@MichaelWaitze How is this making formal verification practical if we can’t even be sure that the Lean is actually tying everything down to the real-world security guarantees that we care about? https://t.co/N4RXmJstqQ
@Gegam245074 Pure bias example. Expectation got higher and more complex after models became steonger and so prompts got more difficult. If you provide same prompt, new models finish faster with less bugs.
@kaepora I don't understand their reaction. A tool with such reach shouldn't hide behind silence when tough questions are asked. This is an opportunity to provide new basis and clear expectations for the new age that is upon us. Nothing can be hidden anymore behind usual complexity wall.
Tamarin is now alleging that my issues asking for a disclosure policy and for revisiting the use of the term "prover" to describe the tool violated their Code of Conduct.
I welcome anyone showing me where the violation occurred; I can't find it.
https://t.co/Fx70f7ofYE
@Angaisb_ Engineers worked well before AI. Good engineers with open models like GLM 5.3 are very productive. We are not talking about newbies vibe coding an app: such people need proprietary models because they don't understand what they're doing.
@bebuu_ai Not sure about this case but many sysadmins are starting to blame agents for their mistakes. Before AI, many such blunders happened. Now AI is becoming the easiest scapegoat.
@lukOlejnik Still, this will train on conversations flagged which happens quiet often. So, independent researchers should be realistic when using closed models for cybersecurity work.
@TheBlueMatt All LTS distros follow this approach for enterprise offering. People should understand how Linux is used in serious business. Of course, it is Canonical responsibility to ensure the security of their LTS offering like all other distros.
@MarkBaggett@cohere You know that this affects all models, right? So why call out qwen specifically? I don't think you are defending Cohere here...it is the opposite!
@Dinosn Pure fake. Bad harness + 100K tries to make it issue a curl command. Really disappointing from them. They should compete on actual LLM intelligence not play the card of fear mongering
@swapnil_sudhir_@Railway There is no coincidence here. AI capabilities have peaked since September and gpt-6 Astra release and so all dormant high value bugs and "backdoors" are starting to fall. This is good for long term but we'll all suffer in near term.