Dimitri

InfoSec professionnal

[Dimitri @popshell]

@samali459 @ow Yep, all using pseudonyms and anonymous posting.​‌

[Dimitri @popshell]

@ow Risks are real: multi-jurisdiction ID theft, multiple EU/Five Eyes citizenships = heightened scrutiny in "certain countries" + make you a prime target for "certain people" seeking those "specific identities", especially knowing you work for a major fintech company.

[Dimitri @popshell]

@ow Showing off all your passports on social media feels good until you realize you're handing out prime identity-fraud intel, advertising an empty house, and broadcasting your children's citizenship status to the world without their consent.

[Dimitri @popshell]

@fs0c131y Why not RTX 50590? Use case?

[Dimitri @popshell]

This aged very well #Qwen 3.8 27b

[Dimitri @popshell]

@gregpr07 What's your "Alt-Tab" app on MacOS?

[Dimitri @popshell]

Qwen 3.8-27B doesn't have a lot of world knowledge.

It's small and very cheap to run.

But super post-trained for agentic tasks and tool use.

So it can reconstruct knowledge through lots of thinking as the need arises.

[Dimitri @popshell]

Qwen 3.8-27B is agentic/thinking maxed, so better to go with RTX 5090 over DGX Spark. Dense models hate low memory bandwidth.
RTX 5090: 1,792 GB/s
DGX Spark: 273 GB/s
32 GB of VRAM is enough and mem. bandwidth is the real bottleneck.

[Dimitri @popshell]

@fs0c131y To be honest I would go with RTX 5090 because GPU memory mem. bandwith is 1,792 GB/s and Spark is 273 GB/s. Yes only 32GB VRAM on RTX but huge difference in tok/sec.

[Dimitri @popshell]

@fs0c131y Why two? Cluster?

[Santiago @svpino]

First, they stole our data.
Then, they sold it back to us.
Now, they watermark it.
Soon, they claim they own it all.

[Dimitri @popshell]

Pentester crisis over time:
2003 Metasploit make pentesters obsolete.
2006 Cloud eliminate infrastructure security jobs
2010 Automated scanners kill manual testing.
2012 Bug bounty platforms race everyone's rates to zero.
2026 AI will replace everyone.
They're still here.

[Dimitri @popshell]

Qwen 3.8 27b shows a new rule: train a small model long enough, and it learns to think. If LLMs were just a "database" of facts, cutting the size 100x would cut intelligence by 100x. Actually only the facts go away. The thinking stays.

[Dimitri @popshell]

Knowledge moved from data into weights, and now from weights into tools. The weights kept the only part that mattered. #Qwen 3.8-27b

[Dimitri @popshell]

If open weights large language models were poisoned or backdoored, closed AI labs would make sure everyone knew.

[Dimitri @popshell]

In 2026, API stands for Agent Programming Interface.

[Naval @naval]

Don’t send me the report, just send me the prompt.

[Dimitri @popshell]

The endgame for AI frontier labs: turning cybersecurity workflows into metered LLM inference.

[Dimitri @popshell]

It is unacceptable for a tool to sacrifice even an iota of clarity, coherence, meaning, or quality in order to embed hidden clues about its provenance. #watermark #claude

[Dimitri @popshell]

Rumor says someone dropped $50K on tokens at the latest @defcon CTF. Rumors tend to exaggerate. Maybe it was only $48K