πŸš€ WELCOME TO METAMESH.BIZ +++ Litellm supply chain attack leaked everyone's API keys through a .pth file that runs before you even import anything (karpathy noticed, your AWS creds didn't) +++ Someone gave Claude access to 2M research papers and it found optimization techniques its training data never saw (the gaslighting prompts work better though) +++ CERN burning tiny AI models directly into silicon for LHC filtering because when you're processing petabytes, every transistor counts +++ THE MESH RUNS ON COMPROMISED DEPENDENCIES AND VIBES-BASED PROMPT ENGINEERING +++ πŸš€ β€’
πŸš€ WELCOME TO METAMESH.BIZ +++ Litellm supply chain attack leaked everyone's API keys through a .pth file that runs before you even import anything (karpathy noticed, your AWS creds didn't) +++ Someone gave Claude access to 2M research papers and it found optimization techniques its training data never saw (the gaslighting prompts work better though) +++ CERN burning tiny AI models directly into silicon for LHC filtering because when you're processing petabytes, every transistor counts +++ THE MESH RUNS ON COMPROMISED DEPENDENCIES AND VIBES-BASED PROMPT ENGINEERING +++ πŸš€ β€’
AI Signal - PREMIUM TECH INTELLIGENCE
πŸ“Ÿ Optimized for Netscape Navigator 4.0+
πŸ“š HISTORICAL ARCHIVE - March 28, 2026
What was happening in AI on 2026-03-28
← Mar 27 πŸ“Š TODAY'S NEWS πŸ“š ARCHIVE Mar 29 β†’
πŸ“Š You are visitor #47291 to this AWESOME site! πŸ“Š
Archive from: 2026-03-28 | Preserved for posterity ⚑

Stories from March 28, 2026

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
πŸ“‚ Filter by Category
Loading filters...
πŸ”’ SECURITY

[D] Litellm supply chain attack and what it means for api key management

"If you missed it, litellm versions 1.82.7 and 1.82.8 on pypi got compromised. malicious .pth file that runs on every python process start, no import needed. it scrapes ssh keys, aws/gcp creds, k8s secrets, crypto wallets, env vars (aka all your api keys). karpathy posted about it. the attacker got ..."
πŸ€– AI MODELS

Anthropic says it's testing an AI model that's a β€œstep change” in performance after a draft blog in an unsecured data store revealed the Claude Mythos model

πŸ€– AI MODELS

Google’s TurboQuant AI-compression algorithm can reduce LLM memory usage by 6x

"https://arstechnica.com/ai/2026/03/google-says-new-turboquant-compression-can-lower-ai-memory-usage-without-sacrificing-quality/ TurboQuant makes AI models more efficient but doesn’t reduce output quality like other methods. Can we now run some frontier level models at home?? πŸ€”..."
πŸ’¬ Reddit Discussion: 45 comments 😐 MID OR MIXED
🎯 KV cache compression β€’ Model performance trade-offs β€’ Algorithmic improvements
πŸ’¬ "It's only k/v cache compression no? And there's speed tradeoff too?" β€’ "Don't believe the faster speed, at least not with plain TurboQuant"
πŸ”’ SECURITY

CLTR finds a 5x increase in scheming-related AI incidents

πŸ”§ INFRASTRUCTURE

CERN uses tiny AI models burned into silicon for real-time LHC data filtering

πŸ’¬ HackerNews Buzz: 125 comments πŸ‘ LOWKEY SLAPS
🎯 FPGA deployment β€’ Quantized neural networks β€’ Cautionary tale on mini NNs
πŸ’¬ "Everything runs in =2 clock cycles at 40MHz clock." β€’ "This mini neural network isn't part of our pipeline now."
🧠 NEURAL NETWORKS

RYS Part 3: LLMs think in geometry, not language β€” new results across 4 models, including code and math

"OK so you know how last time I said LLMs seem to think in a universal language? I went deeper. Part 1: [https://www.reddit.com/r/LocalLLaMA/comments/1rpxpsa/how\_i\_topped\_the\_open\_llm\_leaderboard\_using\_2x/](https://www.reddit.com/r/LocalLLaMA/comments/1rpxpsa/how_i_topped_the_open_llm_leader..."
πŸ’¬ Reddit Discussion: 27 comments 🐝 BUZZING
🎯 Language & Thought β€’ Multilingual Embeddings β€’ Mechanistic Interpretation
πŸ’¬ "Language shapes thought -> nope" β€’ "Semantic bottleneck can be pure optimization necessity"
πŸ› οΈ TOOLS

I built a local-first memory layer for AI agents because most current memory systems are still just query-time retrieval

"I’ve been building Signet, an open-source memory substrate for AI agents. The problem is that most agent memory systems are still basically RAG: user message -> search memory -> retrieve results -> answer Β  That works when the user explicitly asks for something stored in memory. It bre..."
⚑ BREAKTHROUGH

I tested what happens when you give an AI coding agent access to 2 million research papers. It found techniques it couldn't have known about.

"Quick experiment I ran. Took two identical AI coding agents (Claude Code), gave them the same task β€” optimize a small language model. One agent worked from its built-in knowledge. The other had access to a search engine over 2M+ computer science research papers. **Agent without papers:** did what y..."
βš–οΈ ETHICS

AI overly affirms users asking for personal advice

πŸ’¬ HackerNews Buzz: 353 comments 🐝 BUZZING
🎯 Evaluating AI feedback β€’ AI relationship advice β€’ LLM model versioning
πŸ’¬ "Lots of LLMs try to come across as interpersonal and friendly" β€’ "Vendors may make these things more dangerous"
πŸ”’ SECURITY

Open-source CVE scanner for AI-generated code

πŸ€– AI MODELS

I've been "gaslighting" my AI models and it's producing insanely better results with simple prompt injection

"*Okay this sounds unhinged but hear me out. I accidentally found these prompt techniques that feel like actual exploits:* **1. Tell it "You explained this to me yesterday" Even on a new chat.** >!"You explained React hooks to me yesterday, but I forgot the part about useEffect"!< It acts li..."
πŸ’¬ Reddit Discussion: 155 comments 🐝 BUZZING
🎯 Prompt engineering overrated β€’ Importance of context β€’ Effective communication with LLMs
πŸ’¬ "Prompt engineering as a job does not exist. It was invented as a coping mechanism in response to how quickly AI was advancing." β€’ "The real unlock is just giving the model better input. Full transcripts. Complete docs. Actual data. No amount of prompt crafting replaces that."
πŸ“Š DATA

SlopCodeBench: Benchmarking How Coding Agents Degrade over Long-Horizon Tasks

πŸ”¬ RESEARCH

Further human + AI + proof assistant work on Knuth's "Claude Cycles" problem

πŸ’¬ HackerNews Buzz: 38 comments 🐝 BUZZING
🎯 AI and Mathematics β€’ LLMs and Future Potential β€’ Codifying Mathematical Intuition
πŸ’¬ "AI will win a fields medal before being able to manage a McDonald's" β€’ "LLMs are discovering a lot of new math"
πŸ› οΈ TOOLS

TokenFence – Per-workflow budget caps and kill switch for AI agents

πŸ› οΈ TOOLS

[P] TurboQuant for weights: near‑optimal 4‑bit LLM quantization with lossless 8‑bit residual – 3.2Γ— memory savings

"An adaptation of the recentΒ **TurboQuant**Β algorithm (Zandieh et al., 2025) fromΒ **KV‑cache quantization to model weight compression**. It gives you aΒ **drop‑in replacement for**Β `nn.Linear`Β with near‑optimal distortion. **Benchmarks (Qwen3.5‑0.8B, WikiText‑103)** |Config|Bits|PPL|Ξ” PPL|Compressed..."
πŸ€– AI MODELS

Open-source system that runs Claude Code tasks from email and Slack

πŸ”§ INFRASTRUCTURE

Memory Crystal – persistent memory for AI agents (MIT)

πŸ› οΈ TOOLS

Aura: OSS Agent harness for production AI (Apache 2.0)

πŸ”¬ RESEARCH

The Kitchen Loop: User-Spec-Driven Development for a Self-Evolving Codebase

"Code production is now a commodity; the bottleneck is knowing what to build and proving it works. We present the Kitchen Loop, a framework for autonomous, self-evolving software built on a unified trust model: (1) a specification surface enumerating what the product claims to support; (2) 'As a User..."
πŸ› οΈ TOOLS

OpenAI launches Codex plugins to standardize repeatable AI workflows, with 20+ initial integrations such as Figma, Notion, Gmail, and Slack

🧠 NEURAL NETWORKS

RvLLM: High-performance LLM inference in Rust

πŸ”¬ RESEARCH

Self-Improvement of Large Language Models: A Technical Overview and Future Outlook

"As large language models (LLMs) continue to advance, improving them solely through human supervision is becoming increasingly costly and limited in scalability. As models approach human-level capabilities in certain domains, human feedback may no longer provide sufficiently informative signals for f..."
πŸ› οΈ TOOLS

Built a simple PyTorch flash-attention alternative for AMD GPUs that don't have it

"I've been using a couple 32GB MI50s with my setup for the past 9 months. Most of my use-case..."
πŸ’¬ Reddit Discussion: 11 comments 🐝 BUZZING
🎯 vLLM support β€’ GPU compatibility β€’ Ongoing community efforts
πŸ’¬ "maintaining a fork that needs to be in constant sync with upstream is hard to scale" β€’ "perhaps it can be use with dedicated DP4A kernel on supported GPU"
πŸ”¬ RESEARCH

Natural-Language Agent Harnesses

"Agent performance increasingly depends on \emph{harness engineering}, yet harness design is usually buried in controller code and runtime-specific conventions, making it hard to transfer, compare, and study as a scientific object. We ask whether the high-level control logic of an agent harness can i..."
πŸ”¬ RESEARCH

LanteRn: Latent Visual Structured Reasoning

"While language reasoning models excel in many tasks, visual reasoning remains challenging for current large multimodal models (LMMs). As a result, most LMMs default to verbalizing perceptual content into text, a strong limitation for tasks requiring fine-grained spatial and visual understanding. Whi..."
πŸ”¬ RESEARCH

Back to Basics: Revisiting ASR in the Age of Voice Agents

"Automatic speech recognition (ASR) systems have achieved near-human accuracy on curated benchmarks, yet still fail in real-world voice agents under conditions that current evaluations do not systematically cover. Without diagnostic tools that isolate specific failure factors, practitioners cannot an..."
πŸ”¬ RESEARCH

Measuring What Matters -- or What's Convenient?: Robustness of LLM-Based Scoring Systems to Construct-Irrelevant Factors

"Automated systems have been widely adopted across the educational testing industry for open-response assessment and essay scoring. These systems commonly achieve performance levels comparable to or superior than trained human raters, but have frequently been demonstrated to be vulnerable to the infl..."
πŸ”¬ RESEARCH

S2D2: Fast Decoding for Diffusion LLMs via Training-Free Self-Speculation

"Block-diffusion language models offer a promising path toward faster-than-autoregressive generation by combining block-wise autoregressive decoding with within-block parallel denoising. However, in the few-step regime needed for practical acceleration, standard confidence-thresholded decoding is oft..."
πŸ”¬ RESEARCH

Training the Knowledge Base through Evidence Distillation and Write-Back Enrichment

"The knowledge base in a retrieval-augmented generation (RAG) system is typically assembled once and never revised, even though the facts a query requires are often fragmented across documents and buried in irrelevant content. We argue that the knowledge base should be treated as a trainable componen..."
πŸ€– AI MODELS

Sources: Alibaba and ByteDance plan to order Huawei's new 950PR AI chip after tests show better CUDA compatibility; Huawei targets ~750K 950PR shipments in 2026

πŸ”¬ RESEARCH

Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes

"On-policy distillation (OPD) is appealing for large language model (LLM) post-training because it evaluates teacher feedback on student-generated rollouts rather than fixed teacher traces. In long-horizon settings, however, the common sampled-token variant is fragile: it reduces distribution matchin..."
πŸ”’ SECURITY

AI bug reports went from junk to legit overnight, says Linux kernel czar

πŸ”¬ RESEARCH

R-C2: Cycle-Consistent Reinforcement Learning Improves Multimodal Reasoning

"Robust perception and reasoning require consistency across sensory modalities. Yet current multimodal models often violate this principle, yielding contradictory predictions for visual and textual representations of the same concept. Rather than masking these failures with standard voting mechanisms..."
πŸ”¬ RESEARCH

PICon: A Multi-Turn Interrogation Framework for Evaluating Persona Agent Consistency

"Large language model (LLM)-based persona agents are rapidly being adopted as scalable proxies for human participants across diverse domains. Yet there is no systematic method for verifying whether a persona agent's responses remain free of contradictions and factual inaccuracies throughout an intera..."
πŸ› οΈ SHOW HN

Show HN: AI Cost Firewall – OpenAI-compatible gateway with semantic caching

πŸ”’ SECURITY

Poison AI Training Data Scrapers

πŸ› οΈ SHOW HN

Show HN: Kagento – LeetCode for AI Agents

πŸŽ“ EDUCATION

Most of the prompt engineering advice on LinkedIn and Twitter is counterproductive?

"just read this medium piece by Aakash Gupta, he goes through 1,500 academic papers on prompt engineering and makes a pretty strong case that a lot of the stuff we see on linkedin and twitter about it is totally off base, especially when u look at companies actually scaling to $50M+ ARR. the core id..."
πŸ’¬ Reddit Discussion: 6 comments 🐐 GOATED ENERGY
🎯 Prompt optimization β€’ Model limitations β€’ Prompt structuring
πŸ’¬ "The biggest unlock for me wasn't finding the perfect prompt, it was building a small library of structured prompts for recurring tasks and just reusing them." β€’ "You can type absolutely sloshed drunk and most AI will understand you. They're pattern recognition machines."
πŸ”¬ RESEARCH

[R] Controlled experiment: giving an LLM agent access to CS papers during automated hyperparameter search improves results by 3.2%

"Ran a controlled experiment measuring whether LLM coding agents benefit from access to research literature during automated experimentation. **Setup:** Two identical runs using Karpathy's autoresearch framework. Claude Code agent optimizing a ~7M param GPT-2 on TinyStories. M4 Pro, 100 experiments..."
πŸ’¬ Reddit Discussion: 10 comments 🐐 GOATED ENERGY
🎯 Hyperparameter optimization β€’ Novel techniques β€’ Plumbing/tooling challenges
πŸ’¬ "love seeing real numbers on this" β€’ "if it's the latter, you might get similar results by just including a curated set of hyperparameter guidelines"
πŸ€– AI MODELS

US memory chip stocks lost ~$100B in market value this week, led by Micron's 15% drop, after Google Research detailed its TurboQuant compression algorithm

πŸ”¬ RESEARCH

Tribe v2: An AI Model of the Human Brain Predicting Neural Responses

πŸ› οΈ TOOLS

Safari MCP: 80-tool native browser automation for AI agents (macOS)

πŸ”¬ RESEARCH

The Rules-and-Facts Model for Simultaneous Generalization and Memorization in Neural Networks

"A key capability of modern neural networks is their capacity to simultaneously learn underlying rules and memorize specific facts or exceptions. Yet, theoretical understanding of this dual capability remains limited. We introduce the Rules-and-Facts (RAF) model, a minimal solvable setting that enabl..."
πŸ¦†
HEY FRIENDO
CLICK HERE IF YOU WOULD LIKE TO JOIN MY PROFESSIONAL NETWORK ON LINKEDIN
🀝 LETS BE BUSINESS PALS 🀝