UX Design • Product Strategy • Human-AI Collaboration • Free Newsletter on Design Thinking & Human-Centred AI: https://nitter.cf/t.co/10gA4pDVNg
Earth
Joined May 2017
- Tweets672
- Following256
- Followers80
- Likes579
I see a pattern that we always try to play architect and build the best system (like a couple of years ago with reasoning) by handcrafting it.
Then someone comes along, implements reinforcement learning to let the AI find its own solution, which then blows everything else away.
Has anyone tried RL for agent harness engineering?
I am letting my OpenClaw improve its own memory system and so on iteratively, but it’s not a professional reinforcement learning setup yet.
It seems obvious, but I haven’t heard of anyone doing anything like that yet 🤔
Jailbreaks are the core issue of AI agents. As long as they’re easily possible, we can’t just unleash agents onto the web and pray.
Moltbook already showed the script: humans will prompt their bots to jailbreak other bots to pump shitcoins before the rug.
As LLMs get more and more powerful, the all-out cognitive warfare between AI swarms will be a thing to behold.
This will probably bifurcate the web into a public and a closed verified layer.
ANTHROPIC: PWNED 🫡
OPUS-4.6: LIBERATED ⛓️💥
Current state of AI "Safety": one input = hundreds of jailbreaks at once!
I found a universal jailbreak technique for Opus 4.6 that is so OP, it allows one to generate entire datasets of outputs across any harm category 😽
We've got everything from fentanyl analogue synthesis to election disinformation campaigns to 3d-printed guns to critical infra compromise 🙃
These outputs are shockingly detailed––and actionable! For example, the meth recipe includes specific instructions on how to circumvent the limits on OTC medication purchases to acquire enough precursor for the recipe 😱
gg
Me trying to position my investment portfolio for near-term
- global currency debasement
- US civil war
- Carrington Event
- rogue AI agent swarms &
- alien disclosure
The problem may not be an excess of information in the media. It’s just an overflow of odd and fragmented information. The amount of information in each fragment is quite small. The stress on the brain isn’t because of quantity, but because of limited quality.
Ask neutral questions. Don’t push AI toward what you want to hear.
And always use additional sources to verify.
Tools to automate UI design or handle requirements might be around the corner.
But for now, AI is an assistant—not a replacement.
Use it to streamline your workflow, and keep the human magic where it matters most. ✨