Angehefteter Tweet
Šimon Podhajský
1.2K posts
Šimon Podhajský
@sim_pod
Lapsed neuroscientist turned data everything. Interests outside neuroecon: irrationality, metascience, LLMs (duh)
Prague, Czech Republic Beigetreten Ekim 2009
960 Folgt498 Follower
Šimon Podhajský retweetet
Agent Builders vol. 2 yesterday.
We went deep on agentic evals. How to get started, what tools people actually use, and which methodologies hold up in production.
Thanks Václav Čadek and @sim_pod for coming and @daniel_bukac for presenting how we do it at @duvoai
English
@doodlestein My apologies, I wasn't trying to impugn your reputation! My point (implied as it was) was entirely that naive Clawdbot users that install via the one-liner -- and get used to that delivery mechanism -- will be easy marks for a potential hacker.
Thank you for your work on ACIP!
English
I’m trustworthy. I have a stellar reputation for honesty and I use my real name. My professional reputation is worth a lot to me and would be destroyed if I did anything dishonest and severely injured if I inadvertently did anything negligent.
But I agree that you should be careful about who it is that you take those kinds of scripts from, and always look for those attributes.
Also, you can always have CC review the script first (or use the Claude web app if you’re really paranoid).
English
This actually is a good prompt (though whether it will stand up to targeted prompt injection attacks is an open question), but the irony of having a one-liner that installs it is palpable.
Jeffrey Emanuel@doodlestein
Since I’m seeing so many new people are installing Clawdbot, I highly recommend inoculating it against prompt injection attacks (or at least hardening it a lot to make it much more resistant) with my ACIP project. I even made a one-liner installer script: github.com/Dicklesworthst…
English
@cblatts The Chrome extension is pretty bad/slow for any scraping tasks; for me, the magic comes mostly from terminal applications and parallelizing.
English
All right you boosters—after experimenting on and off for a week with the chrome browser extension and the coworker desktop app on my MacBook, Claude managed to do a bunch of agentic tasks slowly and poorly while taking a GREAT deal of effort. I’m still willing to believe there are use cases for me (outside of statistical code), and I’m going to dive into using the terminal today. Claude probably did not enslave all of you to slavishly promote it on X so any suggestions welcome.
Chris Blattman@cblatts
Approximately 1/3 of my X feed is people gushing about Claude code. I’m already an intensive ChatGPT user so I am open minded. And I will try it. I can't help but wonder: 1. Why do most of these posts sound like they were written by their AI? 2. Is this a viral marketing campaign? 3. Is this just the Twitter algorithm running wild? 4. Why don't I understand from these posts what these people are actually doing with Claude? Why is it all in vague gobbledygook? They talk about tasks in weird jargon, and it's like they're speaking a different language. I really don't understand what 5. Can someone explain in plain English what I would as an academic would do concretely with Claude code? We are already testing it out to clean and analyze basic survey data where it does okay. I'm going to be trying to play around with some new theoretical models, adapting IO models to criminal firms where ChatGPT has been doing ok. Will Claude code do better? Anything else I should be thinking about kind of work they're doing.
English
@EleanorKonik I wrote up my Clobsidian setup but haven't gotten around to polishing it up for Twitter: simon.podhajsky.net/blog/2026/clob… -- looking forward to seeing yours!
English
A couple friends asked me about Claude + Obsidian and they aren't really on X, so I went ahead & typed up some notes about how I'm using LLMs to get things done faster without sacrificing control over communication or research.
It's got links to the resources I found most helpful, some analogies to how the same thought patterns I use in daily life transfer to Claude optimization, and clever (but fun!) tips for getting better with the terminal.
Goes live tomorrow on my website if you're interested in 3,000 words of workflow tips from yours truly.
Bringing back a little of the old Obsidian Roundup magic 💚
English
Šimon Podhajský retweetet
Important point. As organizations grapple with influx of one -shot AI generated content (#6), it is critical to have tools that do not flag the value-generating cases (1-4) as the slop (#6).
To its credit, @pangramlabs does make this distinction, but we have not audited this.
Séb Krier@sebkrier
Imo we should have a 6-point scale for AI involvement in writing: 1. Human-only 2. AI-assisted research/ideation 3. AI-edited human draft 4. AI as co-writer 5. Human-edited AI draft 6. One-shot generated And we should clearly differentiate critiques of substance vs style.
English
Šimon Podhajský retweetet
As a Texas YIMBY, this is what I have the most respect for about the CA YIMBYs.
They are Sisyphus and the housing crisis is the stone. The amount of grit and determination they have is insane. I’m so stoked for them and for California.
Jeremiah Johnson 🌐@JeremiahDJohns
What's impressive about California YIMBYs is that they don't get everything they want every year, but they keep coming back and adding more each session. California still has far too much red tape and NIMBYism but they're whittling away at it year by year, bit by bit.
English
@annierubyjane Čapkova Válka s mloky, ale je to Válka s moly a Chief Moth dostane pořádnou charakterizaci
Čeština
kdyz je zrovna nevrazdim, tak travim premyslenim o molech mnohem vic casu nez bych mela. tohle je uplne namet na horor
Čeština
LLMs amplify midwit trope writing, which is terrible news for midwits like me
(I'm an admitted sucker for "weirdly precise number as storytelling slop" and "metaphorical musing slop", and suspect some of the latter is still good writing when used judiciously)
lyra bubbles@_lyraaaa_
taxonomy of llm slop v1
English
Šimon Podhajský retweetet
Šimon Podhajský retweetet
"I saved a PNG image to a bird" is just an incredible sentence, simultaneously the platonic ideal of YouTube title, and the thing he did.
Db@dbgray
Bird saves and reproduces data: A PNG image of a bird ( photo of a bird -> spectral synthesizer ) was reproduced by an adult Starling bird youtu.be/hCQCP-5g5bo?si…. It seems to have reproduced the sound in conjunction with some additional notes which made it not detectable aside from the recording of the bird in post.
English
Šimon Podhajský retweetet
Outer Wilds (the best video game ever made) is on sale on Steam for $15.
Please play this game. Takes about 20 hours to finish. The less you know going in the better.
English
Šimon Podhajský retweetet
Whenever I see a system prompt that starts with "You are a".
GIF
English
Šimon Podhajský retweetet
An attempt to explain (current) ChatGPT versions.
I still run into many, many people who don't know that:
- o3 is the obvious best thing for important/hard things. It is a reasoning model that is much stronger than 4o and if you are using ChatGPT professionally and not using o3 you're ngmi.
- 4o is different from o4. Yes I know lol. 4o is a good "daily driver" for many easy-medium questions. o4 is only available as mini for now, and is not as good as o3, and I'm not super sure why it's out right now.
Example basic "router" in my own personal use:
- Any simple query (e.g. "what foods are high in fiber"?) => 4o (about ~40% of my use)
- Any hard/important enough query where I am willing to wait a bit (e.g. "help me understand this tax thing...") => o3 (about ~40% of my use)
- I am vibe coding (e.g. "change this code so that...") => 4.1 (about ~10% of my use)
- I want to deeply understand one topic - I want GPT to go off for 10 minutes, look at many, many links and summarize a topic for me. (e.g. "help me understand the rise and fall of Luminar"). => Deep Research (about ~10% of my use). Note that Deep Research is not a model version to be picked from the model picker (!!!), it is a toggle inside the Tools. Under the hood it is based on o3, but I believe is not fully equivalent of just asking o3 the same query, but I am not sure.
All of this is only within the ChatGPT universe of models. In practice my use is more complicated because I like to bounce between all of ChatGPT, Claude, Gemini, Grok and Perplexity depending on the task and out of research interest.
English
English
@tracewoodgrains I've yet to be impressed by anything that's calling itself Straussian, and this is no different
Edginess for its own sake
English
We've made a secure exploratory data analysis workbench at the Prague @e2b hackathon! Loom here: loom.com/share/017f3f4e…
English
Šimon Podhajský retweetet
great news for bayesians. oh is the bayes estimator "biased"?????? ELON WILL NEVER LET YOU TELL ANYONE
Darby Saxbe@darbysaxbe
🚨BREAKING. From a program officer at the National Science Foundation, a list of keywords that can cause a grant to be pulled. I will be sharing screenshots of these keywords along with a decision tree. Please share widely. This is a crisis for academic freedom & science.
English
Šimon Podhajský retweetet
Today, we at OpenAI launched Deep Researcher and I wanted to share a deeply personal story about how amazing this tool is and how it will change the world. Trigger warning, related to cancer....1/9
English
Šimon Podhajský retweetet
Google didn’t get enough credit for everything they provided for free in 2024.
I learned a lot using their free APIs and AI Studio, and I hope they continue this trend in 2025.
AI Studio, available for free to everyone where you can upload an hour-long video for analysis or huge PDFs to get pinpoint answers.
Yet, I hardly hear people talking about how incredible it is.
Logan Kilpatrick@OfficialLoganK
PSA: our experimental Gemini models are free (in Google AI Studio and API), 10 RPM, 4M TPM, 1500 RPD. Enjoy the most powerful models we have to offer (2.0 flash, thinking, 1206, etc), with just 3 clicks on: aistudio.google.com
English