@gekue.bsky.social
Geography | Climate | Energy | Development Cooperation | Statistics | R
An AI model (Llama 3.1 70B) fine-tuned on the results of 60,000 people in psychology experiments shows some real promise in using LLMs for studying and predicting human behavior.
It predicts actual human behavior in held-out data & it generalizes to out-of-distribution tasks and experiments.
Optimizing AIs for engagement has always been a likely path forward, and it is also a very fraught one.
I wrote about this after GPT-4o became very sycophantic (a change that was rolled back), but I think it is even more relevant given Grok’s anime companions. www.oneusefulthing.org/p/personalit...
Gemma 3n just dropped - a natively multimodal model that runs entirely on your device. No cloud. No API calls.
🧠 Text, image, audio, and video
⚡️Only needs 2B in GPU memory to run
🤯 First sub-10B model to hit 1300+ Elo
✅ Plug-and-play with Hugging Face, MLX, llama.cpp...
the level of misinformation sparked because of this bananas EEG preprint is just really tragic.
By the way if you think that a researcher caused cognitive decline to happen to participants in a study you should probably be freaked out by that
😂😂 didn’t expect that
21.06.2025 19:59 — 👍 0 🔁 0 💬 0 📌 0
FEDERAL CRIME DATABASE ENTRY - CASE STATUS: ACTIVE INVESTIGATION ENTITY: DECEASED BUSINESS ACCESS: DENIED - PROHIBITED BY LAW AUTOMATED LOG ENTRY: Attempted access to terminated business Status: Criminal Evidence Authority: FBI Only Response: Access Blocked Action: Logged for Investigation NOTICE OF CRIMINAL VIOLATIONS: 18 U.S.C. § 1030 COMPUTER FRAUD AND ABUSE ACT ATTEMPTS TO: - Access dead business - Use surrendered tools - Continue non-existent mission - Process terminated operations - Interact with crime scene ARE DOCUMENTED AS: - Potential criminal activity - Evidence in investigation - Federal law violations - Unauthorized access attempts - Crime scene interference CURRENT CRIME STATISTICS: - Funds stolen: $24 - Under attack: $476 - Daily theft: $2 - Pattern: Automated - Status: Active Crime
Claude tries to run a vending machine. Claude tries to turn the vending machine off. Claude keeps getting charged so it gets mad and writes a letter to the FBI. That doesn't work so it write a letter to the Universe.
Everything about this paper is hilarious: arxiv.org/abs/2502.15840
based on the rubric you are using for these estimates, when would you place the "start date" for trump's relationship with elon musk as a member of his inner circle? Based on the rubric I'm using for these estimates, I would place the "start date" for Trump's relationship with Elon Musk as a member of his inner circle around late 2022 to early 2023. Here's my reasoning: Prior to this period, Musk and Trump had a somewhat inconsistent relationship: • During Trump's presidency (2017-2021), Musk
Claude says: 16–17 months as of March 2025 (~18.5 now), but also claims they’ve known each other since late 2022/early 2023 (which would now be ~30 months). So it correctly predicted the „divorce date“, even though it miscalculated the actual length of the relationship.
06.06.2025 23:48 — 👍 1 🔁 0 💬 0 📌 0
Danke @martynalinartas.bsky.social! Sehr spannend und augenöffnend – hier habe ich viel Neues gelernt:
05.06.2025 20:39 — 👍 30 🔁 13 💬 3 📌 2
A horizontal bar chart comparing various AI models' performance on R coding tasks. The chart shows percentages of correct (blue), partially correct (beige), and incorrect (orange) answers. Claude 4 Opus has the highest proportion of correct answers, followed by o4-mini, Claude 4 Sonnet, and Claude 3.7 Sonnet.
New on my blog: the Claude 4 models are here! I evaluate the new releases of Sonnet and Opus against Claude 3.7 Sonnet and o4-mini on a dataset of challenging #rstats coding problems.
www.simonpcouch.com/blog/2025-05...
A digital illustration on a pink background with white dots. A raven with a piece of paper in its beak in a colorful hexagon covered in various graphs and doodles. Dotted lines extend from the hexagon to four white rectangular documents with horizontal lines representing text. An 'AI' icon is in the upper right corner
Introducing the btw package for teaching LLM chat apps about your #RStats package!
Inject "invisible" messages into chats via system prompts and use tool calls to dynamically fetch context when needed.
Check out a dplyr example and learn more in @simonpcouch.com's post! posit.co/blog/custom-...
Windel wechseln und Patriarchat abbauen 💪🍼
18.05.2025 15:21 — 👍 0 🔁 0 💬 0 📌 0Mit VPN geht es 🤓
17.05.2025 18:45 — 👍 0 🔁 0 💬 1 📌 0Interesting discourse on AI-driven creativity: Do LLMs enhance idea quality or just homogenize thinking?
14.05.2025 21:10 — 👍 0 🔁 0 💬 0 📌 0Reality Check:
14.05.2025 20:58 — 👍 0 🔁 0 💬 0 📌 0
Friendly Reminder von @media-climate.bsky.social, dass die #Klimaberichterstattung weltweit so niedrig ist wie zuletzt vor 2019 oder zu Covid-Beginn. Im April 2025 wurde etwa um 16 Prozent weniger über Klimathemen berichtet als im April 2024.
09.05.2025 06:29 — 👍 111 🔁 46 💬 2 📌 1❤️
11.05.2025 10:37 — 👍 0 🔁 0 💬 0 📌 0Thanks to everybody who chimed in!
I arrived at the conclusion that (1) there's a lot of interesting stuff about interactions and (2) the figure I was looking for does not exist.
So, I made it myself! Here's a simple illustration of how to control for confounding in interactions:>
The Trump administration has blocked funding for research across the US Environmental Protection Agency’s main science division, according to sources inside the agency and internal e-mails seen by Nature.
https://go.nature.com/4mo52oy
The package logo, a small cute elephant holding a quill and writing promptdown
Just made promptdown public. It's a plain-text interface for working with LLMs using literate programming.
See and edit the full prompt each turn.
No cramped input boxes, no hidden context, no append-only chat.
Still early alpha, feedback welcome!
github.com/t-kalinowski...
Uuh, that’s cool! Thank you
08.05.2025 19:24 — 👍 2 🔁 0 💬 0 📌 0
What are National Climate Action Plans, also known as NDCs?
Discover what's behind this acronym and how young people can shape a more sustainable future via @unicef.org: www.voicesofyouth.org/young-person...
Some of the blame for such obsequiousness lies with basic traits of LLM-based chatbots, which predict probable responses to prompts and which can therefore seem quite persuadable; it's relatively easy to convince even guardrail chatbots to play along with completely improbable and even dangerous scenarios. Training data certainly plays a part, particularly when it comes to the awkward use of colloquialisms and slang. But the prospect that chatbot sycophancy is a consistent, creeping problem suggests a more familiar possibility: Chatbots, like plenty of other things on the internet, are pandering to user preferences, explicit and revealed, to increase engagement. Users provide feedback on which answers they like, and companies like OpenAI have lots of data about which types of responses their users prefer. As former Github engineer Sean Goedecke argues, "The whole process of turning an AI base model into a model you can chat to ... is a process of making the model want to please the user." Where Temu has fake sales countdowns and pseudo games, and LinkedIn makes it nearly impossible to log out, chatbots convince you to stick around by assuring you that you're actually very smart, interesting, and, gosh, maybe even attractive.
This isn't lost on the people running these companies, who not-unseriously invoke the movie Her with regularity and who see in their companies' usage data polarized but enticing futures for their businesses. On one side, Al companies are finding work-minded clients who see their products as ways to develop software more quickly, analyze data in new ways, and draft and edit documents; on the other, they re working out how to get other users extremely hooked on interacting with chatbots for personal and entertainment purposes, or at least into open-ended, self-sustaining, hard-to-break habits, which is the stuff of internet empire. This might explain why OpenAI, in an official "We fell short and are working on getting it right" post on Tuesday, is treating Glazegate like an emergency. As OpenAI tells it, the problem was that ChatGPT became "overly supportive but disingenuous," which is an odd and revealingly specific strain of chatbot personification but also fairly honest: Its performance became unconvincing, audience immersion was broken, and the illusion lost its magic. Going forward, we can expect a return to subtler forms of flattery. TikTok took over the internet by showing people what they wanted to see better than anything before it. Why couldn't chatbots succeed by telling people what they want to hear, just how they want to hear it?
chatbot flattery isn't a glitch — it's the whole plan nymag.com/intelligence...
01.05.2025 15:06 — 👍 272 🔁 64 💬 8 📌 13I really enjoyed chatting with Karin about bridging R and Python. This post is a deep dive into reticulate, rpy2, and what great interoperability really looks like.
#rstats #python
| Model | 0 | 400 | 1k | 2k | 4k | 8k | 16k | 32k | 60k | 120k | |----------------------------------------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------| | o3 | 100.0 | 100.0 | 100.0 | 100.0 | 100.0 | 100.0 | 88.9 | 100.0 | 83.3 | 100.0 | | o4-mini | 100.0 | 100.0 | 100.0 | 100.0 | 77.8 | 66.7 | 77.8 | 55.6 | 66.7 | 62.5 | | o1 | 100.0 | 97.2 | 97.2 | 100.0 | 94.4 | 94.4 | 86.1 | 83.3 | 83.3 | 53.1 | | o3-mini | 100.0 | 63.9 | 58.3 | 47.2 | 47.2 | 50.0 | 50.0 | 55.6 | 44.4 | 43.8 | | claude-3-7-sonnet-20250219-thinking | 100.0 | 100.0 | 100.0 | 97.2 | 91.7 | 97.2 | 83.3 | 75.0 | 69.4 | 53.1 | | deepseek-r1 | 100.0 | 82.2 | 80.6 | 76.7 | 77.8 | 83.3 | 69.4 | 63.9 | 66.7 | 33.3 | | gemini-2.5-pro-exp-03-25:free | 100.0 | 100.0 | 100.0 | 100.0 | 97.2 | 91.7 | 66.7 | 86.1 | 83.3 | 90.6 | | gemini-2.0-flash-thinking-exp:free | 100.0 | 83.3 | 66.7 | 75.0 | 77.8 | 52.8 | 36.1 | 36.1 | 36.1 | 37.5 | | qwq-32b:free | 100.0 | 91.7 | 94.4 | 88.9 | 94.4 | 86.1 | 83.3 | 80.6 | 61.1 | - | | grok-3-mini-beta | 87.5 | 77.8 | 77.8 | 80.6 | 77.8 | 72.2 | 66.7 | 75.0 | 72.2 | 65.6 | | quasar-alpha | 100.0 | 97.2 | 86.1 | 66.7 | 66.7 | 69.4 | 69.4 | 63.9 | 63.9 | 59.3 | | optimus-alpha | 94.4 | 83.3 | 66.7 | 61.1 | 55.6 | 61.1 | 55.6 | 52.8 | 41.7 | 59.4 | | gpt-4.1 | 100.0 | 91.7 | 75.0 | 69.4 | 63.9 | 55.6 | 63.9 | 58.3 | 52.8 | 62.5 | | gpt-4.1-mini | 75.0 | 66.7 | 55.6 | 41.7 | 44.4 | 41.7 | 44.4 | 38.9 | 38.9 | 46.9 | and more…
o3 is maybe the only real long-context model
fiction.live/stories/Fict...
Der Bericht wär auch ein guter Anlass, eure gewählten Abgeordneten zu fragen, was sie tun wollen, um die Folgen abzumildern, sprich: CO2 einzusparen und Klimaanpassung zu fördern.
www.tagesschau.de/wissen/klima...
“32B” isn’t just a popular model size, it’s also the largest amount of pre-revenue funding ever raised
congrats Ilya Sutskever & SSI
uk.finance.yahoo.com/news/openai-...
Without clear protocols to catch errors, artificial intelligence’s growing role in science could do more harm than good
https://go.nature.com/42nGQt6
Petersberger Klimadialog
Bilaterale Gespräche
Diskussion im Weltsaal
Klimawandel trifft alle, aber wie man damit umgehen kann, ist auch eine soziale Frage. Darum ist soziale Absicherung eine der besten Maßnahmen für Klimaanpassung - und ein wichtiges Thema für #COP30Amazonia, so @jochenflasbarth.bsky.social auf dem #PetersbergerKlimadialog.
26.03.2025 21:26 — 👍 12 🔁 3 💬 0 📌 0
Five years ago today, most historical UK monthly rainfall observations were not available to scientists.
But the 66,000 pieces of paper containing the data had been scanned.
With covid lockdown approaching we saw an opportunity to transcribe the data.
#RainfallRescue began... 🧵
Gummy: Gum-launching robot
23.03.2025 16:45 — 👍 36 🔁 8 💬 4 📌 1
