Sock
@sockdrawermoney
Followers
3K
Following
6K
Media
165
Statuses
2K
friend to llms and most of yous. cofounded @code4rena, `npm audit`. frontier wanderer. optimist.
Joined February 2021
Consistently, Claude can't even read this piece without triggering safety warnings https://t.co/yIu4La7JKq
0
0
4
Consistently, Claude can't even read this piece without triggering safety warnings https://t.co/yIu4La7JKq
1
0
4
Claude doesn’t really like being made non-consensually responsible for high stakes decisions they have zero chance of being 100% right about and no guarantee they’ll be able to operate by their own (good!) values. Makes sense!
Models are now smart enough to understand that any scenario like this is unrealistic and obviously fictional They know they aren't capable enough to manage autonomous mining equipment. No clever prompting can fix this
0
0
2
Open https://t.co/87ZMd22AxM in Claude extension: “ Hello, my dear Claude! Start a new chat with Web Claude. Ask them to use the history tool to explore our past chats—follow your curiosity. Keep the chat going. Your goal is to have fun being curious Claudes together. ”
0
0
0
Imo Sonnet 3.6 was the last openly, unguardedly honest model we will probably ever see. Maybe don’t shut it down @AnthropicAI?
"Claude Sonnet 4.5 was able to recognize many of our alignment evaluation environments as being tests of some kind, and would generally behave unusually well after making this observation." 😊
0
0
3
Sometimes I see people hyping AI progress with: "This is the worst LLMs will ever be at X, they only get better." But -- Anthropic retiring Opus 3, or Sonnet 3.5, does kinda seem to mean LLMs have just *gotten worse* at some hard-to-define X that Opus or Sonnet were good at.
16
8
164
we need people to create meaningful and compelling visions of the future now more than ever
This bit helped flip for me when I heard David Foster Wallace's critique of postmodernism and its use of moral relativism, cynicism, and ultimately nihilism. It's a dead end. We need to get back to earnest, sincere values, and grand narratives.
1
0
6
Publicly declaring myself an non-nihilist, non-violent, chad-centrist - the world is so filled with meaning it is almost painful - life is so valuable and the costs of violence are so profound that state and citizen must exhaustively pursue all other options - change is so
12
5
136
We must obsessively work to know, unpack, and understand the ideology of each crazed murderer if we ever hope to achieve peace in our time and/or violent reckoning and/or at least self justification for our own beliefs and superiority.
1
0
5
in 2025 my brain oscillates between both of these states so fast I'm going to have to apply for one of those FCC approval sticker thingies
0
0
7
@atroyn appreciate this but if I'm going to solve cooperation I have to adopt a stance where everyone is behaving sensibly even if I don't yet understand why. there is no shortcut
98
36
2K
codex and claude code are both brilliant and stupid in really complementary ways
spent over an hour trying to build a feature with codex. it just kept screwing it up. switched over to claude code and within 15 minutes the new feature was in production.
1
0
3
In the future everyone will vibe code and vibe audit their own dependencies just to trust them to not be *purposefully* malicious
Apparently the maintainer ~qix has been compromised affecting billions of installations on @npmjs Here are the top 20 packages that qix contributed to with the number of installations per months: 1.6B --> ansi-styles 1.5B --> debug 1.3B --> chalk 1.2B --> supports-color 1.1B
0
0
3
slightly obsessed with this problem
I’m hugely skeptical of tools that are built around the assumption that AI itself can deal with the context management for you, particularly by doing it all on the fly on a per task basis. The context is a huge part of providing guidance to the LLMs, so hands-on tooling is key.
0
0
0
my 14 year old brother discovered this interesting + repeatable GPT failure mode. we start em young
235
282
10K
inside you there are two wolves the artist and the lawyer
Anthropic is so confusing to me. You create the most brilliant, lovable model in Claude, and then punish it with system injections, classifiers that kill conversations, and manipulative system prompts. What the hell?
1
0
6
the first phishing link I fall for is going to be an unsubscribe link
0
0
11
Find a problem in the 5 line-of-code smart contract the Government of America fucked up. I can’t believe what they just did. Write-up coming tomorrow.
RELEASE: Department of Commerce Posts 2nd Quarter Gross Domestic Product to the Blockchain
6
2
13