docs: Reddit responses for BillDStrong, ganonfirehouse420, TopChard1274

- BillDStrong: why 1-bit works (softmax robustness, not better MSE)
- ganonfirehouse420: concrete context extension example (32K→128K+)
- TopChard1274: KV compression helps all hardware tiers equally

Total: 6 response drafts covering all substantive Reddit questions.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: unamedkr

docs/pr/2026-04-03-reddit-responses.md

@@ -41,9 +41,57 @@ Clarification:
 
 ---
 
+## @BillDStrong — "What magic is this. I didn't realize there was a 1-bit version"
+
+Good observation — the paper (TurboQuant, ICLR 2026) focuses on 2.5-bit and 3.5-bit configurations. The 1-bit version is our extension of the paper's framework.
+
+The key insight: the paper's RHT (Randomized Hadamard Transform) makes the quantization error **unbiased** for inner products at any bit-width. We pushed this to the extreme — 1 bit = just the sign of each dimension after RHT. Mathematically, this gives a cosine similarity of 2/pi ≈ 0.637 (we measured 0.634), which is the information-theoretic maximum for sign-only quantization.
+
+Why does 1-bit "beat" 2-3 bit? It doesn't in terms of reconstruction quality (MSE is worse). But for **attention scoring** (which only needs inner product ranking, not exact values), the softmax function is surprisingly tolerant of noise. The attention weights after softmax are nearly identical because:
+1. RHT distributes errors uniformly (no systematic bias)
+2. Softmax amplifies the largest scores and suppresses small ones
+3. The top-attended tokens stay the same even with noisy scores
+
+So it's not that 1-bit is "better" — it's that attention is robust enough that 1-bit is sufficient.
+
+---
+
+## @ganonfirehouse420 — "I hope I will be able to have a huge context for my local models in the future"
+
+That's exactly the use case. With 1-bit K + Q4 V, KV cache memory drops ~5x. Concrete example:
+
+```
+Gemma 3 4B at 32K context:
+  FP16 KV: 4,352 MB → barely fits in 16GB with model weights
+  1-bit K + Q4 V: 885 MB → room for 128K+ context on same hardware
+```
+
+For a 16GB Mac or laptop, this means going from 32K → 100K+ context without any hardware upgrade. The limiting factor shifts from KV memory to model weight memory.
+
+This is available today — `./build/tq_run model.gguf -p "your long prompt" -k turbo_kv_1b -v q4 --ctx 131072`. The `--ctx` flag overrides the default context limit.
+
+---
+
+## @TopChard1274 — "big breakthroughs ... seem brutal for people who invested in nearly-unaffordable systems"
+
+Appreciate the perspective. A few thoughts:
+
+1. **KV compression helps everyone** — whether you have 8GB, 24GB, or 80GB. The ratio is the same (5x KV reduction). High-end systems benefit by running longer contexts or larger batches, not just by fitting a model.
+
+2. **This doesn't obsolete hardware** — weight memory is still the bottleneck for model size. A 70B model still needs ~35GB for Q4 weights regardless of KV compression. What changes is that you can push context much further on whatever hardware you have.
+
+3. **The "0.03% quality loss" criticism is fair** — some critics in this thread pushed back on "zero loss" and they're right. We've updated to "almost no quality loss" with exact PPL numbers. The honest framing matters more than hype.
+
+The real unlock is for use cases like RAG with long documents, code analysis with large repos, or multi-turn conversations that previously hit context limits.
+
+---
+
 ## Key takeaway from Reddit feedback
 
 1. **"zero quality loss" was overstated** → fixed to "almost no" with exact PPL
 2. **"why not just integrate into llama.cpp?"** → we have a patch, that's the plan
 3. **Technical curiosity is high** — 5.4K views, people want to understand the math
 4. **Skepticism is healthy** — the Blizado/No-Manufacturer criticism pushed us to be more precise
+5. **1-bit vs 2-3 bit confusion** → clarified: softmax robustness, not better MSE
+6. **Long context is the killer app** — multiple users asking about context extension
+7. **Hardware democratization** resonates — people want more from existing hardware