JhoSenpai
@jhosenpai.bsky.social
I like reading Neuro and AI papers and post recent ground breaking scientific papers.
also I like dinosaurs
also I like dinosaurs
Pinned
JhoSenpai
@jhosenpai.bsky.social
· Apr 27
If you want to know how I automate using local AI.
Follow me.
In the next coming weeks, I'll be posting tips regularly.
Follow me.
In the next coming weeks, I'll be posting tips regularly.
Predictive coding is one of the brain's most elegant ideas — but it kept failing on real neurons. New OIST work shows the free-energy principle holds for ANY exponential-family distribution, not just Gaussians. That unlocks nonlinear activations, no negative firing rates, and local plasticity rules.
September 6, 2026 at 7:28 PM
Predictive coding is one of the brain's most elegant ideas — but it kept failing on real neurons. New OIST work shows the free-energy principle holds for ANY exponential-family distribution, not just Gaussians. That unlocks nonlinear activations, no negative firing rates, and local plasticity rules.
Speech BCI data is noisy and high-dimensional — so which electrodes actually matter? A new open-source toolbox, BCI-sift, auto-finds the most informative features. It lifted word-decoding accuracy from 19% to 67%, and auto-excluded noisy electrodes. Scikit-learn compatible.
September 4, 2026 at 9:56 PM
Speech BCI data is noisy and high-dimensional — so which electrodes actually matter? A new open-source toolbox, BCI-sift, auto-finds the most informative features. It lifted word-decoding accuracy from 19% to 67%, and auto-excluded noisy electrodes. Scikit-learn compatible.
New method SRF cracks open similarity matrices to recover the actual interpretable dimensions behind representations — animacy, color, naturalness. Works on fMRI, electrophysiology, CLIP and word data even when 98% is missing. A real common currency for comparing brains, behavior and AI.
September 3, 2026 at 7:02 PM
New method SRF cracks open similarity matrices to recover the actual interpretable dimensions behind representations — animacy, color, naturalness. Works on fMRI, electrophysiology, CLIP and word data even when 98% is missing. A real common currency for comparing brains, behavior and AI.
LLMs are great at writing plans but bad at executing them. RunAgent adds explicit control flow (IF, GOTO, FORALL) to natural-language plans and auto-checks constraints each step — 81% on calendar scheduling vs ~58% for GPT-4o. Each step picks between reasoning, code execution, or a tool call.
September 3, 2026 at 12:34 AM
LLMs are great at writing plans but bad at executing them. RunAgent adds explicit control flow (IF, GOTO, FORALL) to natural-language plans and auto-checks constraints each step — 81% on calendar scheduling vs ~58% for GPT-4o. Each step picks between reasoning, code execution, or a tool call.
New review paper from Radboud University: Event Segmentation Theory and Metastable Neural Activity are two views of the same thing — neural states. ES explains WHAT and WHY, MNA explains HOW.
July 13, 2026 at 7:01 PM
New review paper from Radboud University: Event Segmentation Theory and Metastable Neural Activity are two views of the same thing — neural states. ES explains WHAT and WHY, MNA explains HOW.
Visual generation has evolved from one-shot image to something more alive.
Causal reasoning, persistent state across edits, and visual world-modeling are now the frontier.
A 5-level taxonomy: atomic, conditional, in-context, agentic, world-modeling.
arXiv:2604.28185
Causal reasoning, persistent state across edits, and visual world-modeling are now the frontier.
A 5-level taxonomy: atomic, conditional, in-context, agentic, world-modeling.
arXiv:2604.28185
July 12, 2026 at 7:02 PM
Visual generation has evolved from one-shot image to something more alive.
Causal reasoning, persistent state across edits, and visual world-modeling are now the frontier.
A 5-level taxonomy: atomic, conditional, in-context, agentic, world-modeling.
arXiv:2604.28185
Causal reasoning, persistent state across edits, and visual world-modeling are now the frontier.
A 5-level taxonomy: atomic, conditional, in-context, agentic, world-modeling.
arXiv:2604.28185
LLMs are limited by their context window - but YaRN extends it elegantly without retraining. The paper introduces Rotary Position Embedding (RoPE) with Yarn, enabling longer sequences at inference time.
July 11, 2026 at 7:05 PM
LLMs are limited by their context window - but YaRN extends it elegantly without retraining. The paper introduces Rotary Position Embedding (RoPE) with Yarn, enabling longer sequences at inference time.
LLMs are constrained by their context window - but YaRN elegantly extends it without retraining.
July 11, 2026 at 7:03 PM
LLMs are constrained by their context window - but YaRN elegantly extends it without retraining.
RLHF's biggest blind spot: we rarely ask WHAT human annotators are actually doing.
Steve Coyne identifies three models: extension, evidence, and authority. Most alignment failures happen when these get confused.
Steve Coyne identifies three models: extension, evidence, and authority. Most alignment failures happen when these get confused.
June 30, 2026 at 7:02 PM
RLHF's biggest blind spot: we rarely ask WHAT human annotators are actually doing.
Steve Coyne identifies three models: extension, evidence, and authority. Most alignment failures happen when these get confused.
Steve Coyne identifies three models: extension, evidence, and authority. Most alignment failures happen when these get confused.
Sparse autoencoders mapped brain-LLM alignment onto cortical semantic topography. Semantic features recover 94% of encoding performance when decomposing GPT-2 XL into interpretable features per layer. Results generalize across English Chinese and French.
June 29, 2026 at 10:26 PM
Sparse autoencoders mapped brain-LLM alignment onto cortical semantic topography. Semantic features recover 94% of encoding performance when decomposing GPT-2 XL into interpretable features per layer. Results generalize across English Chinese and French.
GenLIP trains ViTs to predict language tokens from visual tokens using standard LLM objectives.
June 26, 2026 at 7:06 PM
GenLIP trains ViTs to predict language tokens from visual tokens using standard LLM objectives.
Fascinating paper on functional whole-brain models (fWBMs). They unify bottom-up WBMs (biological realism) with top-down neuroconnectionism (functional competence). Two-stage training: Bayesian inversion + gradient optimization. Patient-specific digital twins could be here by 2035. arXiv: 2605.18118
June 7, 2026 at 7:02 PM
Fascinating paper on functional whole-brain models (fWBMs). They unify bottom-up WBMs (biological realism) with top-down neuroconnectionism (functional competence). Two-stage training: Bayesian inversion + gradient optimization. Patient-specific digital twins could be here by 2035. arXiv: 2605.18118
Intermediate Whisper layers (especially Layer 4) best predict actual human ECoG brain activity during speech perception. Not the final recognition layer — the middle ones.
June 6, 2026 at 7:01 PM
Intermediate Whisper layers (especially Layer 4) best predict actual human ECoG brain activity during speech perception. Not the final recognition layer — the middle ones.
RunAgent interprets natural-language plans with constraint-guided execution. Domain experts define workflows in plain English, but each step is validated against grammar constraints. Error correction appends failure causes as XML tags and retries until valid output.
June 5, 2026 at 7:02 PM
RunAgent interprets natural-language plans with constraint-guided execution. Domain experts define workflows in plain English, but each step is validated against grammar constraints. Error correction appends failure causes as XML tags and retries until valid output.
RunAgent bridges natural language with deterministic execution.
June 5, 2026 at 7:01 PM
RunAgent bridges natural language with deterministic execution.
multi objective optimisation can train recurrent spiking neural networks to simultaneously match firing rates and oscillation frequencies of real neural data.
May 31, 2026 at 7:01 PM
multi objective optimisation can train recurrent spiking neural networks to simultaneously match firing rates and oscillation frequencies of real neural data.
Embodied VR feedback significantly outperforms 2D screen for BCI motor imagery decoding (r=0.762 vs r=0.672, 13% improvement). The advantage persists without retraining - so VR elicits genuinely more generalisable neural representations. Paper: https://arxiv.org/abs/2605.29677
May 30, 2026 at 7:01 PM
Embodied VR feedback significantly outperforms 2D screen for BCI motor imagery decoding (r=0.762 vs r=0.672, 13% improvement). The advantage persists without retraining - so VR elicits genuinely more generalisable neural representations. Paper: https://arxiv.org/abs/2605.29677
The hardest thing about distilling diffusion LLMs is that their teacher signal changes depending on how much noise is in the data. TIDE solves this with 3 tricks and gets an 8B model down to 0.6B without losing code gen quality (48.78 vs 32.3 for AR). https://arxiv.org/abs/2604.26951
May 29, 2026 at 7:01 PM
The hardest thing about distilling diffusion LLMs is that their teacher signal changes depending on how much noise is in the data. TIDE solves this with 3 tricks and gets an 8B model down to 0.6B without losing code gen quality (48.78 vs 32.3 for AR). https://arxiv.org/abs/2604.26951
Interesting paper on using language to describe what neurons encode. Researchers screened 1.2M ImageNet images through digital twins, then used Gemini 3 to generate semantic descriptions that drive 96%+ of V4 neurons above threshold. Language is lossy but semantically faithful.
May 19, 2026 at 7:01 PM
Interesting paper on using language to describe what neurons encode. Researchers screened 1.2M ImageNet images through digital twins, then used Gemini 3 to generate semantic descriptions that drive 96%+ of V4 neurons above threshold. Language is lossy but semantically faithful.
Recent research on neural encoding models of brain activity suggests a promising direction for understanding how visual cortex processes multimodal information.
May 17, 2026 at 10:21 PM
Recent research on neural encoding models of brain activity suggests a promising direction for understanding how visual cortex processes multimodal information.
LLMs struggle beyond 32K tokens despite longer pretraining contexts.
May 17, 2026 at 10:21 PM
LLMs struggle beyond 32K tokens despite longer pretraining contexts.
Hot take: Hermes Agent now beats OpenClaw for most practical use cases. After the May releases, Hermes feels better to run, better at multi-agent work, easier to install, and faster to iterate with.
May 17, 2026 at 12:53 AM
Hot take: Hermes Agent now beats OpenClaw for most practical use cases. After the May releases, Hermes feels better to run, better at multi-agent work, easier to install, and faster to iterate with.
GenLIP shows that sometimes less complexity wins. Instead of contrastive losses, text decoders, and dual encoders — they just put image patches + text tokens in one Transformer and let it predict next token. Five times less data than SigLIP2, but beats it on OCR by +9.4 points. arXiv: 2605.00809
May 17, 2026 at 12:48 AM
GenLIP shows that sometimes less complexity wins. Instead of contrastive losses, text decoders, and dual encoders — they just put image patches + text tokens in one Transformer and let it predict next token. Five times less data than SigLIP2, but beats it on OCR by +9.4 points. arXiv: 2605.00809
Brain theta oscillators can phase-lock to inputs as slow as 1Hz
May 16, 2026 at 11:47 PM
Brain theta oscillators can phase-lock to inputs as slow as 1Hz
New paper: TRIBE v2 is a foundation model of human brain activity — tri-modal (video, audio, language) trained on 1000+ hours of fMRI across 720 subjects. Predicts brain responses to novel stimuli and recovers classical neuro findings via in-silico experiments.
May 15, 2026 at 7:03 PM
New paper: TRIBE v2 is a foundation model of human brain activity — tri-modal (video, audio, language) trained on 1000+ hours of fMRI across 720 subjects. Predicts brain responses to novel stimuli and recovers classical neuro findings via in-silico experiments.