What could be causing this? Here is my guess on what’s going on:

To get v14 to run on HW3 silicon, Tesla likely had to quantize the model (e.g., down to INT8 precision) to fit within HW3’s 8 GB memory and bandwidth limits. On top of that, because the models have grown so large, HW3 is forced to split compute across both onboard chips to run a single stack rather than running redundant checks, which adds processing latency.

When you quantize a model, you introduce “quantization noise.” In simple terms, minor environmental shifts—like early morning sunlight glare, moving shadows, or passing traffic—can cause the model’s confidence scores to jitter wildly right around the decision threshold (e.g., dropping from 51% to 49% and back every few frames). That rounding fluctuation causes the control system to panic, leading to those quick ~500ms brake pulses at stoplights and sudden brake-stabbing on the road. Subtle details like turn arrows or lane lines also get blurred out by quantization noise, which is compounded by HW3’s lower 1.2-megapixel camera resolution compared to HW4.

What frustrates me is how widespread this seems. Between my 2018 Model 3 and my friend’s 2022/23 Model Y (which actually handles it worse), the issue is constant; like every 30min or so.

If quantization noise, split-compute latency, and reduced model capacity are the root causes, this won’t be a quick fix. As it stands, v14 Lite feels less predictable and less acceptable than v12. I hope Tesla can resolve this by heavily tuning their temporal smoothing filters for HW3, because right now, it makes it really hard to trust FSD.