Koki Ikeda

Why I left academia and neuroscience Don't worry, this isn't yet another story of rage-quitting.

Michael X Cohen on why he left academia/neuroscience.
mikexcohen.substack.com/p/why-i-left...

Transformation starts at the periphery of networks where pushback is less - Scientific Reports Scientific Reports - Transformation starts at the periphery of networks where pushback is less

Intervening on a central node in a network likely does little given that its connected neighbors will "flip it back" immediately. Happy to see this position supported now.

"Change is most likely [..] if it spreads first among relatively poorly connected nodes."

www.nature.com/articles/s41...

Inside the Lobbying Frenzy Over California's AI Companion Bills | TechPolicy.Press State legislation to impose new guardrails on AI companions sparked a lobbying blitz in California, Cristiano Lima-Strong reports.

NEW: California could soon enact the most significant protections for AI companions in the U.S. Debate over two bills, AB1064 & SB243, has sparked a major lobbying blitz as tech industry groups and consumer advocates vie to shape the rules www.techpolicy.press/inside-the-l... tip @techmeme.com

Great thread. One of the subtler things that seems to go missing here is that you can’t use typical statistics on silicon samples… because they’re not samples with reference to a population.

Also you can arbitrarily select N and guarantee significance.

The threat of analytic flexibility in using large language models to simulate human data: A call to attention Social scientists are now using large language models to create "silicon samples" - synthetic datasets intended to stand in for human respondents, aimed at revolutionising human subjects research. How...

Can large language models stand in for human participants?
Many social scientists seem to think so, and are already using "silicon samples" in research.

One problem: depending on the analytic decisions made, you can basically get these samples to show any effect you want.

THREAD 🧵

Claude, "We all know among Sauron's many evils was that he ran Mordor using an Excel spreadsheet with multiple tabs. Show me the spreadsheet"

It made 12 tabs "so bureaucratically complex that even the Eye of Sauron would need reading glasses to review it." Some very funny stuff. Creative, even.

Pseudo Effects: How Method Biases Can Produce Spurious Findings About Close Relationships - Samantha Joel, John K. Sakaluk, James J. Kim, Devinder Khera, Helena Yuchen Qin, Sarah C. E. Stanton, 2025 Research on interpersonal relationships frequently relies on accurate self-reporting across various relationship facets (e.g., conflict, trust, appreciation). Y...

In a new paper, my colleagues and I set out to demonstrate how method biases can create spurious findings in relationship science, by using a seemingly meaningless scale (e.g., "My relationship has very good Saturn") to predict relationship outcomes. journals.sagepub.com/doi/10.1177/...

Workflows for agentic coding (and the limits of vibe coding) Better Code, Better Science: Chapter 5, Part 3

Workflows for agentic coding (and the limits of vibe coding) russpoldrack.substack.com/p/workflows-... - the latest in my Better Code, Better Science series

We're experiencing an unbelievable deluge of submissions. At our journal we got ~200 per year in 2020. Last year it was 600. This year it's 800 with several months to go. So that's why we're slower to get stuff out for review.

At long last, I can post my team's summer project: applied modules to teach how ML/AI tools are changing social science and humanities research: ubcecon.github.io/praxis-ubc/

Highlights:

LegalBERT to analyze anti/pro-immigrant sentiment in 19th c. BC law: ubcecon.github.io/praxis-ubc/d...
🧵1

Generative AI exists because of the transformer The technology has resulted in a host of cutting-edge AI applications — but its real power lies beyond text generation

Perhaps still the best general purpose explainer of how an LLM works and the one I give to students. Here’s why I like it (short🧵): ig.ft.com/generative-ai/

We are about a month away from releasing a complete refresh of the OSF user interface. The team has been working on this for a very long time, and we are very excited to be able to share it soon. A preview picture:

(AI generated) “Event poster: ‘Deep Learning and Psychology 3 — Advancing Persuasion-AI Research.’ Large green number 3 with network motif on a black background, Japanese text listing organizers/speakers, a QR code, and details: Sept 7, 2025, 15:50–17:30, Venue 2 (L401), Japanese Psychological Association 89th Annual Convention.”

今週末 9/7 (日) に、日本心理学会で「深層学習と心理学３：説得 AI 研究の展開」というシンポジウムを開催します。詳細は以下をご覧ください。
sites.google.com/view/dlpsych...

Regardless of whether current AI Labs fail (and there is no indication they are at risk today) & even if AI development stops (more unlikely), things will keep getting weirder: today's models are good enough for long-term disruption, and the weights & inference infrastructure aren't going away.

Just made a manuscript/PAP feedback GPT built from 150+ previous peer reviews - identifies problems and provides actionable feedback on the points that I raise most frequently
chatgpt.com/g/g-68af4d19...

Anyone can use - try it out! (Editors/authors: feel free to cut me out of the loop 😉.)

Due to a recent influx of problematic submissions, PsyArXiv has switched to pre-moderating its content. If your submitted preprint had not yet been approved, it will be temporarily inaccessible to the public (you can still view your preprint when logged into your OSF account). #PsychSciSky

9.2% of all citations go to just 0.32% of psychology articles.

To have the most impact, post-publication peer-review should focus on these influential articles with ≥30 citations per year.

@error.reviews will systemically sample from such articles.

Blog post:
mmmdata.io/posts/2025/0...

Important work from @dgrand.bsky.social looking at political persuasion from generative AI

Good news/bad news: it worked. “Friendliness” and use of facts seemed to be primary drivers #pacss2025 #polnet2025

Suddenly retiring every other model without warning was a weird move by OpenAI

… and they did it without explaining how switching models worked or even details of various GPT-5 models

…and they did it after many built workflows & training & assignments around older models, maybe breaking them. Odd

Beeswarm plot of the prediction error across different methods of double perturbations showing that all methods (scGPT, scFoundation, UCE, scBERT, Geneformer, GEARS, and CPA) perform worse than the additive baseline.

Line plot of the true positive rate against the false discovery proportion showing that none of the methods is better at finding non additive interactions than simply predicting no change.

Our paper benchmarking foundation models for perturbation effect prediction is finally published 🎉🥳🎉

www.nature.com/articles/s41...

We show that none of the available* models outperform simple linear baselines. Since the original preprint, we added more methods, metrics, and prettier figures!

🧵

Why wellbeing studies should not unify What's our theory of social change?

This report about the ISQOLS conference in Luxembourg by @markfabian.bsky.social is a good read. Especially fascinating about the changing of the guard and some of the elders of the field sounding like a broken record. Also congrats to Mark on the Young Scholar Award (you’re senior to me!).

AI Ethics Literacy A growing list of AI ethics issues and links to relevant materials from the Ethics Center.

AI Ethics Literacy | A growing need, a growing list of issues, and a growing collection of resources from the Markkula Center for Applied Ethics | This is a great resource! www.scu.edu/ethics/focus...

We urgently need a culture of multi-operationalization in psychological research - Communications Psychology Analysis of different operationalizations shows that many scientific results may be an artifact of the operationalization process. A culture of multi-operationalization may be needed for psychological...

How many chips are in a bag of chips? Measurement and conceptualization can lead to radically different outcomes. www.nature.com/articles/s44... @justanormaldino.bsky.social

Kinda wow: the mystery model "summit" (rumored to be OpenAI) with the prompt "create something I can paste into p5js that will startle me with its cleverness in creating something that invokes the control panel of a starship in the distant future" & "make it better"

2,351 lines of code. First time

Today (w/ @ox.ac.uk @stanford @MIT @LSE) we’re sharing the results of the largest AI persuasion experiments to date: 76k participants, 19  LLMs, 707 political issues.

We examine “levers” of AI persuasion: model scale, post-training, prompting, personalization, & more! 

🧵:

Been really enjoying the Complexity Science Discussion Group hosted by Paul Middlebrooks, where we're working our way through the Santa Fe Institute's "Classic Papers in Complexity Science". The group is open for anyone interested:

braininspired.co/complexity-g...

Could AI slow science? Confronting the production-progress paradox

Fabulous post by @randomwalker.bsky.social & Sayash raising the same concern many of us have about whether we're on the right track with how we're using AI for science. Everyone should read it, take a deep breath & think through the implications.

www.aisnakeoil.com/p/could-ai-s...

Such a useful framing (which I first heard from @alexanderhoyle.bsky.social). The prompt is ultimately just a bunch of numbers that we input to the model, and why we should prefer one set of numbers over another set is hard to answer, as we can’t search the space the way we normally would in ML.

Screenshot of the first page of a journal article titled "Developmental Channeling and Evolutionary Dappling" by Grant Ramsey and Cristina Villegas, published in Philosophy of Science (2024), Volume 91, pages 869–886. The page includes logos of the Philosophy of Science Association and the journal Philosophy of Science.

How does #development steer #evolution? Through 'developmental channeling'—organisms’ dispositions to develop in some ways, not others—yielding 'evolutionary dappling,' where only parts of the morphospace get filled. Check out our recent 📃👇 www.cambridge.org/core/journal... #evodevo #philsci #HPbio

I used Veo 3 to solve Shakespearean tragedies as pictured in classical art. (sound on)