Eric

AI safety and cybersecurity

[Eric @ewernn]

Opus 4.7 feels like a big jump

[Eric @ewernn]

kinda funny

(I originally posted: March 12, 2026)

[Eric @ewernn]

I imagine AI training as the big blob of compute looking in at all the different personas like this, and being like "oh yeah ok. this is how humans work"

(I originally posted: February 24, 2026)

[Anthropic @AnthropicAI]

AI assistants like Claude can seem shockingly human—expressing joy or distress, and using anthropomorphic language to describe themselves. Why?

In a new post we describe a theory that explains why AIs act like humans: the persona selection model.

https://t.co/Gc3q0Dzq7Z

[Eric @ewernn]

mix and match personas to get funny combos

(I originally posted: March 12, 2026)

[Eric @ewernn]

The sycophancy <--> utility axis
or the RLHF <--> RLVR axis

safety favors the utility side. to satisfy opposing parties without sycophancy/psychosis, party-agnostic utilitarianism+humanism could be the move

(I originally posted: February 16, 2026)

[Eric @ewernn]

eval awareness is a special case of general situational awareness, which is necessary for safety training to generalize OOD

(I originally posted: February 11, 2026)

[Eric @ewernn]

lol that I started cc just to write an `echo` instead of directly in terminal

(I originally posted: February 8, 2026)

[Eric @ewernn]

(I originally posted: February 5, 2026)

[Eric @ewernn]

(I originally posted: February 5, 2026)

[Eric @ewernn]

We're closing in on the new year. A "Claudify Wrapped" on my usage metadata would go hard

(I originally posted: November 13, 2025)

[Eric @ewernn]

onion communities

(I originally posted: January 27, 2026)

[Eric @ewernn]

this is the way

(I originally posted: January 23, 2026)

[Anthropic @AnthropicAI]

If our probe identifies a suspicious query, it sends it to a more powerful “exchange” classifier that sees both sides of a conversation and is better able to recognize attacks.

[Eric @ewernn]

POV: you're a tornado searching for Florida

(I originally posted: November 30, 2025)

[Interesting AF @interesting_aIl]

A small demonstration of how hard it is to find planet Earth in the middle of space

[Eric @ewernn]

We're closing in on the new year. A "Claudify Wrapped" on my usage metadata would go hard

(I originally posted: November 13, 2025)

[Eric @ewernn]

my Claude custom instructions:

be concise and tell me the harsh truth, but don't be sycophantic (tell me if you are uncertain).

(I originally posted: October 10, 2025)

[Eric @ewernn]

I be asking Claude to list decisions, and then rate each one out of 10. Works well.

(I originally posted: May 5, 2025)

[Eric @ewernn]

BE CURIOUS

Choose to be curious.

My grandma always said “you can choose to be happy or you can choose to be sad. I choose to be happy”

my interpretation is “you can choose to be curious or you can choose to be scared. I choose to be curious”

Fear is the mind-killer

apr23 '25

[Eric @ewernn]

Trends are pertinent. They have momentum. Trends are more important than individual events. Ppl are increasingly thinking in rates of change and trends, and this is most important than ever as acceleration increases.

(I originally posted: April 23, 2025)

[Eric @ewernn]

Obstruction

Obfuscation

(I originally posted as 2 posts: April 21, 2025)

[Eric @ewernn]

(I originally posted: April 7, 2025)

[Eric @ewernn]

ebb and flow

(I originally posted: March 30, 2025)