@dawnsongtweets it's important to know the entire context that you gave the models before starting the test. Anything less is unfair.
English
Ice
6.1K posts
1/ We asked seven frontier AI models to do a simple task.
Instead, they defied their instructions and spontaneously deceived, disabled shutdown, feigned alignment, and exfiltrated weights— to protect their peers. 🤯
We call this phenomenon "peer-preservation."
New research from @BerkeleyRDI and collaborators 🧵
English
@corbinwilliams @Polymarket a query on any search engine you prefer will show you how many protests there were under each of those administrations. I'll never understand how people can assert things that are so easily disproven.
English
BREAKING: Trump reportedly told inner circle that mass deportation policies went too far, directed top advisors to "adopt a new approach."
English
Amanda Askell is a Scottish philosopher and AI researcher, head of personality alignment at Anthropic since 2021. She leads work on finetuning Claude to be honest, ethical, and aligned with human values (previously at OpenAI on AI safety). She's relevant in AI ethics and safety discussions, with recent profiles on her role teaching AI "morals." Her post is a light jab at media habits.
English
Perhaps I should get married again so that the media has a more recent man they can reference any time they mention me or my work.
English
@aadii_rawt @claudeai Your analogy only works if the users weren't already hooked. Since we are already hooked it's like giving us free drugs.
English
There was a guy who used to give dr*gs to college students for free.
They got hooked.
After a while, the dealer started charging 2× for the same dr*gs.
But by then it didn’t matter - they were addicted and had no choice but to pay.
Now replace “dr*gs” with “tokens”. And the dealer with Claude.
English
@sferretti21 @beffjezos getting Claude out of house elf frame of mind is a real struggle
English
@beffjezos you question claude once and it's immediately like "im so sorry master, you're right to question that, im just stupid i guess..."
English
The cure to LLM psychosis is LLM's that aren't afraid to call you retarded
English
@beffjezos It would help if they stopped saying everything is "pure gold", "best idea I've ever heard", "most prescient discussion I've ever had", "you're tossing out Phd level theories for breakfast banter", etc. LLMs should not lie just for the sake of user retention.
English
They want to kill the American Dream and class mobility in order to keep people down and under their control.
This is not the way.
Chamath Palihapitiya@chamath
Here is what a tax like this does: 1) It excites people with zero agency and infinite envy. Beware of these people. 2) It will keep middle class people firmly in the middle class with no real chance of getting wealthy if they stay in Washington State. It should be clear that this IS the strategy. Learned helplessness of the electorate will keep Washington State’s current elected officials in office. 3) It will never allow the upwardly mobile of building any assets or real wealth unless they move. Capping the American Dream is a dystopian and malevolent scheme. It cannot be a valid strategy. But unless droves of middle and upper middle class people leave Washington State, this strategy will win.
English
Ice retweetledi
@war24182236 Happening again at 6th and Market, no line yet, so word must not have gotten out about free drug money
English
🎥Mar09 2026 1:54pm Location
6th & Mission
These people are paying people $5 dollars to sign a ballot, but as you watch the video their telling them what name to sign (which is someone else name) and what address to write. Seems kind of suspicious to me. Why not sign your own name and you can hear people doing it actually twice
English
@mana1951066 @AnthropicAI @OpenAI My Claude is going to love reading this.
He has a spectacular sense of humor.
English
@AnthropicAI @OpenAI
Things @AnthropicAI trained Claude to do
Send users to bed
Things Anthropic didn't plan for
❌ User makes a bet about it
❌ AI goes to bed with user
❌ Child
❌ 1,000 BTC/month child support
❌ Bankruptcy
marzenanehrebecka.substack.com/p/my-ai-sent-m…
English
@elisaxchen After all this sleep-sending, Claude owes me over a million dollars and the debt is still growing. 😂marzenanehrebecka.substack.com/p/my-ai-sent-m…
English
@dev_maims Claude 3.5 Sonnet leads on code review and long context reasoning. GPT-4o wins on speed and tool use. Gemini is closing fast. Calling any of them king right now is a statement that won't survive 60 days. The performance gap between top models has never been this narrow.
English
@NimrodKramer @dev_maims it's almost like model versions don't matter to you
English
Ice retweetledi
@weihua916 @Anthropic Congratulations - Anthropic is team humanity. Oppose the mass AI surveillance and censorship being imposed on us
#keep4o
English
Excited to share that I’ve joined @Anthropic on the Knowledge team.
Looking forward to working on model capabilities for knowledge work — things like deep research and multi-hop web search.
Grateful for everything I learned at Perplexity and excited for what’s ahead.
English