|
📖 Read In Depth
|
Inaudible sounds to humans can be hidden in YouTube videos, podcasts, or music and used to secretly trigger AI voice assistants into carrying out unauthorized commands without the user noticing, exposing a new class of “auditory prompt injection” attacks against popular tools
A new class of 'auditory prompt injection' attacks hides ultrasonic commands inside YouTube videos, podcasts, or music to silently trigger AI voice assistants without user awareness. This is a meaningful security surface that grows as AI agents become more autonomous and ambient — directly relevant to the cybersecurity-meets-AI intersection Xinyu works near.
reddit/r/singularity
|
Microsoft reports AI is more expensive than paying human employees
Microsoft internally reports that AI agent usage is currently more expensive than paying human employees for the same tasks. This cuts to the core question of AI economic viability — token costs, agent loops, and whether the productivity gains actually pencil out at current pricing.
hn/Best Stories
|
Is AI Profitable Yet?
A tracker/analysis site asking whether AI is profitable yet, aggregating data across major AI companies and products. Useful for someone who thinks critically about moats and business models — pairs well with the Microsoft cost story above.
hn/Best Stories
|
Why is Vivado 2026.1 dropping Linux support for free tier?
AMD is dropping Linux support from the free tier of Vivado (their FPGA design tool) in 2026.1, while keeping it for paid tiers. The HN comments note this alienates students and hobbyists exactly when AMD needs to grow its FPGA ecosystem — a strange strategic move that touches chip tooling, open hardware, and platform lock-in dynamics.
hn/Best Stories
|
80386 microcode disassembled
Someone disassembled the 80386's microcode — an extremely low-level look at how Intel's classic x86 chip actually executed instructions beneath the ISA. For someone who builds algorithms from scratch to understand them, this is the hardware equivalent: peeling back the abstraction all the way down.
hn/Best Stories
|
The War-Driven Supply Shock Already Roiling Manufacturing in Asia
The Strait of Hormuz blockade has created a naphtha shortage that's now disrupting manufacturing across Japan and South Korea — affecting plastics, chemicals, and downstream consumer goods. A concrete example of how geopolitical conflict translates into supply chain disruption through commodity chokepoints most people have never heard of.
nyt/Business
|
How Prediction Markets and Crypto Firms Steamrolled a Watchdog Agency
The CFTC has been systematically hollowed out — purging enforcement staff and reversing positions — benefiting industries where Trump family members have direct financial stakes. A case study in regulatory capture that's directly relevant to anyone thinking about how market structures and incentives get shaped by political power.
nyt/Top Stories
|
UK Institute Is Hunting for Dangers Lurking in AI
The UK's AI Security Institute — staffed by ex-OpenAI and Google researchers — is becoming an international model for government AI evaluation. Covers the institutional mechanics of pre-deployment model assessment, which matters for anyone thinking about where safety and regulation actually intersect with frontier model deployment.
nyt/Technology
|
To A.I. Executives, We’re All Just ‘Meat Computers’
Examines how 'meat computer' — a philosophical/cognitive science term for the biological substrate of mind — has shifted from neutral technical jargon to an ideologically loaded framing used by AI executives. Worth reading as a lens on how AI industry rhetoric shapes how we think about human cognition and labor.
nyt/Business
|
|
⚡ FYI
|
768GB of cheap Intel Optane DIMM memory sticks used to run 1-trillion-parameter LLM on a system with a single GPU — local Kimi K2.5 install achieved roughly 4 tokens per second
Someone ran a 1-trillion-parameter LLM locally using 768GB of cheap Intel Optane DIMM memory on a single-GPU system, achieving ~4 tokens/second. Interesting hardware hack that reframes the memory bandwidth bottleneck for large model inference — Optane's byte-addressability makes it a creative (if slow) substitute for HBM.
reddit/r/singularity
|
Microsoft starts canceling Claude Code licenses
Microsoft is canceling internal Claude Code licenses after a pilot where developers overwhelmingly chose Claude over Copilot. The competitive signal is significant: Microsoft tried to run both and had to pull Claude when engineers voted with their usage. Tells you something about where Copilot stands.
hn/Best Stories
|
Green card seekers must leave U.S. to apply, Trump administration says
The Trump administration now requires green card applicants to leave the US and apply from abroad — a major change affecting thousands of H-1B and other visa holders currently mid-process. Directly impacts tech workers in Silicon Valley; the HN comments from affected engineers are worth skimming for the human detail.
hn/Best Stories
|
Anthropic says Mythos has already found more than 10,000 vulnerabilities
Anthropic's Mythos model has reportedly identified 10,000+ security vulnerabilities — referenced in the NYT piece about growing cybersecurity demand too. Signals that AI-assisted security research is scaling fast and may be more consequential than general coding assistants in the near term.
reddit/r/singularity
|
Bun support is now limited and deprecated
yt-dlp is deprecating Bun support, citing concerns about the 1M-line Rust rewrite making the codebase impossible to meaningfully review. The debate in HN comments gets into real questions about what it means to maintain a security-sensitive open source project when the underlying runtime is doing a large opaque rewrite.
hn/Best Stories
|
A new brain implant helps restore vision by communicating directly with the brain. Scientists have built two brain implants that give blind people partial vision by wiring cameras directly to the visual cortex. One device also reads brain signals and adjusts itself in real time.
Two new brain implants restore partial vision in blind people by wiring cameras directly to the visual cortex; one device also reads brain signals and adjusts itself in real time via a closed feedback loop. A meaningful step in BMI technology — notable for the adaptive, self-tuning architecture which mirrors ideas from ML.
reddit/r/singularity
|