Want to know what all this microsoldering was for and which silicon bugs we had to exploit? Come to our (@stacksmashing, @cryptotx and me) wallet.fail talk at #35c3. fahrplan.events.ccc.de/congress/2018/…
Dmitry Nedospasov
492 posts
Dmitry Nedospasov
@nedos
🚀 Building next-gen AI systems
💎 Auditing Web3's heavy hitters @keylabsio, ex- @walletfail
🎓 PhD in Security Training @advsecio
Streaming online
Joined November 2007
- Today @StackSmashing and I are announcing our most thorough course on glitching ever. We call it “Applied Glitching of Real-World Hardware Targets”. The course has been entirely designed to be taught live with many live demos against actual targets. advancedsecurity.training/training/live-…
- Found myself opening Photoshop just to remove a background from an image. In fact I think statistically background removal is what I use Photoshop for the most. But guess what? There's an open source AI for that (~17 models). Adding to @comput3ai soon:
- A bunch of conferences are moving to an online format. We wrote down our thoughts about how online trainings should be and what you should watch out for. advancedsecurity.training/posts/online-t…
- Updated my list of recommended lab equipment to include the Rigol MSO5074. My current recommendation for an entry to medium level oscilloscope. The refresh is a bit slow at times but everything else is awesome. toothless.co/equipment
- We just improved our TRELLIS Gradio app and docker container. It now boots in about 40 seconds instead of 2 minutes on our GPU instances.
- Happy to see people trying the coding models on @comput3ai. People don't realize how far open source models have come. Open source models get updated as well. The gap to frontier is definitely shrinking. We built some coding agents internally that were already pretty impressive.Gm Degens!! Vibe coding on @comput3ai could not be easier! Stake $COM and run a node Thanks @nedos, this is a eazy low IQ 10X
- Got claude code running on one of @comput3ai's B200's today. The B200s are running a Kimi K2 Q8 with 128K (the max) of context. With 4 "threads".
- Just updated our FPGA and Glitching classes. Get them both for the price of just one training. ast.fyi/bundles
00:00 - Long time coming! We‘ll be offering not only my four day class, but classes people have been asking us for forever. I’ll be teaching an intro to FPGAs class using Lattice FPGAs and the open source tool chain in addition to the other excellent classes we have.For a while we have been offering professional security trainings around the world. Today we are excited to announce our online live trainings!
- We're gonna play around with hosting our own Kimi K2 instances this Thursday on @comput3ai. Noon Eastern. Mark your calendars.
- Replying to @NFTland @untitledxyz_ and 6 othersBoom! Here's where we got those models up and running (and how). We'll be giving away 10 free GPU hours to those who join today, so make sure you join! 👉youtu.be/jrYe2DXozqQ
- People always ask how big a model really is and what it takes to host it. Forget marketing slides—here’s the napkin math you can use in the field. At FP8, one parameter equals one byte. A 1 B model needs 1 GB of vRAM. That’s it. The GPUs that used to push pixels are now just
