Ari Benjamin

Especially relevant as a counterexample to OpenAI’s Prism tool, which will only worsen the science slop problem.

The point isn’t to be anti-AI. It’s to demand a better philosophy behind the tools we build

This tool is extremely helpful and I think all scientists could benefit from it. Good at all stages of the research process, especially the very beginning.

It's also a great of example of how we *should* be building AI tools for science. Not to replace scientific thinking, but to hone it.

lol I know someone like this. Works on their ‘Civil Liberties & Privacy Team’ (I know). Believes they can learn on the inside so they can regulate once they leave. Also, enjoys an upper class lifestyle

Interesting – this is what you need for synaptic plasticity to solve "within-neuron" credit assignment. Backpropagating APs, elevating Ca2+ above the plasticity threshold, should preferentially enter dendrites which recently activated.

Is this the first evidence of this?

What if there are many cells of each cell type? In that case I’d argue evolution is better suited to moulding populations. In cortex, genes don’t specify each neuron’s connectivity individually, nor could they, with far more synapses than bits in the genome

love it. somehow we’ve convinced ourselves that it’s only worth paying ‘the best and brightest’ to be curious for a living. It’s fundamentally an exclusionary vision for science when it ought to be inclusionary.

Instead of an ivory tower, science is shaped like a root system.

Like in a religion, local ‘churches’ provide lifelong learning for adults. Yearly calendar of holidays celebrating the miracles of evolution since the precambrian. Professors become pastors.

More of this type of content please!

The logic of 'rising stars' programs is that science needs to retain top talent. But if anything, I've seen more brilliant minds leave science due to the culture of individualist careerism that these awards contribute to, and are a symptom of.

For example, if you spatially cluster the brain based on transcriptomics (e.g. www.biorxiv.org/content/10.1...) the clusters barely match functional areas.

IMO, data converges: we need to revisit the atlas.

But I agree broadly: brain tissue is far from homogenous, and we neglect it at our peril.

TissueFormer: a neural network for labeling tissue from grouped single-cell RNA profiles Single-cell RNA sequencing technologies have enabled unprecedented insights into gene expression and are poised to transform clinical diagnostics. At present, most computational approaches for interpr...

We've reanalyzed this dataset to predict functional brain areas from transcriptomic data, with some success (www.biorxiv.org/content/10.1...).

Yet if there is controversy around how areas are defined from neural recordings, cell diversity is equally as confusing a signal.... (1/2)

It is possible to work on neural networks as a neuroscientist and actively oppose how AI is developing and how it will affect society.

My confidence in that last sentence was less than 100%. But what was it? And if I had said "absolutely positive" somewhere, what then?

If you can't quantify natural language uncertainty, you can't train for it. At least explicitly. Thus all we have are RLHF approaches

True, but I think that's something different. Among the uncertainties flying around in Bayes stats – epistemic uncertainty over model weights, aleatoric uncertainty reflected in output probabilities, etc – none of them capture hedging in natural language. (1/2)

One hurdle: how do you reliably quantify the uncertainty expressed in a natural language paragraph?

If you gave me this measure for all data in the wild, and it was differentiable wrt the model, one might take a likelihood-max approach (plus calibration etc)

Consciousness science as a marketplace of rationalizations

my commentary on @smfleming.bsky.social and @matthiasmichel.bsky.social's thought-provoking BBS paper, and more generally about the field.

osf.io/preprints/ps...

Truly. ‘The use of state power for personal vendettas and returning personal favors’ is what we see daily. It’s plain corruption and I don’t know why that’s not the front message

Truly crazy - Texas groups are demanding absolute ideological alignment.

There are multiple TT job calls at Texas public universities right now. I will not be applying.

Under the hood, this
1) gets the avg (pseudobulk) gene expression of each type,
2) computes the similarity matrix,
3) projects expression into a 3D space using MDS, and
4) interprets the space as a perceptually uniform color space (LUV)

To make these colormaps, I made an easy python package:

```
!pip install colormycells

# Create a colormap based on cell type similarities
colors = get_colormap(adata, key="cell_type")

# Plot your cells
sc.pl.umap(adata, color="cell_type", palette=colors)
````
github.com/ZadorLaborat...

When colors reflect the actual differences between cell types, you can see things you wouldn't otherwise. For the mouse cortex, see how the layers form a rainbow, indicating a gradient of gene expression.

I also think this makes pretty plots 🌈 so that's nice.

Good scientific plotting uses color to convey meaning. We don't want to distract readers. Yet with default categorical maps,

(-) some cell types jump out for no reason
(-) perceptual similarity between colors is meaningless
(-) fewer colors than types, leading to repeats

Do you plot transcriptomic data, coloring each cell by its cell type? ~ Don't use a default colormap! ~

Instead, use colormaps that capture biological meaning. If two cell types are very similar, their colors should be similar too. Read on 🧵

🧬💻

This is real – Anthropic just agreed to a $1.5B class action settlement for authors of copyrighted works it stole. File your claim here: www.anthropiccopyrightsettlement.com

Unclear if this is just 'books' or journal articles too, but I'm putting mine in anyway
www.reuters.com/sustainabili...

I think this is the right view. But it’s worth adding that a large section of neuroscience defines specialization by what areas ‘encode for’, defined predictively.

Rather than the bell tolling for local specialization, it might be tolling for that ‘functional’ stim-response paradigm

what’s crazy is how they figured this out. makes for good science reading!

“Replication of an alien genome within one’s own cytoplasm echoes the endosymbiotic domestication of organelles [mitochondria]. Clonal males may thus be regarded as organelles at the superorganism level” 🤯

This interview of Monica López-Hidalgo by @analog-ashley.bsky.social is inspiring. All scientists’ job description should include this, great model

who is liable when an error is made by a BCI-using human with AI intervening in the output? The human? The BCI manufacturer? No one?

Get on my wavelength, man

Oh no how is it that "neuromodulation" means such different things across communities? Who let this happen?

For the record, the right answer is serotonin, etc.

The wrong answer is brain stimulation in a medical context, DBS, TMS etc. I will endlessly defend this turf.