining and sampling lean towards a token-level view, while alignment is fundamentally sequence-level. Curious if there is any work trying to unify the two perspectives, and which is the more principled framing. Pretraining Textbook language modeling defines the task as learning a distribution over strings, but all cross-entropy loss implementations I’ve seen operate at the token level. The difference is subtle but real: both compute sum […]

Ver mais

arXiv:2601.02400v1 Announce Type: cross Abstract: Text-based causal inference increasingly employs textual data as proxies for unobserved confounders, yet this approach introduces a previously undertheorized source of bias: treatment leakage. Treatment leakage occurs when text intended to capture confounding information also contains signals predictive of treatment status, thereby inducing post-treatment bias in causal estimates. Critically, this problem can arise even when documents precede treatment assignment, as authors may employ future-referencing language that anticipates subsequent interventions. Despite growing recognition of […]

Ver mais

I’ve casually seen LLMs correctly verify exact quotations that they either couldn’t or wouldn’t quote directly for me. I’m aware that they’re trained to avoid quoting potentially copywritten content, and the implications of that, but it made me wonder a few things: Can LLMs verify knowledge more (or less) accurately than they can recall knowledge? 1b. Can LLMs verify more (or less) knowledge accurately than they can recall accurately? What research exists into LLM accuracy in recalling facts […]

Ver mais

For years, emerging projects were competing for distribution and attention with 2 types of brands: large mature businesses with unlimited marketing budgets (think, Binance, Coinbase, or Ethereum) and new projects with a growth-hacking mindset pulling guerrilla stunts (e.g., Grass, Cluely). But, in 2026, the most powerful distribution channel won’t belong to any of them – it will be owned by governments. Here’s why. The crypto world used to run on one simple emotional loop: greed and fear. Greed […]

Ver mais

Happy to share the latest colqwen3.5-4.5B model in the series. ColQwen3.5-4.5B-v3 is #1 (avg) on the MTEB ViDoRe leaderboard (Pending release) at 75.67 mean, ~half the params, ~13x fewer embedding dims, ~half the memory footprint of the previous #1 model. Thoughts: V3 edges out v2 on V3 English u@5 (0.6034 vs 0.6023), a marginal gain for substantially more compute. The real win was the V2 benchmark jump and surpassing 8B models on V3. That’s where I decided to […]

Ver mais

Hi all, wanted to share the project I’ve been working on: Volga — an open-source data engine for real-time AI/ML. In short, it is a Flink/Spark/Arroyo alternative tailored for AI/ML pipelines, similar to systems like Chronon and OpenMLDB. I’ve recently completed a full rewrite of the system, moving from a Python+Ray prototype to a native Rust core. The goal was to build a truly standalone runtime that eliminates the “infrastructure tax” of traditional JVM-based stacks. Volga is built […]

Ver mais

Machine learning (ML) offers transformative potential for computational fluid dynamics (CFD), promising to accelerate simulations, improve turbulence modelling, and enable real-time flow prediction and control-capabilities that could fundamentally change how engineers approach fluid dynamics problems. However, the exploration of ML in fluid dynamics is critically hampered by the scarcity of large, diverse, and high-fidelity datasets suitable for training robust models. This limitation is particularly acute for highly turbulent flows, which dominate practical engineering applications yet remain computationally prohibitive […]

Ver mais