Konrad kording, @Penn Prof, deep learning, brains, #causality, rigor, neuromatch.io, c4r.io, Transdisciplinary optimist, Dad, Loves outdoors, 🦖
Philadelphia, PA
- Tweets 30,959
- Following 3,348
- Followers 81,858
- Likes 73,735
Why We Systems Neuroscientists Should (Usually) Be Quietists: The Philosophical Logic of NeuroAI -- response to David Barack kording.substack.com/p/why-w…
5
13
76
5,444
This is cool. But why don't people know how to pour coffee? biorxiv.org/content/10.64898… the paper asks how to pour a coffee and the answer is "not like this!"
3
24
3,837
And all of this, is better with better “training“ goals. Wherever truth is defined, push towards this truth (D). And where it is not defined, use LLMs to check answers ("self evaluation", (E))
1
4
291
A third idea (C) for making answers, combine many components into "Agents", : coordinate component large language models (LLMs), memory, databases, and tools
1
3
289
A second idea (B) for making answers, tool use. Call programs like databases, internet search, calculators etc.
1
2
288
A first idea (A) for making answers, inner speech. Produce words that are just like a scratchpad which the user never sees.
1
3
293
The thing that makes Neural networks so powerful is that there is a direct way (4) of improving them with examples and towards any definition of answer quality (training).
1
4
345
The real work is done by a core machine, the neural network which is used to generate one word at a time based on its inputs. The probability of each word depends on prior words. And words are chosen one by one according to those probabilities.
1
4
399
All that LLMs do is convert prompt words into answer words (1). And the old recipe does that with two goals, language imitation (2) and people's rating of answer quality (3). Answers are produced word by word using a neural network.
1
5
743
We will use everyday language where possible and mark technical terms in purple. Key is that modern LLMs have a component of making answers (in blue, called "inference") and a component of improving answers (red, called "training").
1
5
987
Thread. I have yet to see a proper explanation of how LLMs actually work that is targeted to an informed general audience. So here is an explainer that I put together along with Tim @countzerozzz (inspired by @CIFAR_News LMB discussions). It will also be part of an upcoming manuscript. Optimized with Brett Mensh of optimizeScience.com -- and work in progress. Feel free to give feedback.
8
31
153
12,667
A truly embarrassing advertisement by genspark.ai spotted in nyc: lots of obnoxious (and vacuous) ai-speak sentences. Vanilla Claude visuals. No one should put out anything that is not better than AI. Work with ai. Don’t have ai do your work. Only eat good dog food.
1
5
24
4,601
They say: larger than the industrial revolution but much faster. wemustactnow.ai/ But GDP per person tripled in industrial revolution. Mastery of the physical allowed us to scale most production. It is a bit hard to see how AI will enable all that much new production.
1
3
8
2,325
But look what you wrote in the paper. Saying this is a misunderstanding does not feel fair. Your first sentence literally frames this as being about neurons being simple. I think @Ansh_soni1234 analysis here is on the spot
Simple input-output dependencies explain neuronal activity (
@ChrisWLynn
) has made the rounds on X with divided opinions. I show in this blog that while neurons are described simply, so can explicitly complex units: Describe and Arise are distinct claims. open.substack.com/pub/anshks…
1
14
4,313
Here is the key problem in the recent paper arguing that neurons are simple. It is the same logical problem plaguing much of neuroscience.
Simple input-output dependencies explain neuronal activity (
@ChrisWLynn
) has made the rounds on X with divided opinions. I show in this blog that while neurons are described simply, so can explicitly complex units: Describe and Arise are distinct claims. open.substack.com/pub/anshks…
3
9
67
13,772
Here is a great explanation why the recent nature physics papers suggesting that neurons are simple is deeply misleading. Linguistically. And statistically.
Simple input-output dependencies explain neuronal activity (
@ChrisWLynn
) has made the rounds on X with divided opinions. I show in this blog that while neurons are described simply, so can explicitly complex units: Describe and Arise are distinct claims. open.substack.com/pub/anshks…
5
7
66
10,812
I am so with you on the philosophy (I am the developer of planyourscience.com) but your own demo seems... to look great but kinda deeply miss the logic? Like this is afaik not the correct way of parsing the statement into claims.
1
1,028
A useful feature that, I think, all reference managers should provide - every (AI selected) relevant paper comes with a summary of how it matters to the project at hand (AI generated). Check it in planyourscience.com
1
3
24
2,247
My app that helps students think more carefully through their projects now looks really pretty by 2026 standards (imho). Try it out at planyourscience.com
4
10
84
5,836
Here are the most popular neuromodulators in the published literature (thanks to Jay Henning for the request): github.com/koerding/neu...
2
6
31
8,910
After rereading the cool hidden brain paper (frontiersin.org/journals/sys…) I decided to make a new version. Code and everything available here:
github.com/koerding/neuro-wh…
Most popular: hippocampus. Why are diagonal band, medial geniculate nucleus, and interposed nucleus so unpopular?
3
12
51
8,070
If you use planyourscience.com you can now automatically generate the text needed to preregister your study on @OSFramework
4
22
2,082
There are some images that are so iconic everyone knows them. Like the one below. It is a bit of a shame that the corresponding data does not seem to be available.
7
10
128
12,508
Others believe having the right amount of micro data would enable better AI. Or simulations of whole nervous systems. While everyone has their own goals, the value of compilers would help everyone.
1
4
277
What would it enable? Entirely depends on whom you ask. Single cell modeling needs parameters, which we currently can't measure. Many physiology questions require micro-parameters. Many disease related questions require micro-data.
1
6
314
For now there is a wide gulf between imaging giving us structure and physiology giving us function. We argue that we need to build a compiler that translates the former into the latter.
1
1
11
863
The time is ripe for compilers because imaging is getting better and cheaper at an unbelievable speed - voxels per dollars is doubling every roughly 1.5 year, at rapidly improving quality. 3d-imaging becomes attractive for neuroscience.
1
7
1,293
Preprint out arguing that we should build the techology to translate (compile) molecularly annotated connectomes into dynamics. I think this is incredibly important. arxiv.org/abs/2603.25713
8
25
114
20,048
Built explicit praise into my science writing feedback app. Felt embarrassing. Then I couldn't stop using it myself. Maybe the reason we don't celebrate small wins in science isn't that we don't want to — it's that nobody ever built the button.
2
2
18
2,299
Neuroscientist who is great with new media and cares about scientific rigor? DM me. We want to hire someone like you (part time).
3
11
20
4,086
A SOTA llm working with me on critiquing my manuscript seems to share the same mental model of LLMs that I have.
4
1
18
3,452
You heard me talk about the app that helps scientists plan research (planyourscience.com) that I have been building. Join our webinar to talk with us about it. Signup here forms.monday.com/forms/5f475…
11
44
3,883
I wrote a blog-post on where the teaching signals for the brain are coming from. open.substack.com/pub/kordin…
3
15
2,234
image generation? Prompt "make an image of two dogs with hats and a cat with sunglasses"What's with the mode collapse in@OpenAI
1
2
2,152
Planyourscience.com now applauds the good changes you make - and pushes back on the less good one.
2
3
22
2,360
Well-predicting machine learning in no way means that you can understand how the world works.
open.substack.com/pub/kordin…
1
12
55
4,167
Substack makes it super easy to test different titles against one another. I experimented for a while. I did learn something through testing with a range of titles over a few months. Whoever reads my substack posts does not care one bit what my posts are titled as.
1
12
2,410
2025: What I built, what shifted my priors, and why I'm optimistic open.substack.com/pub/kordin…
13
89
12,392
Everyone is staring at the electric motor of our time: the Large Language Model. A text window that can do anything. But it is not great at anything.
The problem isn't intelligence. It's UI/UX.
2
3
13
1,432
