A 9B local model on my laptop scored 45.56% on Aider Polyglot. The same exact weights in vanilla Aider scored 19.11%. Same model, different scaffold. We may be underestimating small coding models by evaluating them in the wrong wrapper. https://t.co/fODjFy6wWk
Excited to share my first paper: Model–Behavior Alignment under Flexible Evaluation: When the Best-Fitting Model Isn’t the Right One (NeurIPS 2025). link below.
Not all neuroscience is fun and games, but some of it is 👀🕹️
I created a mini-games arcade you can play with your thoughts, right in your browser.
It’s open-source, and built so anyone can contribute a game, regardless of their neuroscience knowledge:
https://t.co/f80oG44XNq
Informal j-club at OIST: Dissociating Artificial Intelligence from Artificial Consciousness
https://t.co/aUEyunuz9a
We asked big questions about artificial consciousness. What would it mean? How would we test for it? What does IIT say about it?
This study finds that the visual system is tuned to the specific speed and duration of saccades, so much so that it filters out anything that looks like them.
A law of vision grounded in action.
https://t.co/gl0nhCpqU4
Impressive work, but representational alignment is tricky. Sometimes preserving global geometry is ideal, other times, distinctions matter more.
Philosophically, computational theories demand a more precise notion of representation than alignment alone.
https://t.co/v7AuVfdo9b
Incredible dataset and elegant analysis! Would love to see efforts toward making parts of this dataset openly accessible.
Even anonymized versions of this data could be a goldmine for perception research.
https://t.co/Ob0N11pwfs
Just had the pleasure of joining a panel on academia vs. industry hosted by the neuroscience department at Ben-Gurion University.
Grateful for the invitation and the thoughtful discussion about navigating career paths, and building things that matter.
This new study shows ~0.6mm of error can stem from consistent misalignment in MNI space, detectable using anatomical fiducials (AFIDs).
https://t.co/udGxEZ9mg6
This elegant study shows that eIF2B5 mutations in oligodendrocytes drive the ataxia phenotype, despite astrocytes showing hallmark pathology.
https://t.co/XbQrJWDWXb
Today, the New England Journal of Medicine published what may be a landmark moment: doctors used customized gene editing inside a living baby to treat a deadly genetic disease called carbamoyl-phosphate synthetase 1 (CPS1) deficiency. Is it time to prepare for a world where we don’t just treat genetic diseases, we rewrite their story? Will gene editing change the future of medicine?
Key Points for the Public:
- Think of it as the first real-life genetic fire rescue; editing a baby’s DNA to prevent brain damage and death.
- They used a tool called base editing, that is considered a precise cousin of CRISPR.
- The authors corrected a disease-causing mutation in the liver using lipid nanoparticles.
- You may recognize lipid nanoparticles as the same delivery technology used in mRNA COVID vaccines.
- Their treatment was "tailor-made for a single patient."
- The personalized gene-editing drug called k-abe was made in under 6 months (incredible!).
- The single infant they used it on had a rare and fatal genetic disorder.
- So far, the treatment has been safe and effective though followup has not been very long.
- The infant tolerated more protein, required fewer medications, and recovered from infections without ammonia spikes (known to be bad for the brain).
My take: This isn’t science fiction anymore, gene editing can be performed in real-time. Gene editing is moving beyond sickle cell. Lipid nanoparticle delivery facilitates re-dosing, making it safer and more flexible. What does this mean for the future? Neurogenetic diseases could be next in line? Could we start moving into rare brain conditions in children or even adult-onset disorders like Huntington’s or inherited forms of Parkinson’s? What was cool about this study was moving from diagnosis to therapy, in just months. The platforms like this one seem to be scalable. Many diseases could share delivery systems with only the guiding RNA changing. Ethics and oversight will be critical. Safety, long-term monitoring, and clear boundaries (e.g., no germline edits) must be built into the future of this therapy. Could this change the arc of progression for many diseases? We hope this technology and approach will move from lethal newborn disorders to slowly progressive neurologic conditions. Has gene editing moved past the dream stage? Is it time to prepare for a world where we don’t just treat genetic diseases, we rewrite their story? https://t.co/1jHOp6ziiB #GeneEditing #CRISPR #Neurology #RareDisease #BaseEditing #PrecisionMedicine #NEJM #Parkinsons #Neurogenetics @ParkinsonDotOrg@FixelInstitute
3/
This could be the opening of a richer diagnostic pipeline where blood guides classification and neurotech guides personalization. What do you think? 🧠
1/
The FDA’s clearance of an Alzheimer’s blood test is huge, but it doesn’t mean other tools are obsolete.
If anything, it reveals how much we still need to understand about how pathology meets behavior.
https://t.co/TYKwO2NXO6
2/
We know people with high amyloid but no symptoms, and others with fast decline and “normal” scans.
That mismatch is where multimodal neurotech shines: tracking cognition, affect, attention, and more.
🧵Ketamine and Dissociation - What’s the Mechanism? 🚨1/7
A real-world 2025 EEG study in bipolar depression patients reveals a cascade of changes across brain rhythms, complexity, and excitation-inhibition (E/I) balance.
Let’s break it down: 👇1/7
Neural markers (like N1 amplitude differences) say more about your listening-in-noise ability than behavior does.
Should hearing diagnostics move beyond the audiogram and include cognitive tests or EEG-based attention measures?
https://t.co/BhTw6Og2JQ
A new study shows lipoamide dissolves stress granules in ALS via redox modulation of the methionine-rich SFPQ.
Can redox be a generalized handle on phase-separated biology? 👀