iFeeling Daily
Daily curated AI insights you can't miss.
Anthropic disputes US government suspension of Fable 5 over minor jailbreak
Anthropic’s statement reveals a critical clash between rapid AI deployment and government oversight: a minor, narrow jailbreak found in Fable 5 led to an abrupt export control suspension, even though similar capabilities exist in other models. The company argues that applying this standard across the industry would halt all new frontier model deployments, underscoring the need for transparent, technically-grounded regulation.
ShinyHunters Zero-Day Oracle PeopleSoft Campaign
Mandiant's detailed breakdown of ShinyHunters' zero-day exploit against Oracle PeopleSoft reveals how attackers used MeshCentral agents and SSH credential spraying to breach educational institutions. Immediate endpoint blocking and log auditing are critical defenses.
Profiling in PyTorch (Part 2): From nn.Linear to a Fused MLP
The article traces the journey from a single nn.Linear to a fused MLP, showing that torch.compile fuses GeLU and multiplication into one Triton kernel while leaving the three cuBLAS GEMMs untouched. It then introduces hand-tuned Liger kernels that achieve the same fusion without compile latency or shape specialization, trading a few microseconds of peak performance for robustness across changing input sizes.
Scaling Past Informal AI: Formal Verification as the Path to Math AGI
Carina Hong argues that reaching math AGI requires formal verification in the training loop, not just bigger models. Axiom Math's system scored a perfect 120/120 on the 2024 Putnam exam using Lean-verified proofs as RL reward signals. The big insight: verification is a scaling property, not a bug fix.
Chinese-Language PhaaS: Real-Time Interception and Digital Wallet Exploitation
Chinese-language PhaaS operators have moved from static password harvesting to real-time interception of OTPs and digital wallet tokenization, enabling direct financial control. Using encrypted channels like RCS and iMessage, along with AI-generated pages, they bypass traditional defenses. Defenders must adopt FIDO2/WebAuthn and risk-based verification during wallet provisioning.
Running Reachy Mini’s voice pipeline fully offline with speech-to-speech
This article walks through running a full speech-to-speech pipeline locally for the Reachy Mini robot, replacing cloud dependencies with a cascaded stack of Silero VAD, Parakeet STT, a local LLM via llama.cpp or vLLM, and Qwen3 TTS. It covers the practical tradeoffs between latency, privacy, and model quality that matter when deploying a voice agent on your own hardware.
Google Cloud’s Autonomous SDLC Security with AI Agents
Google Cloud embeds AI agents across the SDLC—from design review to code scanning and patching—to enable autonomous, machine-speed security defense with a self-reflection loop.
FFASR Leaderboard: Open Benchmark for Far-Field ASR
The FFASR Leaderboard reveals far-field WER at low SNR is consistently several times higher than near-field WER on the same speech content across all models.
Agent Factory: 100X Engineering with AI Agents in Google Antigravity 2.0
Rody Davis explains how to use AI agents to reduce cognitive toil, build skills as cheat sheets, and maintain codebase discipline like a Bonsai artist.