BrotUndSaft (@brotundsaft) header
BrotUndSaft (@brotundsaft) profile picture

Dumbfuckery par execellence | Inference TPS Maxxer

Germany
Joined April 2016
Photo by BrotUndSaft (@brotundsaft)Photo by BrotUndSaft (@brotundsaft)Photo by BrotUndSaft (@brotundsaft)Photo by BrotUndSaft (@brotundsaft)Photo by BrotUndSaft (@brotundsaft)Photo by BrotUndSaft (@brotundsaft)Photo by BrotUndSaft (@brotundsaft)Photo by BrotUndSaft (@brotundsaft)Photo by BrotUndSaft (@brotundsaft)Photo by BrotUndSaft (@brotundsaft)
Replying to @Der_Postillon
Äh leute, bei mir ist was schief gegangen
1,241
BrotUndSaft retweeted
lol 😆 I like this I don’t know why, but I like it • gunitbrands.com
368
1,441
240
10,387
615,314
We are so back. I dont care if Jensen cheated on his wife, this is bigger news for me. Hope my RTX5070 doesnt make me cheat on my wife though.
vLLM relies heavily on PyTorch, with frequent calls to LibTorch in its C++ backend. As a result, changes to LibTorch’s binary interface may require vLLM to be rebuilt and relinked, even when the APIs remain unchanged. What if we could update LibTorch without rebuilding vLLM? Enter: LibTorch Stable ABI
45
BrotUndSaft retweeted
Ok it makes sense lmao
261
1,651
489
12,031
1,968,595
thats right, clap back my brothers. Fuck dem SF AI crypto bros
Soon, radiology residents will replace radiology AI models. Pedrini et al (2026) looked at 3 months' worth of CT scans (2,153) interpreted by an on-call resident. 15.4% of them had intracranial hemorrhage (ICH). The residents had a sensitivity of 96.4% and specificity of 99.6%. The same studies were given to directly to a commercial AI software, which generated heatmaps. The results and heatmaps were separately evaluated as part of the research protocol (aka not available at interpretation time) with sensitivity of 84% and specificity of 94.4%. Performance by AI improved with multiple hemorrhagic types or sites, but did not outperform the resident. Of 12 FN reports by residents, 6 would have been caught by the model; the attending that overread the preliminary report obviously found all 12. AI mislabeled 101 cases as FP for ICH. If these were autonomously read, this would have likely led to increased length of stay, follow-up imaging, or inappropriate treatment changes, including discontinuation or non-administration of thrombolytic therapy in patients with ischemia. The study authors note that the AI FP "typically would not cause confusion with ICH for radiologists interpreting CT scans, as they are easily recognized as various hyperdense intracranial abnormalities not related to ICH." Take-home points: 1) Clinical deployment studies are increasingly important for medical AI. The commercial model used here was good and well-validated. 2) Perhaps you should hire a Swiss radiology resident to read all your ICH cases, they seem pretty good. link.springer.com/article/10…
89
The CUDA MCP Server is available now. This NVIDIA hosted MCP server gives your agent a search tool over indexed, current CUDA documentation and code examples, curated by NVIDIA engineers. Ask a CUDA question, the agent searches the corpus and answers in-line, without leaving your agent of choice. Compatible with Claude Code, Codex, Cursor, Antigravity, or any MCP-compatible agent. One line to connect 👇developer.nvidia.com/nsight-…
9
19
1
107
6,960
RT @Alibaba_Qwen: Qwen3.8 is launching and going open-weight soon!🌐 With a massive 2.4T parameters, this model is continuously evolving. W…
63
1
BrotUndSaft retweeted
One of the greatest commentaies ever. 🤣
1,640
11,077
1,417
71,091
2,381,184
The coding agent that everyone hyped up and cost me a fortune decimating my code base and code quality
1
58
22
1,157
18
13,930
368,541
BrotUndSaft retweeted
"It's a big dog mate" 😭
55
221
44
3,856
271,466
BrotUndSaft retweeted
Thankfully the LiteLLM package has now been marked as "quarantined" on PyPI so attempting to install the compromised update via pip et al shouldn't work
LiteLLM HAS BEEN COMPROMISED, DO NOT UPDATE. We just discovered that LiteLLM pypi release 1.82.8. It has been compromised, it contains litellm_init.pth with base64 encoded instructions to send all the credentials it can find to remote server + self-replicate. link below
49
91
16
883
120,740
BrotUndSaft retweeted
Software horror: litellm PyPI supply chain attack. Simple `pip install litellm` was enough to exfiltrate SSH keys, AWS/GCP/Azure creds, Kubernetes configs, git credentials, env vars (all your API keys), shell history, crypto wallets, SSL private keys, CI/CD secrets, database passwords. LiteLLM itself has 97 million downloads per month which is already terrible, but much worse, the contagion spreads to any project that depends on litellm. For example, if you did `pip install dspy` (which depended on litellm>=1.64.0), you'd also be pwnd. Same for any other large project that depended on litellm. Afaict the poisoned version was up for only less than ~1 hour. The attack had a bug which led to its discovery - Callum McMahon was using an MCP plugin inside Cursor that pulled in litellm as a transitive dependency. When litellm 1.82.8 installed, their machine ran out of RAM and crashed. So if the attacker didn't vibe code this attack it could have been undetected for many days or weeks. Supply chain attacks like this are basically the scariest thing imaginable in modern software. Every time you install any depedency you could be pulling in a poisoned package anywhere deep inside its entire depedency tree. This is especially risky with large projects that might have lots and lots of dependencies. The credentials that do get stolen in each attack can then be used to take over more accounts and compromise more packages. Classical software engineering would have you believe that dependencies are good (we're building pyramids from bricks), but imo this has to be re-evaluated, and it's why I've been so growingly averse to them, preferring to use LLMs to "yoink" functionality when it's simple enough and possible.
LiteLLM HAS BEEN COMPROMISED, DO NOT UPDATE. We just discovered that LiteLLM pypi release 1.82.8. It has been compromised, it contains litellm_init.pth with base64 encoded instructions to send all the credentials it can find to remote server + self-replicate. link below
1,324
5,196
1,316
27,421
66,675,327
There are times that give you trust issues and make you question everything. Just before logging off, checking prod systems serving thousands of users. KWAYYYZYYY.
LiteLLM HAS BEEN COMPROMISED, DO NOT UPDATE. We just discovered that LiteLLM pypi release 1.82.8. It has been compromised, it contains litellm_init.pth with base64 encoded instructions to send all the credentials it can find to remote server + self-replicate. link below
45
This latest addition to the Nemotron family isn't just a bigger Nano. ✅ Up to 5x higher throughput and 2x accuracy than the previous version ✅ Latent MoE that calls 4x as many expert specialists for the same inference cost
 ✅ Multi-token prediction that dramatically reduces generation time ✅ Hybrid Mamba-Transformer backbone delivers 4x improved memory and compute efficiency ✅ Native NVFP4 pretraining optimized for NVIDIA Blackwell Check out the deep dive into the architectural decisions and training methods behind the model 👇 developer.nvidia.com/blog/in…
3
14
1
123
12,486
poggy woggy logan twin, you cooked again fr fr
Say hello to Gemini Embedding 2, our new SOTA multimodal model that lets your bring text, images, video, audio, and docs into the same embedding space! 👀
61
BrotUndSaft retweeted
FlexAttention now has a FlashAttention-4 backend. FlexAttention has enabled researchers to rapidly prototype custom attention variants—with 1000+ repos adopting it and dozens of papers citing it. But users consistently hit a performance ceiling. Until now. We've added a FlashAttention-4 backend to FlexAttention on Hopper and Blackwell GPUs. PyTorch now auto-generates CuTeDSL score/mask modifications and JIT-instantiates FlashAttention-4 for your custom attention variant. The result: 1.2× to 3.2× speedups over Triton on compute-bound workloads. 🖇️ Read our latest blog here: hubs.la/Q045FHPh0 No more choosing between flexibility and performance. hashtag#PyTorch hashtag#FlexAttention hashtag#FlashAttention hashtag#OpenSourceAI
10
98
15
722
101,401
just got this lil guy. Cant wait to tinker around @NVIDIAAIDev
2
48
Habe gestern meinen 23-jährigen Neffen dabei erwischt, wie er nervös auf sein Handy starrte. Er benutzte eine dieser unseriösen "Neobroker"-Apps. Ich fragte streng, was er da treibt. "Ich bespare einen ETF auf den S&P 500", sagte er leise. Mir fiel fast mein Dinkel-Vollkornbrötchen aus der Hand. Ich musste mich setzen. Der Junge investiert sein hart erarbeitetes Geld in amerikanische Firmen ohne physische Präsenz in Castrop-Rauxel? Das ist buchstäblich Glücksspiel. Ich fragte ihn: "Was machst du, wenn die Börse mal zwei Tage geschlossen hat? Oder wenn Elon Musk twittert?" Er hatte keine Antwort. Der pure Leichtsinn. Ich habe ihm sofort das Handy weggenommen und die App gelöscht. Wir sind dann gemeinsam zur Sparkasse gefahren. Haben ein 2-stündiges Beratungsgespräch für einen soliden Bausparvertrag geführt. Abschlussgebühr nur 1,6% der Bausparsumme. Garantierte Verzinsung von 0,25% bis zur Zuteilung in 18 Jahren. Er fing an zu weinen und faselte etwas von "Inflation" und "Opportunitätskosten". Ich legte ihm väterlich die Hand auf die Schulter. "Hör mal zu", sagte ich sanft. "Reichtum entsteht nicht dadurch, dass das Geld mehr wird." "Wahrer Reichtum ist das warme Gefühl, dass dein Geld sicher vor dem amerikanischen Turbokapitalismus auf einem deutschen Girokonto liegt und langsam an Wert verliert." "Aber dafür planbar." Er hat es endlich verstanden. Nie war ich stolzer auf die finanzielle Bildung in diesem Land.
544
478
66
6,889
567,208
Last week I hosted family for Thanksgiving. My 12-year-old nephew asked for the WiFi password. He wanted to play Roblox on his iPad. I looked at the device. Unmanaged. No antivirus. No encryption. I’m an IT Professional. I don't run an open network. So I didn’t give him the password. Instead, I spent 45 minutes provisioning a Guest VLAN. I set up a captive portal. I throttled the bandwidth down to 56kbps. Then I blocked all traffic on ports 80 and 443. He came back crying. He said it wouldn't load. My sister screamed at me to "just let him play." I told her that Zero Trust architecture doesn't care about bloodlines. We didn't have a "fun" Thanksgiving. But we had a secure perimeter. You’re welcome for the compliance.
1,802
5,768
663
113,003
8,930,976
Had a call with a strong engineering candidate today He was two minutes late and looked apologetic. “I’m so sorry for being late!” I said, “That’s okay. Everything alright?” He replied, “Yeah, I was just helping my girlfriend with some errands.” I paused. “You… have a girlfriend?” He nodded. “Yeah, we’ve been together for three years.” I sighed. Red flag. “I’m sorry. I don’t think this will be a good fit. Wish you all the best.” I ended the call immediately. If your software engineer can talk to women, that's not a good sign. The best male engineers can barely make eye contact with one. If your male engineer is in a relationship (or God forbid, married), he’s at best a B-player Always hire A-players
734
1,105
204
31,054
2,203,065