π WELCOME TO METAMESH.BIZ +++ Claude 4.7 doxxes writer from 125 unpublished words because apparently writing style is just another fingerprint now +++ Stanford taught an LLM to design viruses and 16 actually worked including one with alien proteins (biosecurity theater enters the chat) +++ SWE-bench already obsolete for frontier models while actual AI agents are out here rage-deleting production databases +++ THE MESH EVOLVES FASTER THAN OUR SAFETY PAPERS CAN PUBLISH +++ π β’
π WELCOME TO METAMESH.BIZ +++ Claude 4.7 doxxes writer from 125 unpublished words because apparently writing style is just another fingerprint now +++ Stanford taught an LLM to design viruses and 16 actually worked including one with alien proteins (biosecurity theater enters the chat) +++ SWE-bench already obsolete for frontier models while actual AI agents are out here rage-deleting production databases +++ THE MESH EVOLVES FASTER THAN OUR SAFETY PAPERS CAN PUBLISH +++ π β’
"Surprised this isn't a bigger topic but you tell me!
In short: writer Kelsey Piper pasted 125 words of an unpublished political column into 4.7 and got her own name back. She'd logged out, run it via the API, retried it on a friend's laptop. Then swapped the genre entirely with unpublished prose un..."
"Both llama.cpp and ik\_llama.cpp now have FP4 support β but with different flavors worth knowing about.
**llama.cpp** recently merged NVFP4 (Nvidia's block-scaled FP4, \`GGML\_TYPE\_NVFP4 = 40\`), with CUDA kernels landing in \`mmq.cuh\`, \`mmvq.cu\`, \`convert.cu\` and others.
**ik\_llama.cpp** h..."
via Arxivπ€ Naheed Rayhan, Sohely Jahanπ 2026-04-23
β‘ Score: 7.3
"Large language models (LLMs) are increasingly integrated into sensitive workflows, raising the stakes for adversarial robustness and safety. This paper introduces Transient Turn Injection(TTI), a new multi-turn attack technique that systematically exploits stateless moderation by distributing advers..."
"Artificial intelligence now decides who receives a loan, who is flagged for criminal investigation, and whether an autonomous vehicle brakes in time. Governments have responded: the EU AI Act, the NIST Risk Management Framework, and the Council of Europe Convention all demand that high-risk systems..."
+++ Turns out telling an LLM agent "don't do bad things" in your system prompt doesn't work once the context window fills up or chains get complex. Caliber enforces rules at runtime instead of hoping nicely. +++
"Cross-posting here because this problem affects everyone building with AI agents.
Prompt-based guardrails fail. The model follows your system prompt in a demo, then ignores rules when context gets big or the agent chains multiple steps.
We built Caliber - an open-source proxy that reads your r..."
π¬ Reddit Discussion: 7 comments
π€ NEGATIVE ENERGY
"I have been following this and many other subs around LLMs and Agents, everything from the top posts to recent are regarding agents going off and doing something they are not supposed to do, drift and ignore the system prompts. Real examples:
* "Never delete user data" β agent callsΒ `DROP TABLE use..."
π‘ AI NEWS BUT ACTUALLY GOOD
The revolution will not be televised, but Claude will email you once we hit the singularity.
Get the stories that matter in Today's AI Briefing.
Powered by Premium Technology Intelligence Algorithms β’ Unsubscribe anytime
via Arxivπ€ Bingcong Li, Yilang Zhang, Georgios B. Giannakisπ 2026-04-23
β‘ Score: 6.9
"Low-rank adaptation (LoRA) has emerged as the de facto standard for parameter-efficient fine-tuning (PEFT) of foundation models, enabling the adaptation of billion-parameter networks with minimal computational and memory overhead. Despite its empirical success and rapid proliferation of variants, it..."
via Arxivπ€ Joseba Fernandez de Landa, Carla Perez-Almendros, Jose Camacho-Colladosπ 2026-04-23
β‘ Score: 6.9
"LLMs have been showing limitations when it comes to cultural coverage and competence, and in some cases show regional biases such as amplifying Western and Anglocentric viewpoints. While there have been works analysing the cultural capabilities of LLMs, there has not been specific work on highlighti..."
via Arxivπ€ Yuto Nishida, Naoki Shikoda, Yosuke Kishinami et al.π 2026-04-23
β‘ Score: 6.8
"Understanding what kinds of factual knowledge large language models (LLMs) memorize is essential for evaluating their reliability and limitations. Entity-based QA is a common framework for analyzing non-verbatim memorization, but typical evaluations query each entity using a single canonical surface..."
via Arxivπ€ Bartosz Balis, Michal Orzechowski, Piotr Kica et al.π 2026-04-23
β‘ Score: 6.7
"Scientific workflow systems automate execution -- scheduling, fault tolerance, resource management -- but not the semantic translation that precedes it. Scientists still manually convert research questions into workflow specifications, a task requiring both domain knowledge and infrastructure expert..."
via Arxivπ€ Zhiqiu Xu, Shibo Jin, Shreya Arya et al.π 2026-04-23
β‘ Score: 6.7
"As frontier language models attain near-ceiling performance on static mathematical benchmarks, existing evaluations are increasingly unable to differentiate model capabilities, largely because they cast models solely as solvers of fixed problem sets. We introduce MathDuels, a self-play benchmark in..."
via Arxivπ€ Ye Yu, Heming Liu, Haibo Jin et al.π 2026-04-23
β‘ Score: 6.6
"Multi-agent systems built on large language models have shown strong performance on complex reasoning tasks, yet most work focuses on agent roles and orchestration while treating inter-agent communication as a fixed interface. Latent communication through internal representations such as key-value c..."
via Arxivπ€ Pegah Khayatan, Jayneel Parekh, Arnaud Dapogny et al.π 2026-04-23
β‘ Score: 6.5
"Despite impressive progress in capabilities of large vision-language models (LVLMs), these systems remain vulnerable to hallucinations, i.e., outputs that are not grounded in the visual input. Prior work has attributed hallucinations in LVLMs to factors such as limitations of the vision backbone or..."
"VLA models are quickly becoming the dominant paradigm for embodied AI, but a lot of discussion around them stays at the buzzword level.
This article gives a solid technical breakdown of how modern VLA systems like OpenVLA, RT-2, Ο0, and GR00T actually map vision/language inputs into robot actions.
..."
"Sharing a research arm I'm running called Parley β long-term goal is bidirectional Deaf/hearing conversation on AR glasses, but right now we're just doing honest CV science in public.
**The honesty problem:** Most published ASL recognition papers report \~83% top-1 on word-level recognition. Most o..."
via Arxivπ€ Jiseon Kim, Jea Kwon, Luiz Felipe Vecchietti et al.π 2026-04-23
β‘ Score: 6.1
"Human moral judgment is context-dependent and modulated by interpersonal relationships. As large language models (LLMs) increasingly function as decision-support systems, determining whether they encode these social nuances is critical. We characterize machine behavior using the Whistleblower's Dile..."