From Neighborhood Watch to Bird Watching
A blogger points his existing security cameras at the birds instead of the intruders and gets a local, cloud-free identification system out of BirdNET-Go. The Hacker News thread responds with frog misidentifications, e-ink displays, and a running joke about Flock cameras finally living up to their name.
OpenAI Built Its Own Chip. It's Called Jalapeño.
OpenAI designed an inference chip from scratch in 16 months and claims it beats Nvidia's Blackwell on performance per watt. SemiAnalysis got to benchmark it - and Hacker News split between 'token prices will plummet' and 'they'll keep the margin.'
The Robot That Played Tennis Against a Champion
A humanoid robot returned 50 km/h serves from former Grand Slam champion Zheng Jie, fell mid-rally, and got back up. Galaxy General calls it embodied AI's AlphaGo moment.
At the Robot Fair, Everyone's Doing the Math Now
China's World Robot Conference stopped being a talent show. The booths that drew crowds were the ones proving they could work a full shift - sorting parcels, peeling cucumbers, paying for themselves.
The First Humanoid Robot Stock Opened at +629%
Unitree listed on Shanghai's STAR Market and closed day one up 460%. The market's yardstick for robots just moved from demos to delivery.
A Ten-Cent Chip Is RISC-V's Best Defense
Dmitry Grinberg's RISC-V takedown made the rounds this week. An embedded engineer in Trinidad answered with a shipping bill: $60 to move a one-dollar chip, and a ten-cent chip that settled the argument.
One Man, a Mac, and a 33B Video Model
antirez - the Redis creator who stepped away in 2020 - just shipped a native Metal inference engine for MiniMax H3. On a 128 GB M5 Max it renders end-to-end video in about 75 seconds with a 40 GB footprint, and Hacker News noticed.
The Phone That Replaced the VPS
One self-hoster rooted his CMF Phone 1, kept Android, and turned a drawer phone into his whole server stack - Termux, runit, Tailscale, and no systemd in sight.
Too Fast to Be Thinking
AMD just bought Taalas, the Toronto startup that etches model weights directly into silicon. Its test chip served Llama 3.1 8B at 16,960 tokens a second - 48x faster than Nvidia's GPUs when it was announced - and the live demo was so instant that people stopped asking how fast it was and started asking whether it was thinking at all.
The SSD Is the New VRAM
An 80B Qwen runs in 4.3 GB of RAM on a Mac, and a 35B runs natively on an iPhone. Swiftlet streams a model's MoE experts straight from SSD - and Hacker News can't decide if it's progress or a NAND-burner.
A 27B Model That Fits on a Phone
PrismML's Bonsai 27B uses ternary weights to squeeze a 27-billion parameter model into 5.9 GB — running natively on iPhones and laptops with surprisingly little capability loss.
WiFi That Sees Through Walls — RuView Turns Signals Into Spatial Intelligence
A 79K-star GitHub project does what cameras can't: reading human presence, movement, and even vital signs from ambient WiFi signals.