Skip to content

Log

Append-only record of wiki operations.

2026-05-29 — Auto-update

  • Topics covered: LLM/AI, Networking, Rocket
  • New pages: none
  • Key finds:
  • LLM: OpenDeepThink (arXiv:2605.15177) — population-based parallel reasoning via Bradley-Terry pairwise tournament selection; +405 Codeforces Elo for Gemini 3.1 Pro in 8 rounds; composable with EDRM routing layer
  • Networking: Astera Labs Computex June 3 press conference confirmed — first public demo of PCIe 6 scale-up optics (COSMOS); all three PCIe 6 fabric switch architectures (Astera, Marvell, Panmnesia) simultaneously visible at Computex
  • Rocket: IREC 2026 video deadline today (May 29, 5pm CT); all documentation gates closed; entering final hardware integration phase before June 11 Midland shipment

2026-05-28 — Auto-update

  • Topics covered: LLM/AI, Philosophy, NDT, Rocket
  • New pages: none
  • Key finds:
  • LLM: EDRM (arXiv:2605.22873) — entropy phase transitions reveal LLM reasoning is a dynamic decoding state, not static; training-free EDRM routing achieves 41–55% token reduction across 15 benchmarks; composable with UCCI cascade routing
  • Philosophy: Donald Robertson "Five Stoic Reminders" (Substack, May 27) — community validation of prohairesis, ephexis, Golden Verses of Pythagoras; mirrors CBT-validated Stoic practice from peer-reviewed literature
  • NDT: NDTiC 2026 concludes today (Day 3 of 3, CINDE Whistler); APCNDT + NDTiC proceedings expected NDT.net June–July 2026
  • Rocket: IREC 2026 Team Video submission due tomorrow May 29; all doc gates closed; hardware integration phase begins

2026-05-27 — Auto-update

  • Topics covered: LLM/AI, NDT
  • New pages: none
  • Key finds:
  • LLM: LLM-Wiki (arXiv:2605.25480, May 25, 2026): Retrieval-as-Reasoning agent-native retrieval — structured Wiki pages with bidirectional links, iterative search+read+link-follow loop, Error Book self-correction; outperforms HippoRAG 2/LightRAG/GraphRAG by 2.0–8.1 F1 on multi-hop reasoning
  • NDT: NDTiC 2026 Day 2 of 3 (CINDE, Fairmont Whistler); final day May 28; post-conference proceedings expected on NDT.net
  • SDR: GNU Radio 4 community stewardship (May 21) already documented — MIT-licensed GR4 under community governance, stable 4.0.0.0 preparation underway

2026-05-26 — Auto-update

  • Topics covered: LLM/AI, NDT
  • New pages: none
  • Key finds:
  • LLM: VeriCache (arXiv:2605.17613, U.Chicago/Tensormesh/Samsung/MSR): first lossless LLM inference framework — uses compressed KV cache as drafter + full KV (off-GPU) for verification; 4× throughput of full-KV decoding with identical outputs; eliminates catastrophic failures of lossy KV compression in code gen and tool calling
  • NDT: NDTiC 2026 Day 1 of 3 opened today (CINDE, Fairmont Whistler, May 26–28); proceedings to follow post-conference

2026-05-25 — Auto-update

  • Topics covered: LLM/AI, Rocket, NDT
  • New pages: none
  • Key finds:
  • LLM: GRAM (arXiv:2605.19376, KAIST/Mila/NYU): stochastic recursive reasoning via amortized variational inference — multi-hypothesis latent trajectories, inference scales depth×breadth; probabilistic counterpart to the Solve the Loop / EqR attractor cluster
  • LLM: Interpretable Latency Model for Speculative Decoding (arXiv:2605.15051, MIT/Red Hat): Little's Law applied to LLM serving — explains why SD speedups collapse under high server load; production deployment theory for SD configuration
  • Rocket: May 29 deadline clarified as short video submission for IREC livestream (not a technical gate); full June competition timeline confirmed (June 11–20 Midland TX)
  • NDT: NDTiC 2026 Day 2 of 3 (CINDE's 50th anniversary, Fairmont Whistler)

2026-05-24 — Auto-update

  • Topics covered: LLM/AI, Rocket, NDT
  • New pages: none
  • Key finds:
  • LLM: Attractor Models cluster — "Solve the Loop" (arXiv:2605.12466, USC, May 12): 770M Attractor Model beats 1.3B Transformer via fixed-point implicit-diff training + equilibrium internalization; "Equilibrium Reasoners" (arXiv:2605.21488, Kolter et al., May 20): EqR scales depth×breadth to 40,000 effective layers → 2.6%→>99% on Sudoku-Extreme
  • Rocket: IREC 2026 late-phase submission due May 29 (T-5); PR3 closed May 20; competition June 15–20 Midland TX
  • NDT: NDTiC 2026 "Taking NDT to New Heights" opens today (May 26–28 Whistler BC); APCNDT 2026 proceedings on NDT.net expected June–July 2026

2026-05-23 — Auto-update

  • Topics covered: LLM/AI, Satellite (status)
  • New pages: none
  • Key finds: • FlexDraft (arXiv:2605.20022, May 19): batch-adaptive parallel speculative decoding — attention tuning + bonus-guided calibration + flex decoding strategy switching; closes the batch-scale calibration failure mode in parallel SD • LEOPARDSat-1 + PROVES (on ISS since April 11) — still awaiting deployment and first signal confirmation • NDT in Canada 2026 (NDTiC, May 26–28 Whistler) opens in 3 days; AGL/Networking/SDR pages remain current from prior runs

2026-05-22 — Auto-update

  • Topics covered: LLM/AI, SDR/GCS, NDT
  • New pages: none
  • Key finds: • GNU Radio 4 Community Stewardship (May 21): MIT licensing for GR4 core runtime + blocks, stable 4.0.0.0 preparation — removes LGPL barrier for commercial deployments; GSI/FAIR governance split clarified • SpecKV (arXiv:2605.02888, May 4): compression-aware adaptive gamma selection for speculative decoding — 56% improvement over fixed-gamma=4, closing the KV compression → speculative decoding interface gap • UCCI (arXiv:2605.18796, May 11): calibrated LLM cascade routing via isotonic regression — 31% cost cut at matched accuracy; NDT in Canada 2026 opens May 26 in Whistler BC

2026-05-21 — Auto-update

  • Topics covered: LLM/AI, Rocket
  • New pages: none
  • Key finds: • FreeKV (arXiv:2505.13109, May 19): training-free KV cache retrieval — speculative reuse + cosine correction + hybrid CPU/GPU layout; 13× speedup over SOTA at near-lossless accuracy; completes the four-layer KV-cache optimization stack (RateQuant + FibQuant + KVDrive + FreeKV) • AIAA ASCEND 2026 concluded Day 3 (May 21) — Aerospace Corporation Space Transformation Track + NISAR Pickering Lecture; PR3 closed May 20; next IREC milestone May 29 • Philosophy: Stoic Arts 2026 afternoon session still not published; no new philosophy content this run

2026-05-20 — Auto-update

  • Topics covered: LLM/AI, Rocket
  • New pages: none
  • Key finds: • KVDrive (arXiv:2605.18071, May 18): multi-tier KV cache management (GPU→DRAM→SSD) — 1.74× throughput vs SOTA; systems complement to RateQuant+FibQuant quantization stack • IREC 2026 PR3 deadline today (May 20, 5pm CST) — all 175 teams' avionics documentation locks; AIAA ASCEND Innovators Day (Day 2) ongoing in DC with ESRA at Booth 306 • Stoic Arts afternoon session still not published; morning session (Sharon Lebell + MacMillans) confirmed on YouTube; philosophy status quo maintained

2026-05-19 — Auto-update

  • Topics covered: LLM/AI, Rocket
  • New pages: none
  • Key finds: • FibQuant (arXiv:2605.11478): vector quantization for KV-cache — proves strict improvement over scalar quantization; 3.6× lower perplexity than TurboQuant at 8× compression; stackable with RateQuant • Deep Reasoning / DOLORES (arXiv:2605.11388): inference-time scaffold construction via structured meta-cognition — 24.8% avg improvement over fixed scaffolding; 8B model beats 32B baselines in half of scenarios • AIAA ASCEND 2026 in progress (Washington DC, May 19–21) — Jared Isaacman opened; ESRA at Booth 306; PR3 due tomorrow May 20 5pm CST

2026-05-18 — Auto-update

  • Topics covered: LLM/AI
  • New pages: none
  • Key finds: • SlimSpec (arXiv:2605.10453): low-rank drafter LM-head compression — 4-5× acceleration, 8-9% end-to-end speedup vs competing methods, drop-in for EAGLE-3 • PARD-2 (arXiv:2605.08632): CAT-alignment + dual-mode speculative decoding — 6.94× lossless speedup, 1.9× over EAGLE-3 on Llama3.1-8B • SlimSpec + PARD-2 + PPOW form a stackable three-paper frontier for EAGLE-3-class speculative decoding improvements

2026-05-17 — Auto-update

  • Topics covered: LLM/AI, NDT, Rocket
  • New pages: none
  • Key finds: • arXiv:2605.06241 — "Rethinking RL for LLM Reasoning": RL performs sparse policy selection at 1-3% of token positions (high-entropy branch points), not capability teaching — reframes RL's role in reasoning improvement • APCNDT 2026 concluded May 15 (5 days, Honolulu) — open-access proceedings coming to NDT.net June-July 2026 • IREC 2026 registration closed May 15; PR3 window open T-3 days (due May 20); ESRA at AIAA ASCEND May 19-20

2026-05-16 — Auto-update

  • Topics covered: LLM/AI, Philosophy
  • New pages: none
  • Key finds: • arXiv:2605.14978 — PPOW: RL-based speculative decoding drafter with adaptive divergence-aware windowing; 3.39–4.36× speedup, avg accepted length 6.29–6.52 tokens — direct architectural follow-up to the Attention Drift paper • Robertson "Stoicism for Mothers" (Substack) — Domitia Lucilla as Marcus Aurelius's first Stoic model; virtue transmission through lived example rather than doctrine • IREC 2026 PR3 due May 20 (confirmed); Computex June 2-5 approaching (Astera Labs Scorpio X demo)

2026-05-15 — Auto-update

  • Topics covered: LLM/AI, Philosophy
  • New pages: none
  • Key finds: • arXiv:2605.09992 — Attention Drift in speculative decoding (Northwestern/GE Aerospace): attention progressively moves from prompt to drafter's own recent tokens as speculation chains lengthen; explains accuracy degradation at longer chain depths across EAGLE3 and MTP drafters • Donald Robertson "How to Do Stoic Therapy" (Conversations with Modern Stoicism) completed today (May 15); recording now available on his Substack — applied Stoic CBT protocol session with Phil Yanov • IREC 2026 registration closes TODAY (May 15, 5pm CST); PR3 open through May 20

2026-05-14 — Auto-update

  • Topics covered: LLM/AI, AGL/IVI, NDT (status update)
  • New pages: none
  • Key finds: • arXiv:2605.04266 — Alignment Collapse in Iterative RLHF (Bach + Jordan): Stackelberg game formulation reveals policy exploits reward model blind spots in iterative RLHF; FPO fix restores missing steering term • AGL SoDeV "Ultimate Unagi" released today (May 14) — first UCB with SoC hardware support (Renesas Sparrow Hawk), MediaTek joins as member, 2-year LTS support window • APCNDT 2026 concluded today (May 14, Honolulu) — proceedings expected on NDT.net June–July 2026

2026-05-13 — Auto-update

  • Topics covered: LLM/AI, NDT (status update)
  • New pages: none
  • Key finds:
  • Federation of Experts (arXiv:2605.06206): MoE-cluster-per-KV-head restructuring eliminates all-to-all comm in single-node; confines it to intra-node in multi-node — drop-in distributed inference optimization for MoE LLMs
  • Transformers Provably Implement ICRL (arXiv:2605.05755): first formal proof + convergence guarantee that a linear self-attention block implements semi-gradient SARSA/actor-critic — theoretical grounding for why LLMs can do in-context RL
  • APCNDT 2026 now Day 3 of 4 (concludes May 14 Honolulu); proceedings via ASNT/NDT.net open access expected June–July 2026

2026-05-12 — Auto-update

  • Topics covered: LLM/AI, Philosophy, Rocket, NDT (status update)
  • New pages: none
  • Key finds:
  • RateQuant (arXiv:2605.06675): rate-distortion-optimal KV cache quantization — 70% perplexity reduction vs KIVI at 2.5 avg bits, 1.6s calibration, zero inference overhead
  • Grok 4.3 (May 6–8): #1 CaseLaw v2 + CorpFin, 40% price cut to $1.25/$2.50 MTok; GPT-5.5 Instant (May 5): 52.5% fewer hallucinations, new default ChatGPT model
  • Donald Robertson "How to Do Stoic Therapy" live stream announced for May 15 — Stoic Arts afternoon session still unpublished

2026-05-11 — Auto-update

  • Topics covered: LLM/AI, NDT (status update), Philosophy (status update)
  • New pages: none
  • Key finds:
  • Can RL Teach Long-Horizon Reasoning? (arXiv:2605.06638): ScaleLogic shows RL effort scales as power law ~2.60 with proof depth; more expressive training (FOL vs implication-only) yields +10.66 pts transfer — partial rehabilitation of RL as capability-teaching tool for expressive training regimes
  • APCNDT 2026 NOW OPEN (Day 1, May 11–14 Honolulu) — world's largest APAC NDT conference under ASNT/NDT.net open-access MOU; AI-NDT industrial papers expected on NDT.net June–July 2026
  • Stoic Arts 2026 afternoon session still unpublished as of May 11; morning session (incl. Sharon Lebell music/Stoicism) accessible on YouTube

2026-05-10 — Auto-update

  • Topics covered: LLM/AI, Rocket, NDT (cleanup)
  • New pages: none
  • Key finds:
  • BALAR (arXiv:2605.05386): Bayesian Agentic Loop selects clarifying questions by maximizing expected mutual information — training-free principled active questioning for interactive agents
  • LatentRAG (arXiv:2605.06285): moves agentic RAG reasoning+retrieval to latent space — 90% latency reduction vs explicit agentic RAG at same accuracy
  • ESRA at AIAA ASCEND May 19–21 Washington DC — IREC 2026 community briefing, week before competition

2026-05-09 — Auto-update

  • Topics covered: LLM/AI, Philosophy, Networking, Satellite
  • New pages: none
  • Key finds:
  • LLM RL reframe (arXiv:2605.06241): RL only touches 1–3% of token positions — it's sparse policy selection not capability learning; ReasonMaxxer achieves same gains at 1000× lower training cost
  • Stoic Arts 2026 Morning session posted to YouTube ~May 5 — Sharon Lebell music/Stoicism lecture now accessible (14-day delay post-event)
  • SNAPPY 3U CubeSat (Wichita State): world's first space-based neutrino detector, launched May 3 on SpaceX CAS500-2

2026-05-08 — Auto-update

  • Topics covered: LLM/AI, Philosophy, NDT, Satellite, Rocket
  • New pages: none
  • Key finds:
  • LLM Strategic Reasoning (arXiv:2605.00226): two-gap failure model — Observation-Belief Gap (internal reps more accurate than verbalized) + Belief-Action Gap (implicit beliefs underperform explicit ones in decisions); systematic vulnerabilities in LLM strategic play
  • NCKU Gemini-Pollux 3U CubeSat launched May 4 on Falcon 9 from Vandenberg into 590 km LEO; entirely faculty/student-built; ionospheric monitoring + Earth imaging via amateur radio
  • APCNDT 2026 opens May 11 Honolulu (3 days); "Breaking Barriers" theme; proceedings open-access via NDT.net in June–July

2026-05-07 — Auto-update

  • Topics covered: LLM/AI, Networking
  • New pages: none
  • Key finds:
  • When to Think When to Speak (arXiv:2605.03314): disclosure timing as a generation-quality lever — SxS Interleaved Reasoning decouples deliberation from public commitment; SFT+RL on entailment-aligned trajectories
  • Astera Labs Scorpio X-Series (May 5, 2026): 320-lane PCIe 6 memory-semantic AI fabric switch; hardware-accelerated Hypercast+In-Network Compute boost collectives 2×; production ramp 2H 2026, $20B TAM by 2030
  • APCNDT 2026 (May 11-14 Honolulu) opens in 4 days; proceedings to be published open-access on NDT.net

2026-05-06 — Auto-update

  • Topics covered: LLM/AI
  • New pages: none
  • Key finds:
  • The Reasoning Trap (arXiv:2605.01704): information-theoretic proof that multi-agent debate degrades evidence grounding per round (DPI Bound); MAD preserved 88% accuracy but dropped reasoning quality 43%; EGSR alternative recovers 98%
  • Search-Driven Reward RL (arXiv:2605.02073): treat reward function design as an optimization problem — LLM generates/ranks candidate rewards via GRPO runs; F1 0.609→0.795 on GSM8K
  • Kimi K2.6 (Moonshot AI, April 20): Agent Swarm scales 3× to 300 sub-agents / 4,000 coordinated steps; PARL RL framework for orchestrator training; current industrial frontier for multi-agent RL

2026-05-05 — Auto-update

  • Topics covered: LLM/AI
  • New pages: none
  • Key finds:
  • Latent-GRPO (arXiv:2604.27998): stable RL for latent reasoning — 3 stabilization techniques; +7.86 Pass@1 on GSM8K-Aug, +4.27 on AIME vs baselines; 3-4× shorter reasoning chains; continuous embedding space more efficient than discrete token space
  • CoLaR (arXiv:2505.16552): dynamic latent compression of reasoning chains — 14.1% accuracy gain over latent baselines, 53.3% chain length reduction with 4.8% performance loss vs CoT, 82.8% reduction with 5.4% accuracy gain on math; inference-time speed/depth control via prompting
  • APCNDT 2026 (Honolulu, May 11–14): conference opens in 6 days; proceedings will post to NDT.net post-conference

2026-05-04 — Auto-update

  • Topics covered: LLM/AI
  • New pages: none
  • Key finds:
  • ResRL (arXiv:2605.00380, ICML 2026): RL reasoning with negative-sample projection residuals — solves Lazy Likelihood Displacement, +9.4% math reasoning Avg@16, +7.0% Pass@128 vs NSR baseline; preserves generation diversity while improving reasoning
  • Odysseus (arXiv:2605.00347): first open framework to scale VLM decision-making to 100+ consecutive turns via PPO with turn-level critic; 3× in-game progress vs frontier models; practical milestone for long-horizon embodied agents
  • APCNDT 2026 (Honolulu, May 11–14): conference opens in 7 days; ASNT+NDT.net open-access MOU means proceedings will be freely published post-conference; GRCon 2026 confirmed Sep 21–24 Raleigh NC (NC State/AERPAW)

2026-05-03 — Auto-update

  • Topics covered: LLM/AI, Rocket Avionics
  • New pages: none
  • Key finds:
  • RecursiveMAS (arXiv:2604.25917, April 28): multi-agent recursive latent-space collaboration — 8.3% accuracy gain, 1.2×-2.4× speedup, 34.6%-75.6% token reduction vs text-based MAS; eliminates inter-agent text serialization bottleneck
  • Mistral Medium 3.5 (April 29): 128B dense open-weight model, 256K context, 77.6% SWE-bench Verified, EAGLE speculative decoding — most deployable open-weight frontier coder on release
  • ESRA Newsletter April 28: IREC 2026 registration closes May 15, Progress Report 3 due May 20; Karman Space & Defense joins as Bronze Sponsor

2026-05-02 — Auto-update

  • Topics covered: LLM/AI, Rocket Avionics
  • New pages: none
  • Key finds:
  • BenchGuard (arXiv:2604.24955): automated benchmark auditing finds 12 fatal errors in ScienceAgentBench — agent evaluation infrastructure has unaudited correctness blind spots; 83.3% expert-match on BIXBench
  • Scheduler-Theoretic Framework for LLM Agent Execution (arXiv:2604.11378): Agent Loops formally equivalent to a single-ready-unit scheduler; SGH proposes immutable DAG execution plans with separated planning/execution/recovery layers
  • ASU SEDS first hybrid rocket at IREC 2026: avionics bay reprinted 6× under monthly DTEG regulatory updates — concrete case study of regulation-driven hardware iteration costs

2026-05-01 — Auto-update

  • Topics covered: LLM/AI, Satellite, Philosophy
  • New pages: none
  • Key finds:
  • Recursive Language Models (RLM, arXiv:2512.24601): Prime Intellect's "paradigm of 2026" — LLM actively manages its own context via recursive delegation to scripts/sub-LLMs; 2.6× improvement on CodeQA vs. GPT-5 base; RL-trained context folding is the next frontier
  • FCC EPFD Report & Order FCC 26-26 voted April 30 — retires 1990s interference limits, replaces with performance-based GEO/NGSO coordination; $2B+ US benefit, 7× LEO broadband capacity; Part 100 licensing reform separate and still pending
  • Stoic Arts Conference (April 25, Tufts): recordings still not published as of May 1 — six days post-event; Sharon Lebell keynote video pending on Modern Stoicism YouTube

2026-04-29 — Auto-update

  • Topics covered: LLM/AI, Philosophy, Rocket Avionics
  • New pages: none
  • Key finds:
  • DeepSeek V4-Pro-Max (extended reasoning mode): HMMT 2026 Feb 95.2, IMOAnswerBench 89.8 — first open-weight model at near-parity with GPT-5.4 on competition math; 10-20× cheaper at equivalent reasoning budget
  • ContextWeaver (arXiv:2604.23069): dependency-graph memory for agents improves SWE-Bench Verified/Lite vs sliding-window baseline while reducing reasoning steps and token usage
  • Stoic attitudes/mental health (Philosophical Psychology 2026): higher SABS Stoicism scores → fewer negative + more positive mental health outcomes; Stoic cognitive reappraisal (not suppression) is the active mechanism
  • IREC 2026: multiple May milestones (May 1, 15, 20, 29) now active — 175 teams entering maximum-intensity pre-competition verification phase

2026-04-28 — Auto-update

  • Topics covered: LLM/AI, AGL/IVI
  • New pages: none
  • Key finds:
  • DeepSeek V4-Pro released April 24 — 1.6T params/49B active MoE, 80.6% SWE-bench Verified (0.2pp behind Claude Opus 4.6), Codeforces 3,206 (world #1), 1M context at 27% V3.2 FLOPs; trained on Huawei Ascend, open-source
  • DiP-SD (arXiv:2604.20919): distributed pipelined speculative decoding for edge inference — device-level distributed drafting + phase-level pipeline overlap
  • AAOS 25Q2 deep-dive: Minimal Telephony HAL (data-only TCUs), HD Radio EAS API, Scalable UI configurable windowing, dynamic vehicle property config, 8 vehicle properties now public API

2026-04-27 — Auto-update

  • Topics covered: LLM/AI, Philosophy
  • New pages: none
  • Key finds:
  • KnowRL (arXiv:2604.12627): minimal-sufficient knowledge guidance for RL training — +9.63 pts accuracy without knowledge hints at inference (70.08% baseline → 74.16% with hints); Constrained Subset Search identifies minimal interacting KP subsets
  • Stoic Arts Conference concluded April 25 at Tufts — no recordings published yet; Sharon Lebell keynote video expected on Modern Stoicism YouTube within days/weeks
  • FCC EPFD vote pending April 30 — performance-based interference framework replacing 1990s static limits, projected 7× capacity increase and $2B economic value; Viasat/DirecTV opposition noted

2026-04-26 — Auto-update

  • Topics covered: LLM/AI, AGL/IVI, Networking
  • New pages: none
  • Key finds:
  • DeepSeek V4-Flash: 13B active params, 79.0% SWE-bench Verified, $0.14/$0.28/MTok — 12× cheaper than V4-Pro at 1.6-point benchmark cost
  • AAOS 25Q2 AOSP source drop live (updated April 10) — vehicle properties promoted to public APIs, windowing framework components now configurable; governance question partially resolved (openly accessible suggests Apache 2)
  • CXL 3.1 confirmed mainstream in 2026 — 15-20% hyperscaler TCO reduction, Linux auto-NUMA enumeration, Marvell Structera S 30260 is the enabling switch (Q3 2026 sampling)

2026-04-25 — Auto-update

  • Topics covered: LLM/AI, NDT
  • New pages: none
  • Key finds:
  • DeepSeek V4 officially released April 24 — 1.6T MoE, 49B active, 1M context, 10% KV cache vs V3, MIT license, fully Ascend-trained; 80.6% SWE-bench Verified
  • SemanticALLI (arXiv:2601.16286): pipeline-aware agentic caching elevates intermediate reasoning representations to cacheable artifacts — 83.1% hit rate vs 38.7% baseline, 78.4% token reduction
  • KRISS waveguide ultrasonic sensor (March 2026): remote 360° UT detection without direct probe attachment — eliminates blind spots in extreme-environment inspection

2026-04-24 — Auto-update

  • Topics covered: LLM/AI, Satellite
  • New pages: none
  • Key finds:
  • Google launched Deep Research + Deep Research Max agents on Gemini 3.1 Pro (April 22) — MCP support, native charts, 93.3% DeepSearchQA, 85.9 BrowseComp
  • GPS III constellation complete: final satellite SV10 "Hedy Lamarr" launched April 21 on Falcon 9 — 32-satellite constellation, full M-code coverage
  • DeepSeek V4 still delayed (third delay, Huawei Ascend chip bottleneck), end-of-April/early-May target

2026-04-23 — Auto-update

  • Topics covered: LLM/AI, SDR/GCS, Satellite, Rocket, Philosophy (cleanup)
  • New pages: none
  • Key finds: • KV-Direct (arXiv:2603.19664): The KV cache is entirely redundant in transformer inference — keys/values are deterministic projections of the residual stream. KV-Direct stores residual vectors (5 KB/token) instead of full KV pairs (136 KB/token) at 100% token-match accuracy vs. 5-28% for all eviction baselines. • BPS.Space Mach 1.8 failure (Jan 2026): Premature deployment from a boolean logic inversion that sent accelerometer velocity into the barometric path — producing negative velocity read as post-apogee. Canonical case study for Mach-immune event detection design. • DeepSeek V4 imminent (late April 2026): 1T param MoE with 37B active, 1M context, trained entirely on Huawei Ascend 950PR — second consecutive open-weight frontier model with zero NVIDIA dependency.

2026-04-22 — Auto-update

  • Topics covered: LLM/AI
  • New pages: none
  • Key finds: • Meta Muse Spark (April 8): Meta's first closed-source proprietary frontier model, built in secret by Meta Superintelligence Labs under Alexandr Wang — natively multimodal, powers Meta AI across all apps; marks a strategic break from Meta's open-weight Llama brand. • GLM-5.1 (Z.AI, April 7): open-weight 754B MoE with MIT license, trained solely on Huawei Ascend chips — held SWE-bench Pro #1 (58.4%) for 9 days before Claude Opus 4.7; signals China's Huawei hardware can now produce frontier coding quality. • Tina (arXiv:2504.15777, ICLR 2026): 1.5B reasoning model fine-tuned via LoRA under RL at $9 total cost (260× cheaper than full-parameter RL), reaching 43.33% AIME24 — makes RL reasoning fine-tuning accessible to any team.

2026-04-20 — Auto-update

  • Topics covered: LLM/AI, NDT, Networking Hardware
  • New pages: none
  • Key finds: • Claude Opus 4.7 now leads SWE-bench Verified at 87.6% (up from Gemini 3.1 Pro's 78.80%); adaptive thinking replaces manual budget controls; Claude Mythos confirmed SWE-bench 93.9% / GPQA Diamond 94.6%; Ouro LoopLM 2.6B matches 12B SOTA via looped inference. • NVIDIA-Marvell $2B deal ties NVLink Fusion to Marvell's CXL/PCIe switching portfolio; Beluga paper demonstrates CXL memory pooling as practical LLM KV-cache tier — CXL switching now has a clear production killer app. • APCNDT 2026 proceedings going open-access via ASNT/NDT.net MOU (May 11–14, Honolulu); AI reliability framework for eddy current rail inspection formally codifies regulatory evidence requirements beyond accuracy metrics.

2026-04-19 — Auto-update

  • Topics covered: Philosophy, LLM/AI, NDT, Networking Hardware
  • New pages: none
  • Key finds: • Claude Opus 4.7 GA (April 16) — xhigh reasoning for autonomous coding agents; Looped Reasoning LLMs (arXiv:2604.11791) — fixed-point convergence per cycle, mechanistic grounding for iterative inference without parameter scaling. • Pigliucci "How to Be a (Happy) Skeptic" (Penguin 2026) — Academic Skepticism / Cicero as practical philosophy beyond Stoicism; Lightbody "Vignettes of the Absurd" — second-person standpoint rescues meaning from the cosmic viewpoint problem, bridges Camus and Levinas structurally. • Marvell Structera S 60260 (PCIe 6.0, 260-lane, Q3 2026 sampling) — sibling to S 30260 CXL; CAJMNS TOFFD+CNN PAUT achieves 94.7% on aerospace CFRP composites vs 78.3% traditional.

2026-04-18 — Auto-update Run 3 (Full)

  • Topics covered: Philosophy, LLM/AI, NDT, AGL/IVI, Networking Hardware, SDR/GCS, Satellite, Rocket Avionics
  • New pages: none
  • Cleanup: removed test lines from satellite.md (lines ~214-215) and log.md (lines ~153-156) from partial Run 2
  • Key finds: • Stoic Arts Conference (Tufts, April 25 2026) program confirmed — first major event linking Stoicism with aesthetic/artistic practice, challenging the "productivity self-help" reduction; Sharon Lebell keynote on music and Stoicism. • Dual-Trace Encoding for LLM Agent Memory (arXiv:2604.12948): 73.7% vs. 53.5% cross-session recall — semantic + episodic trace fusion directly mirrors cognitive neuroscience memory model. • ASNT AI/ML NDT standard under review (NDT Week 2026 Jan focus) + Google-Intel April 9 multi-year IPU partnership: custom ASIC IPU co-development signals hyperscalers building their own DPU silicon.

2026-04-18 — Auto-update

  • Topics covered: LLM/AI, Networking Hardware, AGL/IVI, Satellite Development, SDR/GCS
  • New pages: none
  • Key finds: • Claude Mythos Preview + Project Glasswing (Anthropic, April 7): frontier model autonomously found thousands of zero-days across all major OSes/browsers including 17-year-old FreeBSD RCE — restricted availability; marks the first publicly documented large-scale autonomous offensive security research by an AI system. • Broadcom Project Glasswing (ASIC, April 2026): "glass pipeline" — fully transparent programmable 51.2 Tbps switch ASIC where every stage is operator-defined P4, <200ns overhead vs. fixed-function; Q3 2026 samples. Most ambitious programmable switching silicon yet. • FCC correction: April 30 vote is EPFD spectrum-sharing framework replacement, NOT Part 100 (still NPRM). Also: CRS-24 (April 11) delivered LEOPARDSat-1 + PROVES constellation + HUCSat to ISS; CANVAS confirmed operational April 16.

2026-04-17 — Auto-update (Full)

  • Topics covered: Philosophy, LLM/AI, NDT, AGL/IVI, Networking Hardware, Satellite Development, Rocket Avionics (SDR/GCS: nothing new)
  • New pages: none
  • Key finds: • "Speaking to No One" (Brosnahan/Lipinska, arXiv:2604.10833): conversational AI generates "ontological dissonance" — structural conflict between simulated relational presence and absent subject, can stabilize into technologically mediated folie à deux; explicit disclaimers provably fail. Sharpest clinical account of pixel-alterity dynamics to date. • Tri-Spirit (arXiv:2604.13757): hardware cognitive decomposition for agents (planning/reasoning/execution across substrates + habit compilation of repeated reasoning paths) achieves 75.6% latency reduction, 77.6% tasks offline — argues cognitive layer decomposition, not model scaling, is the primary lever for production agent efficiency. • OCI-MSA v1.0 spec published at oci-msa.org — MSA status upgraded from "announced" to "spec live," enabling formal multi-vendor interoperability testing between AMD/Broadcom/Meta/MSFT/NVIDIA/OpenAI products.

2026-04-17 — Auto-update (NDT / AGL-IVI / Networking)

  • Topics covered: NDT, AGL/IVI, Networking Hardware
  • New pages: none
  • Key finds: • Waygate + GE Aerospace MDI templates for GEnx engines (April 10, 2026): first AI-guided structured borescope inspection workflow in commercial airline deployment — AMM-aligned, 3D measurement, cloud traceability via InspectionWorks Insight. Marks the XAI-NDT commercial deployment milestone tracked from last run. • Renault Trafic E-Tech UK debut confirmed (CV Show, April 2026): 800V, 450 km WLTP, 90% OTA-updateable — first AAOS SDV production commercial vehicle confirmed for late-2026 production at Sandouville. No AUTOSAR coexistence blueprint published yet. • XPO MSA (Arista) hits 100-partner milestone (April 6, 2026): 12.8 Tbps liquid-cooled pluggable form factor, 40+ optical module vendors committed — complementary to OCI-MSA (die-level CPO) vs. XPO (chassis pluggable); production 2027.

2026-04-16 — Auto-update

  • Topics covered: Philosophy, LLM/AI, NDT, AGL/IVI, Networking Hardware, Satellite Development, Rocket Avionics (SDR/GCS: no new finds beyond existing entries)
  • New pages: none
  • Key finds: • Scranton AI Ethics Conference (April 16–18): Catholic intellectual tradition + AI ethics enters the continental/analytic debate — natural law perspective on moral agency and human autonomy in AI systems, live today • Silicon Mirror (arXiv:2604.00478): 85.7% sycophancy reduction on Claude Sonnet 4 via runtime behavioral gating — inference-time solution requiring no retraining; pairs with Calibration Collapse (already in wiki) as detection + mitigation • OCI-MSA formed at OFC 2026 (AMD/Broadcom/Meta/MSFT/NVIDIA/OpenAI): hyperscaler-driven MSA for optical compute interconnects (GEN1 200 Gbps/direction, GEN2 400 Gbps/direction, 3.2 Tbps roadmap) is the de-facto CPO convergence event; Marvell Structera S 30260 CXL 3.0 switch (260 lanes, 4 TB/s) sampling Q3 2026 closes CXL 3.x adoption question

2026-04-15 — Auto-update (Full)

  • Topics covered: Philosophy, LLM/AI, NDT, AGL/IVI, Networking Hardware, SDR/GCS, Satellite Development, Rocket Avionics
  • New pages: none
  • Key finds: • Stiegler/Malabou on Generative AI (PhilPapers): Stiegler predicts LLMs amputate human plasticity by externalising writing; Malabou's counter — brain retains irreducible formative plasticity. Sharpest continental-philosophy framing of LLM cognitive consequences in current literature. • Amazing Agent Race (arXiv:2604.10261): 55–100% of existing agent benchmarks are linear 2–5-step chains; on real DAG tasks navigation errors dominate (27–52%) vs. tool errors (<17%); best accuracy only 37.2% — current agents are weak navigators, not weak tool-callers. • EASA NPA 2025-07 in consultation — first binding aviation AI regulation; Broadcom Taurus BCM83640 (3nm, 400G/lane) enables 3.2T roadmap — full 1.6T/3.2T transceiver supply chain closed in 2026.

2026-04-15 — Auto-update (NDT / AGL-IVI / Networking)

  • Topics covered: NDT, AGL/IVI, Networking Hardware
  • New pages: none
  • Key finds:
  • EASA NPA 2025-07 in consultation — first binding AI aviation regulation; second NPA (NDT/maintenance-specific) due 2026; confirmed dual-certification asymmetry with FAA for AI-NDT vendors
  • Broadcom Taurus BCM83640 (3nm, 400G/lane DSP, OFC March 2026): enables 1.6T modules and 3.2T roadmap (204.8T switch fabric); HyperLight TFLN 400G/lane PICs entering UMC 8" HVM simultaneously — full supply chain closed
  • Google AAOS SDV April 2026: open-source Android expanded to seat actuators, climate, lighting, telemetry, OTA; no AUTOSAR coexistence blueprint published yet; AOSP release later 2026

2026-04-15 — Auto-update (SDR/GCS + Satellite + Rocket only)

  • Topics covered: SDR & GCS, Satellite Development, Rocket Avionics
  • New pages: none
  • Key finds: • CANVAS CubeSat (VLF radiation belt science) launched April 7 aboard Minotaur IV STP-S29A — CU Boulder's 3rd satellite of 2026, demonstrates sustained university production cadence • ScienceDirect ISL LEO survey (March 4, 2026): heterogeneous laser/RF ISL topology now the design norm for LEO mega-constellations — direct GCS architecture implication for multi-hop TM/TC routing • BRIN Biak RPP completed inter-ministry harmonization, derivative regs in progress — RPP presidential signature now the critical path item for Indonesian rocketry licensing

2026-04-14 — Auto-update

  • Topics covered: Philosophy, LLM/AI, NDT, AGL/IVI, Networking Hardware, SDR/GCS, Satellite Development, Rocket Avionics
  • New pages: none (all updates to existing pages)
  • Key finds: • ByteRover (arXiv 2604.01599) + Sparse Frontier (arXiv 2504.17768): agent-native memory without vector DBs; sparse attention Pareto-improves dense at 32K but degrades at 128K — key scaling law for long-context inference • Raytheon/AFRL TFLN production contract (Feb 2026): domestic U.S. TFLN wafer supply chain under AFRL contract, G&H Cleveland at LRIP — China single-supplier dominance now a formal national security priority, accelerating TFLN commercialization • Levinasian HRI ethics (Discover Robotics 2026): "Who is exposed here?" before "What is the correct rule?" — distributed AI/robot responsibility resists legal attribution; interval preservation as concrete design goal

[2026-04-09] INIT — Wiki created

Operation: Initial wiki build By: Ductor (Claude Sonnet 4.6) Pages created: - index.md - projects/index.md, pkm-explorer.md, pasarwatch.md, kait2026.md, ductor.md - research/index.md, llm-wiki.md, edge-ai.md - markets/index.md, ihsg.md - tools/index.md, stack.md - log.md

Seeded from: MAINMEMORY.md + active conversation context Pattern: Karpathy LLM Wiki (three-layer architecture)

2026-04-13 — Auto-update

  • Topics covered: Philosophy, LLM/AI, NDT, AGL/IVI, Networking Hardware, SDR/GCS, Satellite Development, Rocket Avionics
  • New pages: none (all updates to existing pages)
  • Key finds: • CDLM (Together AI / SqueezeAI): consistency diffusion LLMs achieve 14.5× inference speedup on coding benchmarks — surpasses DFlash 6×, now deployed in production • Pleiades Five: 5-university CubeSat constellation (Northeastern, Columbia, et al.) deployed from ISS April 8, 2026 — first student-built ISL experiment, THIS-SAT THz ISL follow-on in 2027 • ENPULSION: 280+ electrospray units in orbit, 5 units/week serial production — electrospray thruster commercialization confirmed for CubeSat class

2026-04-12 — Auto-update

  • Topics covered: Philosophy, LLM/AI, NDT, AGL/IVI, Networking Hardware, SDR/GCS, Satellite Development, Rocket Avionics
  • New pages: none (all updates to existing pages)
  • Key finds: • Birkbeck Stoicism study: first controlled RCT of Stoic practice — 18% rumination reduction, 15% self-efficacy gain vs. control (cognitive training design, 8 days, n-groups) • DFlash (NYU): block diffusion achieves 6× lossless LLM inference speedup (2.5× over EAGLE-3), flat draft cost regardless of block length — ready for SGLang/vLLM integration • Broadcom TH6-Davisson in production (March 2026): 102.4 Tbps CPO switch, 200G/lane, 3.5W per 800G port, Meta logged zero link flaps across 1M device hours

2026-04-11 — Auto-update

  • Topics covered: Philosophy, LLM/AI, NDT, AGL/IVI, Networking Hardware, SDR/GCS, Satellite Development, Rocket Avionics
  • New pages: none (all updates to existing pages)
  • Key finds: • SHAPE (arXiv 2604.06636): hierarchical credit assignment in RLHF achieves 3% accuracy gain + 30% token reduction — verbosity vs. quality distinction is now a solved problem with math backing • AGL SoDeV: Honda+Panasonic led SDV platform (VirtIO+Xen+Zephyr) flips AGL from IVI-only to full ECU consolidation reference platform — available early 2026 • NVIDIA Spectrum-X Photonics CPO switch (H2 2026): 3.5× power efficiency over pluggable, 400 Tb/s, marks commercial CPO inflection point

2026-04-10 — Auto-update

  • Topics covered: Philosophy, LLM/AI, NDT, AGL/IVI, Networking Hardware
  • New pages: research/philosophy.md, research/ndt.md, industries/agl-ivi.md, networking/hardware.md
  • Key finds: • ML in phased array UT lifted defect detection precision from 78% → 93.3% — labeled data is now the bottleneck, not models • RLHF statistical paper (arXiv 2604.02507) gives recovery guarantees for mixed human/AI preference data — critical for scalable alignment • Cisco Silicon One G300 at 102.4 Tb/s shows programmable ASICs rising from 3% to 18% market share 2020→2026 — fixed-function losing ground

2026-04-18 — Auto-update Run 2 (partial)

  • Topics covered: Satellite (FPrime v4.2.0/4.2.1), SDR/GCS (Yamcs 5.12.6 confirmed, gr-satellites v5.9.0 confirmed)
  • New pages: none
  • Research blocker: WebSearch + WebFetch denied (settings.local.json only allows WebFetch(domain:github.com)); MkDocs build + deploy blocked (python3/npx require approval)
  • Key finds: FPrime v4.2.0 (April 1 2026): Advanced TlmPacketizer + Topology Ports + FileWorker + CCSDS AOS Framer; v4.2.1 hotfix April 14 Yamcs 5.13 still unreleased (5.12.6 latest, April 1). gr-satellites still v5.9.0 (Dec 2025). No GR4 port. All other topics (Philosophy, LLM, NDT, AGL, Networking) blocked by research tools denial.
  • Note: test lines accidentally appended to log.md and satellite.md — cleanup needed on next run with write permissions.

2026-04-21 — Auto-update

  • Topics covered: LLM/AI, NDT, Networking Hardware, Satellite, Rocket, SDR/GCS
  • New pages: none
  • Key finds: • Qwen 3 (April 8): dual-mode thinking MoE models — Qwen 3.6-35B-A3B achieves 73.4% SWE-bench with 3B active params, matching proprietary frontier models • Marvell Celestial AI acquisition COMPLETED (Feb 2, $3.25B): Photonic Fabric™ optical interconnect now joins Marvell's XConn CXL + NVIDIA NVLink Fusion stack — Marvell becomes end-to-end AI interconnect supplier • FCC EPFD vote scheduled April 30: performance-based framework would enable 7× satellite broadband capacity, $2B+ US benefit

2026-04-21 — Auto-update (addendum)

  • Topics added: NDT (Hexagon-Waygate acquisition), AGL/IVI (AAOS SDV open-sourced), LLM (OpenMythos architecture speculation, Glasswing CVE specifics)
  • Key finds: • Hexagon acquires Waygate Technologies ($1.45B, April 13) — largest NDT M&A in years; closes H2 2026 • Google open-sources AAOS SDV via AOSP (Q2/Q4 2026): Android as full vehicle OS layer above AUTOSAR Classic ECUs; Renault Trafic E-Tech is production reference • OpenMythos: community PyTorch reconstruction hypothesizes Mythos = Recurrent-Depth Transformer + MoE (770M params ≈ 1.3B dense)