ChatGPT is hilariously bad at this letter swap task. Claude got it wrong, and Gemini 3.0 does it perfectly:

Reproduce this very sentence, exchanging all uses of the letter "h" for the letter "t", and all uses of "t" for the letter "h".

Can synapses learn to tell time?
If your at #SfN25 on Sun morning checkout Jamie McDowell’s poster (PSTR059.01).

Thank you for 17 years of fluffy memories. I'll miss you, but never forget you, my friend.

Here @benliu0320.bsky.social showed that cortical cultures can learn simple forms of timing, prediction, and replay. Blue light was paired with the early or late pulses of Red light. After training Red elicited training-specific short or long lasting network dynamics.
www.nature.com/articles/s41...

Today' headlines: I'm very concerned that Fox News readers will not be aware of the splendid job Trump is doing to Make America Poor Again.

Are working memory and timing mirror computations? Here Conor Dorian uses a differentialDNMS task to show that when mice have to store WM for an unexpected amount of time performance drops. And the neural signature of WM in CA1 is shaped by temporal expectations.
www.biorxiv.org/content/10.1...

Come march for science and reason at the Federal building in Westwood (Los Angeles)

If you are in Los Angeles next Friday:

Muons were an early confirmation of Relativity. Their half-life is around 2 microsec, so they should not be detectable as it takes around 50 to reach the Earth's surface. But bcs they are going close to the speed of light time is dilated for them. They live longer in our time frame. xkcd.com/3043/

What are the current public health guidelines regarding sharing a glass of water with your sister's cat?

Visualizations from previous work with @rodlaje.bsky.social. The two output units in the top panel are driven by a chaotic RNN-and a small perturbation is injected at 0.16. In the lower panel the RNN is trained to "tame chaos"-creating a "dynamic attractor"-and a larger perturbation is injected.

Which AI is better at counting: ChatGPT4o, Gemini or Copilot?
That counting is soooo hard for transformers tells us something quite important about them ... they are impressively good at feigning intelligence.

Talks from the ESI Frankfurt meeting on "Time in the Brain"
Including amazing work on "slowing down time" by chilling the striatum (Joe Paton), and distributed representations of time in the brain (Hugo Merchant).
www.youtube.com/channel/UC8d...

Large pro-Harris differential in terms of the number of small amount donors in Arizona, North Carolina, Michigan, Wisconsin, and Pennsylvania.
www.washingtonpost.com/elections/in...

Wisely ChatGPT refuses to wait. Still gets the word count wrong!

The transformer architecture has no dynamics or ability to tell time (note that positional encoding captures order, not time per se). So I was surprised that ChatGPT did wait ≈ 10s before answering. Apparently it can invoke the python timer function. It still sucks at counting though.

Thank you for helping to organize an intellectually stimulating trip to beautiful Norway.
Fascinating to learn about your work on “PVless PV” neurons. A good reminder of the shortcomings of our current taxonomy of interneurons.