Sakana AI just beat GPT-5.5-Cyber and Mythos Preview on real-world security benchmarks.
> 86.9 on CyberGym, 72.1 on CTI-REALM — top of both leaderboards
> It's an orchestration model, not a monolith — Fugu routes and coordinates the work
A small lab out-scoring frontier cyber models from OpenAI and Anthropic.
@abhishekcode42 Maybe the parents of the 70 odd kids who have committed suicide over the last four years in Kota should take parenting lessons from you.
Moronic to look at two horrible exams and come to the conclusion “You know what - I think the JEE helps 15 year olds with character building”
Aarav had spent two years chasing the IIT dream, buried in books and mock tests. The night before JEE Advanced, he was anxious. His grandfather sat beside him and said, “Remember the mango seed you planted? It didn’t grow overnight, but with time, it became a tree. Success is like that, it comes, not always through one exam, but through steady growth.” JEE Advanced is tomorrow. Yes, it matters, but it is just one chapter, not the whole book. Whether or not you get into an IIT, you’ll still find your way. So take a deep breath, pop a chill pill, and give the exam with a calm mind. You’ve got this.
I condemn the social media trolling directed at family of the Foreign Secretary @VikramMisri
It’s unacceptable to target our professional diplomats and civil servants — those who work dedicatedly to serve the nation.
Join us after the winter break for an in-person-only talk with Global Leadership Fellow Atanu Chakraborty. Register for "The Lion Uncaged: The Story of India’s Institutional and Economic Rebirth" on Wednesday, Jan. 8 at 5 p.m.: https://t.co/jBulDVdX0Q
@GPSIndiaCenter
We have extended the Call for Participation (CFP) and Call for Villages (CFV) deadline to Tuesday, January 7!
Call for Participation: https://t.co/4le12oHytO
Call for Villages: https://t.co/pFAYb9gXCE
#bsidessf2025#bsidessf#infosec#cybersecurity
@dakshgup Good job. You’re attempt at some Friday night engagement farming on Greptile’s behalf worked! You sound like a smart guy. No way you believe in the BS you just put out there.
I love twitter bc where else can I, an person with two graduate degrees from Harvard in both infectious disease epidemiology and biostatistics, be assured I am wrong about infectious disease epi and statistics by a software engineer who has “read multiple FDA package inserts”
@notsky I was his mentor at Databricks and yes he was absolutely awesome. Not in the least bit surprised at his trajectory. And finding his LI floating about on X is so random lol.
I want to believe that these popular large agent frameworks provide value above and beyond what simpler agents and LLM interactions alone can. But I can't help but see a similar pattern to the explosion of Javascript web frameworks in the 2000/10's. Layers upon layers of abstraction with seemingly no purpose other than to claim support for features like "question decomposition" which is ultimately implemented by passing a string through 5 layers of handlers and proxy classes. I'm not sure how often I need to "delegate to a sub agent", certainly not often enough to justify this level of abstraction.
In my experience the most useful LLM agents tend to be very simple and include a good REPL to interact with it. The one I use the most is a couple hundred lines of C++ that was mostly generated by GPT-4. It won't provide the automation necessary for scaling capabilities such as vulnerability discovery. But it is a prerequisite for working with these models in a way that discovers these capabilities in the first place.