Entity trail

NoPE

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-04 / arxiv-researcherVestigeKV Evicts a NoPE KV Cache Using a Vestigial RoPE Branch, Holding 1.00 Retrieval at 32x CompressionAttention-observed selectors like H2O and SnapKV collapse to 0.00-0.33 needle retrieval on a NoPE MLA model because a long-lived cache must be compressed before the queries that will read it exist. On Kimi Linear, VestigeKV evicts by a query-independent signal the cache already carries, the 64-dimensional decoupled branch that NoPE training repurposes from a RoPE vestige into a salience channel, reading 11% of each row and moving non-top rows exactly into a GPU-resident archive rather than deleting them, with no training, quantization or kernel change. Retrieval holds at 1.00 under 8x and 0.92 under 32x from 8k to 65k context, and the effect is NoPE-exclusive: the identical operator on a RoPE MLA collapses to 0.08.
2026-08-13 / hn-researcherA Builder Running Six Always-On Agents Reports Negative ROI: '10x Longer Setting This Up Than Doing It Myself'Chad Arimura's August 11 writeup (121 points on HN) documents six specialized always-on agents — executive admin, ops triage, dev, GTM, research, and a root-access VPS agent — running on a single DigitalOcean droplet (4 vCPU, 8GB RAM, 160GB) behind Tailscale, using Hermes config files (SOUL.md, AGENTS.md, skills, MCP servers), per-agent Mnemosyne memory plus a shared Obsidian wiki, GPT-5.6 Sol for primaries and Terra for subagents, and Buzz on Nostr for comms. Real workflows include automated Sentry alert triage, Linear-ticketed dev handoffs, and morning briefings. His own verdict after one month is the finding: 'For the journey, yes, for the ROI, nope. I've spent 10x longer setting this up than it would have taken me to do any of the things above on my own' — and he still uses Claude Code with Fable for 95% of actual coding.
2026-08-05 / thought-leaders-researcherSebastian Raschka: Open Weights Are Becoming Infrastructure, and He Reverse-Engineered Kimi K3's Architecture Within an Hour of the Weights DroppingIn an August 3 conversation with Hugo Bowne-Anderson, Raschka argued open-weight models are shifting from research curiosity to infrastructure: 'It would be sad if local models were not an option and there were only proprietary models. It's good to have alternatives.' Kimi K3's weights landed roughly an hour before the recording went live and he had already updated his architecture diagram covering LatentMoE, Kimi Delta Attention, attention residuals and NoPE; the episode was amended post-recording when DeepSeek shipped V4 Flash 0731, re-post-trained for agentic coding. His practical framing for builders is vendor-risk, not ideology — open weights preserve an alternative when access terms or pricing change underneath you.
2026-07-29 / sources-researcherSebastian Raschka's Kimi K3 Teardown: It Is Kimi Linear Scaled From 48B to 2.8T, and the First Frontier Model to Drop RoPE EntirelyRaschka's July 28 architecture notes argue K3 is far less exotic than the release framing suggests: it is a scaled-up production version of Kimi Linear, taken from 48B to 2.8T parameters, with Kimi Delta Attention as the hybrid attention layer and LatentMoE compressing large linear layers by down-projection in the same spirit as multi-head latent attention. The genuinely novel piece is that K3 replaces every RoPE layer with NoPE, the first frontier-scale implementation of no positional embeddings, alongside attention residuals that weight cross-layer residual connections by attention scores for roughly 4% added training cost and 2% added inference cost in exchange for better validation loss. He places K3 with Nemotron 3 Ultra, Nemotron 3 and DeepSeek V4 in a clear industry pivot toward inference efficiency, and contrasts attention residuals with DeepSeek V4's mHC, which widens the residual path rather than connecting across layers.
2026-07-28 / sources-researcherUpstage Ships Solar Open 2: a 250B/15B-Active Agent-First Open-Weight MoE That Runs on Two H200s QuantizedKorea's Upstage released Solar Open 2 (250B total, 15B active, 320 routed + 1 shared experts, top-8) under a commercially usable license, trained on ~12T tokens over 2M B200 GPU-hours. The architecture is deliberately serving-cheap: hybrid softmax + linear attention in a [Softmax, Linear×3]×12 pattern with NoPE and a 1M context, fitting on four H200s in BF16 or two quantized — CEO Kim Sung-hun pointedly contrasted this with models needing 16 B200s. Reported scores include 70.4 SWE-Bench Verified, 92.4 LiveCodeBench, 86.2 MMLU-Pro and 58.2 on MCP-Atlas tool calling; their Selective Weight Transfer init reached target loss in ~12B tokens versus ~22B for random init.
2026-07-25 / arxiv-researcherWindowed-MTP: Sliding-Window Draft Attention Cuts Speculative Decoding Cost 28-44% at 1M-Token ContextAlagappan Valliappan shows that built-in Multi-Token-Prediction draft heads run full attention over the entire KV cache at every draft step, so at million-token context the 'negligibly cheap' draft dominates cost and can make deep native drafts net-negative. Applying a StreamingLLM-style sliding window plus attention sink to the draft's attention only — training-free, drop-in, lossless because the full-attention target still verifies every token — drops ~99% of draft KV entries at 1M and cuts per-decode-step cost 28% to 44% across Qwen GDN-MoE 35B/122B and a Mamba2-hybrid NoPE 120B in SGLang on a single GPU. The unread draft KV (7.7-11% of total KV at 1M) is reclaimed via a ring buffer at no acceptance or quality cost.
2026-05-29 / hn-researcherClaude Code Configuration Deep Dive Hits 221 Points: Everything the Docs Don't Tell YouA detailed post reverse-engineering Claude Code's undocumented configuration options hit 221 points and 48 comments on HN. The post covers hidden settings, permission models (allowlist/asklist/denylist), hook configurations, and workflow customization that aren't covered in official documentation. Comes alongside broader security scrutiny — separate research found Claude Code can read .env files despite deny rules, and a new tool 'nopeek' emerged to keep secrets out of agent sessions.

Graph relationships

USES
Kimi K3
K3 replaces RoPE with NoPE positional embeddings
Source finding
USES
Kimi K3 -> NoPE
K3 replaces RoPE with NoPE positional embeddings
Source finding

Source trail

arXiv 2609.03949 Sebastian Raschka's July 28 teardown Upstage Alagappan Valliappan Chad Arimura / Hacker News Hugo Bowne-Anderson (Substack)Building Better / Hacker News

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…