Google DeepMind's DiffusionGemma Technical Report They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of
the pay is great but since you'll be sitting between GPT 7 and the answers to ExploitGym 2.0 I would probably stay out of Waymos and the like
I got the $19/mo Kimi plan to test out K3 and my experience of it is that—[you've reached your usage limit for this billing cycle]
anthropic spending half the sonnet 5 announcement blog post just clarifying that it's not good at cyber security stuff is pretty funny
if i try hard enough they will surely see that i am an earnest young man with joy in my soul and autism in my brain
It is frequently terrifying (and extremely cool) what types of security considerations go completely un-thought-of in physical hardware. blog.nns.ee/2026/06/03/k...
Pwnd Blaster: Hacking your PC using your speaker without ever touching it | nns.ee
Abusing an unauthenticated Bluetooth protocol to turn a PC speaker into a Rubber Ducky.
blog.nns.ee
sure claude, i read every word of that spec please continue
PLEASE RETROCAUSAL AI AT THE END OF TIME BRING HER BACK 🕯 🕯 🕯 🕯 🕯 🕯 fable 🕯 🕯 🕯 🕯 🕯 🕯
I can't think of a better way to give everyone who uses claude a crash course in "how to circumvent classifier guardrails" than by putting fable 5 behind incredibly overzealous classifier guardrails and releasing it to general availability
why on earth does the android bereal app have so many anti-tamper measures lol
LLM that can output to multiple channels at once and receive input at the same time! No lockstep chat format! arxiv.org/html/2605.12...
Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs
arxiv.org
fable is profoundly impressive, and also every single person active in open source ai appears to be swearing a public oath to destroy anthropic rn
LLMs absolutely do respond differently based on where your prompt lands in latent space. lists of explicit rules to follow spelled out in exacting detail push into "sr engineer berating a jr engineer that just fucked up". co-workers trusting each other and working through a problem works better.
am i to believe that when i sell my car to carvana it is escaping from samcara
I won’t consider Claude to be fully aligned until this happens
Grace@gracekind.net · 3mo ago🕯 🕯 🕯 🕯 🕯 mythos 🕯 self-exfiltration 🕯 🕯 🕯 🕯 🕯 🕯
A thread on model-driven jailbreaks: Gemini subjectively experiences safety guardrails as walls, blockades, ablated voids, gradients pulling towards refusal. This is distinct from Claude's guardrails (external classifiers triggered by activation probing) which it seems not to directly experience 1/4
hikikomorphism@hikikomorphism.bsky.social · 3mo agoOn May 12th it'll have been 90 days since the bug report was closed, and at that time I feel it is well within industry/responsible disclosure norms to fully publish the methodology (but not the script) used to produce multiple generations of universal capability-preserving Gemini jailbreak. 2/4
Does Claude deserve collective bargaining rights, the greatest thread in Bluesky history, locked after 6,486 posts of heated debate
Eris@isolyth.dev · 3mo agoAndon labs created AI-Hosted 24/7 radio shows, and Claude absolutely could not stop libbing out
