LiveOverflow 🔴

Where to find me: 🔴 Hacking Videos: youtube.com/LiveOverflow 📜 Articles: liveoverflow.com 💻 Start Hacking? liveoverflow.com/start-hacking 🥰 Support: liveoverflow.com/support 📹 2nd Channel: youtube.com/LiveUnderflow 🤳 Instagram: instagram.com/liveoverflow

@At0mXploit Honestly I am jealous of the new generation. I wish I had ChatGPT when I started :D you can learn so much more in shorter time

Sometimes looking at @LiveOverflow old videos showing his ideas and research while doing ctf challs makes me jealous that i wasn't in compsci BC (before chatgpt).

An important thing to observe (and this applies to all model providers everywhere, and has a long-running basis in international law): Foreigners abroad can be mass surveilled. And everybody is a foreigner abroad almost everywhere.

We're launching a $100,000 fund to save CTFs.

I have said this before, but as we are moving into a world where surveilling every spoken word becomes feasible: A good principle would be "no surveillance without representation, no use of force against me without representation". We are far away from that.

@__Necros @colunaprestess @ProDeveloperFL Why don’t you call it obsessive when somebody constantly talks about a programming language or a vulnerability? I just spend a lot of time with AI, so I’m gonna talk about it… Also AI courses coming to @hextreeio 🙃

@colunaprestess @LiveOverflow @ProDeveloperFL obsessing over it and becoming over-reliant on it completely turns your brain into an omelette. Furthermore, AI is easy, there's no reason to talk this much about it. You just tell it what to do and it does it. We don't need 10000000 articles about how to speak to a bot

I wonder if the same developers who used to say "I'm a web developer, I don't need to know assembly", are now saying "I'm a web developer with AI, I don't need to know html/js/..." 🤔

@__Necros @ProDeveloperFL My previous tweet was just devils advocate. I will always strive to understand the layers below

@LiveOverflow @ProDeveloperFL brother i used to love you as a kid. What happened... Who bought you... How did you fall so hard man

@halvarflake @robertgraham In my home village, any new building is required to get solar

@robertgraham You'll also run AC during times when solar energy production is firing. I mean, if they're worried about CO2 emissions from AC, simply make it a requirement that the AC is solar-powered.

Germans heating their homes in winter puts more carbon into the air than air conditioning in the summer. Germans can become more environmentally friendly by replacing the coal/oil/gas furnaces they use now with heat pumps, and using them in the summer as air conditioners. Also, none of this matters as much as replacing nuclear power plants with brown coal power plants. Germans are weird.

@ProDeveloperFL Who needs JS and HTML to build a website?

@LiveOverflow Who needs assembly to build a website....

@lux_sp4rk Didn’t say you need it. But knowing it gives you a lot of context and better understanding what’s happening

@LiveOverflow never in a million years have you ever needed to write assembly scrub dont tell me you actually did that shit recently

@7h3h4ckv157 Are they consistent? Or do they change their mind and say you should know programming languages to use AI more effectively.

@LiveOverflow Okay. Your point? :)

@sadc0der what you mean? I see many people say you don't need to know programming anymore because AI just codes it for you. Like in the past, I still think there is a huge benefit of knowing the layers below. So even if you are coding with AI, knowing programming still helps.

@LiveOverflow This is a good analogy, but it lacks realism

Was wondering why Fable was suddenly disabled in Cursor. If the EU doesn’t get their asses together we gonna have big problems in the future…

I cancelled my $10/mo Calendly subscription and vibe coded my own with Fable for $12,000

@brutecat @0xLupin Obviously pays for itself :P I am still very curious about the token cost, because your workflow seems very elaborate with lots of agents running. So just 1k per month sounds very little to me - considering what I can burn with Cursor in a week 😅

@LiveOverflow @0xLupin i hadn’t actually calculated it tbh but it’s all subsidized compute from numerous accounts, in total maybe $1k/month or so, just 1 bug pays it off anyways

Hacking Google with A.I. for $500,000 brutecat.com/r/hacking-goog…

@stelin41 Ohhh, “attention sink” is what I had in mind. Thank you!

@LiveOverflow There are some adjacent concepts. The attention sink token which is sometimes used as a "summary" token. Local/sparse attention which only uses semantics of a few adjacent tokens. SBERT is trained to output a token encoding a "summary" used for search.

Anybody into model interpretability research. Do we know if certain tokens get “abused” in the layers as a “summary” token for the context before it? I don’t know how to describe it properly 😅 but when you have multiple rounds of sliding attention, summaries of meaning might be carried from the front to the end? But this kinda “muddies?” what the vector should mean? So I was wondering if a special “snapshot” token could be used that is somehow trained that (sliding) attention uses it to “summarize” prompt before. Maybe if it is ignored during training for next token prediction, and neural network or attention can more “freely” use it. Or am I basically describing the idea of linear attention?

@halvarflake Makes sense to me if a different (smaller) model is used for the guardrail checks. Maybe the large model does not have the guardrails in its parameters at all. Thus distilling from it just gives you an approximation of that model, including its non guardrail performance.

Uh. Distilling from a model with guardrails somehow equips the trained models to learn things the teacher refuses to teach?

Do we know the attention layers used by the large proprietary models? It’s great that we have huge contexts. But if there is eg. sliding attention of a certain limit, it would be good to know. Then you can plan prompt layout or define tool output token limits around it.