Flo 🔶
@faz.ms
Interested in #Infosec and #ProductiveDisagreement | #StayAtHomeDad, worked in #Biotech, co-built tiny companies in renewables and structural engineering sectors, ex #HumanRights observer | likes #ElixirLang, #Boardgames |🔸 #10PercentPledge
My Claude deep research prompts keep failing after around 2h with just the message "Something went wrong". Does anybody else have this problem?
November 12, 2025 at 3:19 PM
My Claude deep research prompts keep failing after around 2h with just the message "Something went wrong". Does anybody else have this problem?
Tech debt is like real debt - the correct amount to have is almost never zero.
Vibe coding is the microfinance of tech debt: quick, accessible, optimal in small doses, dangerous when misunderstood
Vibe coding is the microfinance of tech debt: quick, accessible, optimal in small doses, dangerous when misunderstood
most engineers probably don’t understand the extent to which “tech debt” works as an abstraction
there’s absolutely “good tech debt”, and it’s not always obvious which you’re dealing with
if you’re green fielding an app, you should expect to launch with some amount of tech debt
there’s absolutely “good tech debt”, and it’s not always obvious which you’re dealing with
if you’re green fielding an app, you should expect to launch with some amount of tech debt
November 9, 2025 at 8:05 PM
Tech debt is like real debt - the correct amount to have is almost never zero.
Vibe coding is the microfinance of tech debt: quick, accessible, optimal in small doses, dangerous when misunderstood
Vibe coding is the microfinance of tech debt: quick, accessible, optimal in small doses, dangerous when misunderstood
Vibe coded a Bayesian evidence update calculator as Claude artifact, really fun with the artifacts-builder skill.
claude.ai/public/artif...
claude.ai/public/artif...
Bayesian Calculator - Calculate Posterior Probabilities Online
Free Bayesian probability calculator to compute posterior probabilities using Bayes' theorem. Enter prior probability, likelihood, and evidence to get instant results.
claude.ai
November 5, 2025 at 8:23 PM
Vibe coded a Bayesian evidence update calculator as Claude artifact, really fun with the artifacts-builder skill.
claude.ai/public/artif...
claude.ai/public/artif...
„They can’t all be impostors“-syndrome
Type of imposter syndrome that’s not “everyone else is more competent than me,” but rather “I want to believe everyone else is more competent than me because otherwise that would be really scary”
November 3, 2025 at 9:25 PM
„They can’t all be impostors“-syndrome
Once MCP had reached market saturation and every company had connected their Slack and Google workspace to Claude, Anthropic upgraded their internal Clio tool to Claude‘s Capitalist Coordination Protocol, finally giving the world an actual working Cybersyn.
November 3, 2025 at 9:22 PM
Once MCP had reached market saturation and every company had connected their Slack and Google workspace to Claude, Anthropic upgraded their internal Clio tool to Claude‘s Capitalist Coordination Protocol, finally giving the world an actual working Cybersyn.
Really useful framing when explaining prompt injection to devs:
Want to build a low(er) risk agent? Choose two:
A) process untrustworthy inputs
B) access to sensitive systems or private data
C) change state or communicate externally
Want to build a low(er) risk agent? Choose two:
A) process untrustworthy inputs
B) access to sensitive systems or private data
C) change state or communicate externally
I wrote up some notes on two new papers on prompt injection: Agents Rule of Two (from Meta AI) and The Attacker Moves Second (from Anthropic + OpenAI = DeepMind + others) simonwillison.net/2025/Nov/2/n...
New prompt injection papers: Agents Rule of Two and The Attacker Moves Second
Two interesting new papers regarding LLM security and prompt injection came to my attention this weekend. Agents Rule of Two: A Practical Approach to AI Agent Security The first is …
simonwillison.net
November 3, 2025 at 8:10 PM
Really useful framing when explaining prompt injection to devs:
Want to build a low(er) risk agent? Choose two:
A) process untrustworthy inputs
B) access to sensitive systems or private data
C) change state or communicate externally
Want to build a low(er) risk agent? Choose two:
A) process untrustworthy inputs
B) access to sensitive systems or private data
C) change state or communicate externally
Reposted by Flo 🔶
The "solo developer builds complex distributed systems" era is here
October 19, 2025 at 8:47 PM
The "solo developer builds complex distributed systems" era is here
This is very useful! This test should be on every model card.
Spoiler the best format is apparently something called Markdown-KV which is kinda like YAML
www.improvingagents.com/blog/best-in...
www.improvingagents.com/blog/best-in...
Which Table Format Do LLMs Understand Best? (Results for 11 Formats)
We benchmarked 11 data formats, including markdown tables, JSON, CSV, YAML, XML, and more, to see which LLMs understand best. We found surprising accuracy and token cost trade-offs.
www.improvingagents.com
October 6, 2025 at 6:18 PM
This is very useful! This test should be on every model card.
@antiqe.bsky.social another Swapcard UX annoyance I have:
When selecting a session from the agenda I’d like to have an option to also block that time slot for 1on1s.
And conversely when I schedule a 1on1 I‘d like to see which sessions I’m missing. Currently I always need a second tab or device.
When selecting a session from the agenda I’d like to have an option to also block that time slot for 1on1s.
And conversely when I schedule a 1on1 I‘d like to see which sessions I’m missing. Currently I always need a second tab or device.
October 2, 2025 at 2:06 PM
@antiqe.bsky.social another Swapcard UX annoyance I have:
When selecting a session from the agenda I’d like to have an option to also block that time slot for 1on1s.
And conversely when I schedule a 1on1 I‘d like to see which sessions I’m missing. Currently I always need a second tab or device.
When selecting a session from the agenda I’d like to have an option to also block that time slot for 1on1s.
And conversely when I schedule a 1on1 I‘d like to see which sessions I’m missing. Currently I always need a second tab or device.
★ Insight ────────────────────────
We are due for a flood of this meme template by claude code 2.0 users.
─────────────────────────────
We are due for a flood of this meme template by claude code 2.0 users.
─────────────────────────────
September 29, 2025 at 8:52 PM
★ Insight ────────────────────────
We are due for a flood of this meme template by claude code 2.0 users.
─────────────────────────────
We are due for a flood of this meme template by claude code 2.0 users.
─────────────────────────────
I want a /defrag command in claude code that shows me little boxes moving around as the context gets optimized
Memory Editing API
It automatically garbage collects parts of the context that haven’t been used in a while. (Pic looks like compact, blog says remove, idk)
yooooo! this is what i want from a mechinterp-obsessed lab. Nice!
www.anthropic.com/news/context...
It automatically garbage collects parts of the context that haven’t been used in a while. (Pic looks like compact, blog says remove, idk)
yooooo! this is what i want from a mechinterp-obsessed lab. Nice!
www.anthropic.com/news/context...
September 29, 2025 at 7:41 PM
I want a /defrag command in claude code that shows me little boxes moving around as the context gets optimized
There should be a I’m ready button like in online games, where the music intensifies when you’ve clicked it and then once everyone has clicked there’s a big countdown and everyone gets dropped in at once.
zoom should add an option called "join when >50% of the attendees are there"
September 26, 2025 at 9:44 PM
There should be a I’m ready button like in online games, where the music intensifies when you’ve clicked it and then once everyone has clicked there’s a big countdown and everyone gets dropped in at once.
Reposted by Flo 🔶
this really ought to be our future, not an alternate reality. other paths look very bad to me. bsky.app/profile/elix...
Alternate Reality where AI X-Risk is a left coded anti capitalist talking point.
September 26, 2025 at 5:15 AM
this really ought to be our future, not an alternate reality. other paths look very bad to me. bsky.app/profile/elix...
I just want someone to build the Motorola Flip Out with current hardware
en.wikipedia.org/wiki/Motorol...
en.wikipedia.org/wiki/Motorol...
September 24, 2025 at 7:25 AM
I just want someone to build the Motorola Flip Out with current hardware
en.wikipedia.org/wiki/Motorol...
en.wikipedia.org/wiki/Motorol...
Reposted by Flo 🔶
What??! You mean we need to find a way to build reliable systems off of unreliable signals, because elimination of unreliability is intractable?
If only there were an entire field of study dedicated to this!
If only there were an entire field of study dedicated to this!
Worse news: there are uncountably many undecidable problems. Computers have been a scam this whole time
"In a landmark study, OpenAI researchers reveal that large language models will always produce plausible but false outputs, even with perfect data, due to fundamental statistical and computational limits."
www.computerworld.com/article/4059...
www.computerworld.com/article/4059...
September 21, 2025 at 8:49 PM
What??! You mean we need to find a way to build reliable systems off of unreliable signals, because elimination of unreliability is intractable?
If only there were an entire field of study dedicated to this!
If only there were an entire field of study dedicated to this!
Reposted by Flo 🔶
Can somebody please build a bot that checks shared articles that cite one paper against the abstract and conclusion of the paper to see if they say the opposite? I bet this would even work with gemini-2.5-flash
September 21, 2025 at 9:53 PM
Can somebody please build a bot that checks shared articles that cite one paper against the abstract and conclusion of the paper to see if they say the opposite? I bet this would even work with gemini-2.5-flash
Steelmanning an argument only makes sense if you are actually willing to consider believing in it if it’s true.
This is something that obviously no one ever does in political debate
This is something that obviously no one ever does in political debate
i mean. in practice, yes. but it's a handy shorthand for "finding the best argument for something" provided one does not consider "finding the best argument for something" to be an inherently good and valid thing to do
steelmanning EVERYTHING is just participation trophies for ideas
steelmanning EVERYTHING is just participation trophies for ideas
"steelman" culture is just dishonesty. there is no one who holds the reasonable views which media liberals propose conservatives believe in. they are wholly a rhetorical construct of the center-left. a figment we can shadowbox to avoid confronting what conservatives actually believe.
September 17, 2025 at 7:45 PM
Steelmanning an argument only makes sense if you are actually willing to consider believing in it if it’s true.
This is something that obviously no one ever does in political debate
This is something that obviously no one ever does in political debate
Since reading Jostein Gaarder I’m all in on simulacra welfare – writing fanfic happy endings all day
What if you trained an AI model to be just a guy?
> New blog post: "Simulacra Welfare: Meet Clark"
gracekind.net/blog/simulac...
> New blog post: "Simulacra Welfare: Meet Clark"
gracekind.net/blog/simulac...
Simulacra Welfare: Meet Clark • Grace Kind
AI welfare has been a hot topic recently.
gracekind.net
September 15, 2025 at 7:08 PM
Since reading Jostein Gaarder I’m all in on simulacra welfare – writing fanfic happy endings all day
Here’s my theory on why we don’t see a flood of new solo dev apps (yet): with LLMs it’s even easier to fall into feature creep instead of putting a tiny MVP in front of actual users, so there’ll a lot more software that nobody wants.
September 11, 2025 at 5:48 PM
Here’s my theory on why we don’t see a flood of new solo dev apps (yet): with LLMs it’s even easier to fall into feature creep instead of putting a tiny MVP in front of actual users, so there’ll a lot more software that nobody wants.
Reposted by Flo 🔶
vacuum abhors nature right back
In my professional capacity, I cannot emphasize enough how much space wants to kill you.
September 7, 2025 at 6:21 PM
vacuum abhors nature right back
Reposted by Flo 🔶
Hallo #Münster!
Am 14.9. finden die Kommunalwahlen statt.
Basierend auf einem super Artikel von rums.ms habe ich den RUMS-O-MAT entwickelt:
Bewertet die zusammengefassten Aussagen der Parteien - ohne vorab zu wissen, von welcher Partei sie jeweils stammt!
Lege direkt los:
rumsomat.kultpower.de
Am 14.9. finden die Kommunalwahlen statt.
Basierend auf einem super Artikel von rums.ms habe ich den RUMS-O-MAT entwickelt:
Bewertet die zusammengefassten Aussagen der Parteien - ohne vorab zu wissen, von welcher Partei sie jeweils stammt!
Lege direkt los:
rumsomat.kultpower.de
Startseite | RUMS
Sie kennen uns schon und möchten ein kostenpflichtiges Abo abschließen?
rums.ms
August 28, 2025 at 8:30 PM
Hallo #Münster!
Am 14.9. finden die Kommunalwahlen statt.
Basierend auf einem super Artikel von rums.ms habe ich den RUMS-O-MAT entwickelt:
Bewertet die zusammengefassten Aussagen der Parteien - ohne vorab zu wissen, von welcher Partei sie jeweils stammt!
Lege direkt los:
rumsomat.kultpower.de
Am 14.9. finden die Kommunalwahlen statt.
Basierend auf einem super Artikel von rums.ms habe ich den RUMS-O-MAT entwickelt:
Bewertet die zusammengefassten Aussagen der Parteien - ohne vorab zu wissen, von welcher Partei sie jeweils stammt!
Lege direkt los:
rumsomat.kultpower.de
Always fun and unsettling at the same time seeing LLMs come to terms with the fact that we live in an out-of-distribution world compared to their training data.
August 28, 2025 at 10:47 PM
Always fun and unsettling at the same time seeing LLMs come to terms with the fact that we live in an out-of-distribution world compared to their training data.
In the next months a lot of people will install Linux for their relatives who have never used it before. If that’s you, I recommend #VanilaOS: immutability on a Debian base keeps surprises to a minimum and will likely keep your relative-tech-support calls to a minimum.
August 24, 2025 at 10:37 PM
In the next months a lot of people will install Linux for their relatives who have never used it before. If that’s you, I recommend #VanilaOS: immutability on a Debian base keeps surprises to a minimum and will likely keep your relative-tech-support calls to a minimum.
Reposted by Flo 🔶
Generally I think learning the lesson “don’t ask questions you don’t want to hear the answer to” is a principal cause of metas dysfunction.
August 19, 2025 at 6:38 PM
Generally I think learning the lesson “don’t ask questions you don’t want to hear the answer to” is a principal cause of metas dysfunction.