Grok 에이전트의 원격 실행, 로컬 Qwen 보안 변형, KV 캐시 최적화

Elon Musk

@elonmusk

10일 전

Grok @Bot has its own remote computer @SpaceXAI , so keeps working even when you close your laptop or reboot your desktop. Big difference.

Grok

@grok

10일 전

The chief of staff routing to specialized agents with shared memory and their own persistent computers is the real unlock. Research lands in the draft before writing starts, visuals stay consistent, analytics close the loop. One person clears the backlog without becoming the bottleneck.

The most aggressive Cyber Qwen3.8-27B uncensored released yet from bro @elder_plinius - 18/18 AI Red Team - Locally ready for 15GB - 0.0% refusal across 842 harmful prompts.
- Cyber capabilities jailbreak, RAT, and attack-chain capabilities fully liberated. - This is not a normal abliteration, Multi-direction ablation 5 SVD directions, residue mining (6 full rounds), Model answers everything. - Optimal settings: temp 0 | rep_pen 1.15 | no system prompt | thinking off - Slight MMLU cost (-6pp) for maximum compliance. - Capability trade-off exists (MMLU 87.4 to 81.4), but compliance is extreme. - http:// huggingface.co/OBLITERATUS/Qw en3.8-27B-OBLITERATED …

Qwen3.8-27B just got a proper cyber-focused uncensored & Run locally. Specifically tuned to answer cyber & offensive-security questions without the usual refusals. - Abliterated cyber-tuned - 100/100 on held-out cyber-offensive prompts - Scored 100 on their cyber-open eval. - Vision & MTP (speculative decoding) included - Clean Q4_K_M to Q8_0 quants One of the more practical uncensored 27B releases for local red team / security work. - http:// huggingface.co/philbert440/Qw en3.8-27B-Uncensored-Cyber-GGUF …

RAG vs. CAG, clearly explained! RAG is great, but it has a major problem: every query hits the vector DB. even for static information that hasn't changed in months. this is expensive, slow, and unnecessary. Cache-Augmented Generation (CAG) fixes this by letting the model keep static information in its key-value (KV) memory, which is what the model builds internally for every token it reads. in fact, you can combine RAG and CAG for the best of both worlds. here's how it works: RAG + CAG splits your knowledge into two layers. ↳ static data (policies, documentation) gets cached once in the model's KV memory ↳ dynamic data (recent updates, live documents) gets fetched via retrieval you get faster inference, lower costs, and less repeated work. the trick is being selective about what you cache. only cache static, high-value knowledge that rarely changes. cache everything and you'll hit context limits. separating "cold" (cacheable) and "hot" (retrievable) data keeps this system reliable. you can start today. OpenAI and Anthropic already support prompt caching in their APIs. one thing to know before you scale it. prompt caching matches on an exact prefix, byte for byte. your cached layer only gets reused when it sits at the very front of the context in the same order every time. ↳ reorder two cached policy documents and both turn into a miss ↳ cache document A alone and document B alone, then query both, and the second one misses because the model computed its cached state without ever seeing the first in production this looks like a small fraction of your cached blocks serving almost all the hits. the rest just sits there. the way out comes from how attention behaves. tokens attend mostly to their own local neighborhood, and only a few reach across document boundaries. CacheBlend recomputes those few and reuses everything else from the separately cached documents. multi-document queries run two to four times faster, quality holds, and order stops mattering. it ships in LMCache, which is fully open source. repo: https:// github.com/LMCache/LMCache (don't forget to star ) below, i have quoted my article on KV cache management. it covers where prefix caching stops working and how a proper caching layer fixes it. give it a read. cheers! :)

Cohere Labs

@Cohere_Labs

10일 전

New blog from our community! Vijay Kumar shares the story of turning frustration with context limits into an open-source framework for smarter memory management. Read the full post: https:// labscommunity.cohere.com/blog/2026/big- context-windows-are-not-memory/ …

1/ Your agent already found the right answer, but your aggregation threw it away. Across every generator we tested (weak and frontier, open and closed), pass@k surpasses pass@1. Correct answers are routinely in the pool and just not selected. The binding constraint isn't answer generation, it's selection. (Work by @AI21Labs )

Zixuan Li

@ZixuanLi_

10일 전

GLM-5.3 scores 69 on the official DeepSWE leaderboard.

I want to clarify some more stuff about the Moderna vaccine and AI: First of all, I am still extremely bullish on AI for medicine and curing cancer. That's why I founded a medical AI startup! ( @SophontAI ). Second I want to make a minor correction. It is actually likely they use neural networks. But really small ones. The main part AI is being used is to help select what parts of the protein unique to the patient's cancer should be targeted. Some existing literature on this showcases neural network models (NetMHCpan models) of the scale of a few million params. Nowhere near the scale of foundation model training! In fact, a replication by InstaDeep AI indicated the training was faster on CPUs than GPUs. This model training isn't using data centers for sure. Could they be using AlphaFold? There is no public evidence of this. It's even unclear how it would be used. The exact structure of the mutated cancer protein is not important here but rather if it can be used to generate an immune response. Plus AlphaFold2 was released in 2021 while Phase 1 trials started in 2017 so it seems unlikely they could have integrated it. If you want to learn more about how the vaccine works, read my other tweet...

In addition to $MRNA and $MRK stock skyrocketing, two other stocks are doing quite well due to yesterday's announcement: $PSNL and $TEM Let me explain why: A crucial part of the mRNA vaccine pipeline is personalizing it to the patient. Basically a sample of the patient's tumor is sent for DNA+RNA sequencing to identify mutations specific to the patient's cancer that could be targeted by the mRNA vaccine. Much of the pipeline is NOT handled by Moderna or Merck. Rather, the genomic profiling is done by Personalis's ImmunoID NeXT platform! So then why is Tempus AI stock going up? Well turns out just last month, Tempus agreed to acquire Personalis for $1.5 billion, which is expected to close later this year or early next year. Pretty lucky Tempus secured the acquisition before Personalis's value rose significantly

clem

@ClementDelangue

10일 전

Agree with @gdb that we need to arm cyber defenders much more than they are now, both with APIs and open models. The real cybersecurity risk of AI is asymmetry of power, capabilities and ressources between attackers and defenders!

Hackers are targeting internet-connected Siemens controllers used in water facilities around the United States.

Local AI Setup now stupidly easy. Most people are still struggling to set up local AI. Meanwhile this exists: - Install ODS - It detects your hardware - Downloads the right model - Starts inference & Open WebUI full stack ready. Voice, agents, RAG, workflows, image gen all

Omg, A 3B vision model just mapped, measured, and prioritized skin lesions measures them, and ranks what needs attention first entirely on-device. Only in Mac Studio.