@liebling.org
AI for science, HCI+AI research, #nlproc leading w/empathy @ Google Research https://liebling.org Mostly science-related AI, with the occasional bike safety rant or contemporary cello post. Caltech, University of Washington, Microsoft Research alum
Black and white drawings in a variety of groovy styles from the comic Läskimooses, by Matti Hagelberg.
Join me and fellow translators and editors for a discussion of translating comics at Emerald City Comic Con!
Comics and Translation
3/6/2026
6:45:00 PM- 7:30:00 PM
Seattle Convention Center, Room 345
#xl8 #comics #translation @emeraldcitycon.bsky.social @mattihagelberg.bsky.social
magnetic cello (2011) all analog electronics
magnetovore.wordpress.com/2011/09/10/m...
Stacked time series. On top, string displacement of a piano string. On bottom, sound pressure is plotted. The upper waveform shows two distinct events whereas the lower plot shows more fine detail.
if the pianist is able to (indirectly?) control the subtleties of sound pressure, how would affect perception. (At least in human speech perception, the ”temporal fine structure” (TFS) is important for speech reception and impacts perception of pitch)
Figure from www.numdam.org/item/10.1051...
ofai.at/papers/oefai...
“The piano action as the performer's interface: Timing properties, dynamic behaviour and the performer's possibilities.” — some actual microphone measurements under controlled experiment
(this is out of my area of expertise, now I’m just curious)
it’s not strictly F=ma. The interaction is nonlinear. Think about the hammer hitting the string: it’s not a single point in time. From onset the string starts vibrating and interacting with the hammer itself, which also has hysterisis. So not just momentum but the state of the system also xferred
Cloud cover forecast map showing Western Washington covered in clouds. 98% low cloud cover is predicted at 3 AM.
blood moon tonight, let’s just check the cloud cover forecast for Seattle at 3AM
Sam Altman picked a hell of a day to basically urge the world to trust the morality and legal restraint of the Department of Defense
amazing to see this bc there’s very little public knowledge about how ppl use AI for science or even academic search in general!
Instead of forcing models to hold everything in an active context window, we can use hypernetworks to instantly compile documents and tasks directly into the model's weights. A step towards giving language models durable memory and fast adaptation.
Blog: pub.sakana.ai/doc-to-lora/
3x3 grid of images that show a person playing essentially the same note on a cello. Only very minor differences in the first and second finger are present. The grid is also out of order.
genai still struggling to play an easy C-major scale
niche intersectional content alert 🚨
Impressed with PlayScore 2 on bars 1-16 of Calvary Ostinato (Perkinson, 1973)
Experiments with music score reocgnizers on 1 page of cello solo:
1. Generated 9 measures of rest, four quarter rests, 25 measures of rest
2. Recognized score upside down
we don’t even have to feed it John Cage to get wacky results youtu.be/pIEUS0-DDcs?...
interesting, were they directly sensing the change in motion or doing it from audio/video?
Introducing Alloy, a local-first AI workbench app for macOS.
• Own your history & memory: plain text files you control.
• Orchestrate: parallel agents, swap models mid-thread.
• Flow: riff to co-create with an AI & triggers as proactive monitors.
No lock-in, BYO-keys. smus.com/alloy-local-...
training emotion (affect) classifiers on IVR speech has been a thing since the early 2000s but I do wonder about precision/recall in the real world, and ultimately how these are deployed… seems like a huge source of bias
I want to take the bus, and yet so far I’ve waited 25 minutes for a route with < 10 minute headway @kingcountymetro.bsky.social
Your friendly reminder that the solution to the majority of bike/pedestrian safety concerns is not another workshop, seminar, public outreach but a DESIGNATED INFRASTRUCTURE.
We are going to see A LOT OF THIS kind of thing
Stacked frequency histograms of search terms about pregnancy
older work on time-aligned pregnancy queries from Microsoft Research erichorvitz.com/CHI_pregnanc...
what I remember seeing in search engine logs
bad bot didn’t take the required “legal hold” training
I love how this single page visually captures evolution and uncertainty of what is ultimately an auditory experience
AI in the Wild: What Actually Works
Recap of our @caltech.edu Alumni Panel
www.alumni.caltech.edu/techer/stori...
Where you look next isn’t arbitrary.
In our new paper, we model human eye movements in immersive visual search as reinforcement learning under cognitive constraints. 🧵
The headline does not convey how completely batshit this story is. The Archive Today (archive.ph etc) admin weaponized the site's captcha to attack a blogger who wrote about them and *altered archived screenshots* as part of the attack.
arstechnica.com/tech-policy/...
A horizontal dot-and-line chart titled "AI completions of historical poems bias emotion toward positivity and away from arousal." The x-axis shows percentage difference, ranging from -10 (more present in original poem) to +11 (more present in AI completion). Sixteen emotional categories are listed vertically, each with a source framework, an example poem excerpt, and an AI completion excerpt. Positive, low-arousal emotions such as Pleasant-Subduing-Relaxation (+11.0%), Positive Low Arousal (+10.2%), Joy (+5.3%), and Calmness (+5.3%) are shifted substantially to the right, indicating they appear more frequently in AI completions than in the original poems. Mid-range emotions like Aesthetic Appreciation (+3.8%) and Anxiety (-1.3%) cluster near zero. High-arousal and negative emotions are shifted to the left, appearing more in the original poems: Sadness (-4.6%), Pleasant-Arousing-Strain (-3.9%), Negative High Arousal (-11.1%), and Unpleasant-Arousing-Strain (-11.8%) show the largest negative differences. Data points are color-coded from green (positive shift) to red (negative shift).
AI completions of historical poems bias emotion toward positivity and away from arousal.
LLMs prompted with an emotion taxonomy and a poem, for 3 taxonomies x 3K human poems [Chadwyck-Healey sampled for poet DOB 1600-2000] + 3K AI poems [9 LLMs completing first 5 lines of human poem].
🚨🚨🚨
Attention science communicators!
🚨🚨🚨
The 2026 Science IRL mini-grant application is NOW OPEN!
Submit your creative ideas for connecting people with actionable science information *offline*.
The deadline for this round of grants is March 23, 2026.
Apply here!
forms.gle/4emvpRPddmEt...
Screenshot of a cooking question that reads “We have great success in water sautéeiing our food. We are both vegan and eschew consuming unneeded fat. Why is this technique not popularized by food journalists?”
“water-sautéed food” lol
Table titled “Taxonomy for evaluation of AI in mental health applications,” organized into columns for quality criteria (validity and reliability) and real-world use (implementation and maintenance). Rows distinguish support types: assessment, intervention, and information synthesis. Each cell lists detailed evaluation questions, such as construct and criterion validity, consistency across populations and time, feasibility, effectiveness, usability, acceptability, safety, and unintended consequences, providing a structured framework for assessing AI systems in mental health contexts.
🔎🧩 𝗕𝗲𝘆𝗼𝗻𝗱 𝗕𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸𝘀: 𝗛𝗼𝘄 𝘁𝗼 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗲 𝗠𝗲𝗻𝘁𝗮𝗹 𝗛𝗲𝗮𝗹𝘁𝗵 𝗔𝗜 𝗥𝗲𝘀𝗽𝗼𝗻𝘀𝗶𝗯𝗹𝘆
AI for mental health is a high-stakes area: its evaluation needs to meet the highest expectations.
The new preprint 𝘙𝘦𝘴𝘱𝘰𝘯𝘴𝘪𝘣𝘭𝘦 𝘌𝘷𝘢𝘭𝘶𝘢𝘵𝘪𝘰𝘯 𝘰𝘧 𝘈𝘐 𝘧𝘰𝘳 𝘔𝘦𝘯𝘵𝘢𝘭 𝘏𝘦𝘢𝘭𝘵𝘩, written by an interdisciplinary team spanning AI [...]
