Vasilios Mavroudis
@mavroudisv
Principal Research Scientist @turinginst. Autonomous Cyberdefence 🤖 and AI cyber risks ⚡
'water is transparent only within a very narrow band of the electromagnetic spectrum, so living organisms evolved sensitivity to that band, and that's what we now call "visible light". ' (found via HN)
My friends, I want to organise Secure AI Club in London -- gig for people interested in (practical!) AI Security. Not just academic toy setups, but actually making systems reliable. Trying to gauge interest, please sign up here: forms.gle/zSUMh6ykthQwtt…
WOW! DeepMind *just* dropped Magenta Real-time - Apache 2.0 licensed 🔥 > 800M params transformer, trained on ~190K hours of instrumental stock music > adapts MusicLM for real-time generation via 2s audio chunks (conditioned on prior 10s context) > 48 KHz Stereo > MusicCoCa: New…
As a scientist I love elegant solutions but Agentic AI security will be achieved by several complementary interventions rather than a silver bullet. I unpack this argument in my commentary: techmonitor.ai/technology/cyb…
Man, I love legacy tech. So many cool ideas that made it to products but never saw widespread use.
In 2001, Sony Japan released an adapter allowing the PlayStation to connect to mobile phones for network services.
"Krapivin was not held back by the conventional wisdom for the simple reason that he was unaware of it" They achieved a constant average query time regardless of the hash table's fullness (for non-greedy hash tables): arxiv.org/abs/2501.02305 quantamagazine.org/undergraduate-…
I realised I spend increasingly more time in websites with the "open" prefix: *.openai.com, openrouter.net @OpenRouterAI, the dreaded openreview.net, but openwrt.org will always have a special place in our hearts.
1.5 yrs ago, we set out to answer a seemingly simple question: what are we *actually* getting out of RL in fine-tuning? I'm thrilled to share a pearl we found on the deepest dive of my PhD: the value of RL in RLHF seems to come from *generation-verification gaps*. Get ready to🤿!
Configuring chatGPT to be upfront has 100% upgraded my experience.
Tired of ChatGPT's excessive politeness? Worry not; here's a built-in "jailbreak". You'll be surprised how direct it gets.
My name seems to be an "adversarial example" for google search. I have nothing to do with insurance underwriting but maybe I should consider changing careers.
Apparently gyms operate on the basis that not all customers will use their membership. Which means my subscription is actually subsided by all these people who sign up impulsively. Thank you all ♥️
Quasar Alpha and Optimus Alpha Optimus alpha were OpenAI as speculated.
Stealth model reveal 👀 Many of you have been excited to know who made Quasar Alpha & Optimus Alpha, which both topped the charts during testing. Today we can announce that they were early test versions of...
Today is the start of a new era of natively multimodal AI innovation. Today, we’re introducing the first Llama 4 models: Llama 4 Scout and Llama 4 Maverick — our most advanced models yet and the best in their class for multimodality. Llama 4 Scout • 17B-active-parameter model…