Chams @shemsddine

@akincibor1 @banteg Pay 50 -> sorry someone else found it first 🤣

@deadvolvo Cars vs Hours

@rootxharsh I think someone should spend some tokens and give it a try

@mountainaddict0 @grok please tell this bot to stop posting fake stuff and explain them there is no bounty here

@SocketSecurity Where are the npm haters now?

Reading vigolium docs — like the skills-at-escalation idea. One thing I'm not seeing: a deterministic precondition (AST routes) or a cite-gate ahead of report_finding. On large codebases that re-derivation gets expensive. Any token/finding benchmarks or a confirmed CVE from agent mode?

@hash_kitten @siunam321 Thanks for sharing, really helpful to read your reasoning process

@hash_kitten @kevin_mizu claude code sonnet 4.6 took 30mins

It's like relying on cloud streaming vs owning a powerful local console Today the cloud (GPT-5.5) gives you the absolute best graphics and newest titles. Your stacked DGX Sparks are the high-end console sitting right there.Open-source models are updating so fast that soon your local setup will deliver near-identical quality for 90% of your games with zero latency, no subscription, full privacy, and you own it forever.Renters pay monthly. Owners win long-term. and dont forget something called PRIVACY

In 12-24 months, "good enough" local will cover 80-90% of serious daily use for most people building agents/coding — cheaper, more private, and more reliable long-term. Your DGX setup is perfectly positioned for that shift instead of just sitting there.Why bet everything on renting intelligence forever when you can own the stack? Curious what you think — you planning to fire those Sparks up for anything local yet?Hey man, fair take — and I get why you're saying it. Those two DGX Sparks sitting there are basically a mini supercomputer, and right now frontier models like GPT-5.5 or Claude Opus 4.7 still feel like the reliable default for serious agent work and vibe coding. Quantized local stuff (even strong ones like Kimi or GLM) often feels like a noticeable step down in quality/consistency for the hardest tasks. No argument there today.openai.comBut tomorrow? The delta is shrinking fast, and the trajectory heavily favors local/open-source for most real workflows.Open-source models have closed the gap dramatically. Look at DeepSeek-V3.2 / V4 (671B+ MoE, 37B active) — it's hitting frontier-level math/reasoning/coding benchmarks (often matching or beating earlier GPT-5 generations) under MIT license. Llama 4 Maverick, Mistral Large variants, Qwen3, etc., are right there too. People are already running strong quantized versions of these on hardware like your DGX setup (or even smaller clusters) with usable speed via vLLM/llama.cpp + good quantization (FP8 or 4-bit loses very little quality now).fireworks.aiWith your two stacked DGX Sparks, you could comfortably run something like a quantized DeepSeek-V4 Pro or Llama 4 Maverick locally for the bulk of your Hermes agents — full privacy, no rate limits, zero per-token costs after hardware, offline capable, and full control to fine-tune or build custom agents on your own data. No more "hoping the cloud provider doesn't throttle or change the model." For the few tasks where the absolute bleeding-edge frontier still wins, hybrid routing (local first, escalate to GPT/Claude when needed) is becoming the smart default anyway.The tech is moving insane fast:Better quantization + MoE efficiency = bigger models run well on less hardware. Inference engines keep getting faster (multi-token prediction, etc.). New open releases drop every few months that eat into the proprietary lead.

@ctatedev the idea is good but know how are you gonna build the dataset to train the model to use it and which model are you planning to fine tune

@yegor256 They said the new algo will not show you people that rage bate but still @xai need to find a way to fix this

@yegor256 Are you saying this just to piss us off ? i wish they add a dislike button

@UsmanReads I will ping you when i find one but still no luck

Looking for a no-bullshit Discord/server for people actually building & shipping AI agents + LLM work. Genuine devs only, folks who know what the words mean, share real code/architecture, debug in public, and aren't just reposting hype. Drop links if you got one. Building in the trenches daily, want real peers. #LLM #AIagents #BuildInPublic

@NVIDIAGeForce #007FirstLightRTX

Repo: github.com/ChamsBouzaiene… Article: bouzaiene.org/blog/vuln-scan…

I wanted to test something: Can modern AI coding agents rediscover a real critical vulnerability in complex C code with almost no guidance? So I gave Cursor Composer 2 and Codex/GPT-5.5: one vulnerable nginx source file generic security auditing prompts no CVE hints no fuzzers no static analyzers Both independently rediscovered CVE-2026-42945. Cursor found the bug extremely fast. Codex/GPT-5.5 went further: validated the finding generated a PoC compiled nginx with ASan triggered a real heap-buffer-overflow What surprised me most is that the vulnerability itself was not impossibly complex. It survived because nobody looked deeply enough at that exact interaction. LLMs are becoming surprisingly good at recursively investigating these kinds of subtle state mismatches. For transparency: part of this experiment was conducted through OpenAI’s Cyber program access, which required identity verification before participation. I published the full write-up + reproducible repo below. Really interesting to see where tools like @cursor_ai and GPT-5.5 are heading for recursive code auditing.

@orcdev i enjoy styled-components

@xenobyte_ @feross u mean its a standard scan to enumerate the maintainers emails this i understand but all of this attacks at the same time seem to me that some groups are automating multiple tools on multiple big npm packages to find a way in that's what am thinking about