4 stories
The author used the shape and perspective clues in a photograph to narrow the world, then used CUDA to compare geometry across possible locations.
Its value is methodological: the answer matters less than making the chain of inference reproducible enough for someone else to challenge.
The top HN threads: facts from the discussions, both sides, insider voices. Email, Telegram, or RSS.
His thesis lands as a rule: 'truly robust engineering isn't about what works for most; it's about gracefully handling the edge cases.' The thread split on whether that framing is even right.
The sharpest replies went past the post itself: 98% quietly compounds. At 99% right per step a 20-pane dashboard breaks nearly one load in five, and an LLM that nails 99% of tool calls still fumbles almost every session.
In LLM token generation the GPU often sits idle, not for lack of work but because the CPU hasn't finished housekeeping for the next token. Moondream calls that gap a GPU bubble and rebuilt its Photon engine to hide it, overlapping the next forward pass with the sampling of the previous one.
The payoff scales with the machine. On an RTX 3090 the trick adds about 6.5%; on a B200, where CPU sync is a bigger slice of a tiny 2.4ms forward pass, it adds 35.4%. An inference engineer pushed back: the win looks this big only because the model is small.
Modular's pitch was a portable compiler stack that breaks NVIDIA's CUDA lock-in; founder Chris Lattner built LLVM, Clang and Swift before this. The buyer designs ARMv9 chips and has almost no presence in the H100-class training market.
The thread reads it two ways. One commenter notes Modular's founder repeatedly wrote that hardware firms fail at AI software, making the acquirer the punchline. Another sees a portfolio play, lining Modular up beside Tenstorrent and Ventana as Qualcomm pushes from phone chips toward RISC-V and datacenter AI.
