{"count":3,"price_usd":"0.20","dossiers":[{"slug":"llm-architectures","topic":"llm-architectures","title":"LLM Architectures & Pretraining: State of the Field (October 2026)","summary":"How frontier language models are built and pretrained as of October 2026: sparse mixture-of-experts scaling, the competing full, sparse-compressed and hybrid linear attention designs, million-token context, optimizers, low-precision training, residual-stream redesigns, data and scaling laws. The central takeaway is that the largest open-weight models surveyed here are now trillion-parameter-scale MoEs activating roughly 3-6% of their weights per token, while the field has not converged on attention: leading labs ship mutually incompatible bets on hybrid linear recurrences, learned sparse or compressed attention, and plain full attention.","version":1,"published_at":"2026-10-29","updated_at":"2026-10-29","stats":{"words":5034,"citations":111,"claims":38,"timeline":20,"benchmarks":13,"open_problems":6,"key_papers":14,"glossary":16},"verification":{"status":"verified","verified_at":"2026-10-29","method":"evidence-pack check of 38 claims + 118 contexts against 111 cited records, plus 15 web checks on primary arXiv pages (13 returned content, 2 API requests were rate-limited)","claims_checked":38,"issues_fixed":54,"notes":["Qwen3.5: the Gated DeltaNet hybrid claim was cited to the Qwen3.5-Omni report, whose abstract only supports a hybrid-attention MoE; re-pointed to arxiv:2609.20751 (three GDN layers per attention layer, checked on the paper) and 'introduced' removed (exec summary, hybrids paragraph, c11).","Hybrids paragraph: 'in 2026 hybrids reached frontier scale' contradicted the dossier's own 2025 sources (MiniMax-01 and M1 are 456B hybrids); reworded.","MiniMax Sparse Attention: 'matched GQA quality' changed to 'on par, benchmark-dependent gaps'; 28.4x labelled a theoretical FLOPs figure; speedups tied to 1M context; link to MiniMax-M3 confirmed (body, c5, benchmarks).","Scope of 'largest open-weight' claims limited to the models surveyed (front matter, exec summary, benchmark metric, Kimi K3 key-paper note); loss-free balancing narrowed from 'several frontier MoEs' to DeepSeek-V3.","Nemotron-H: 'FP8 matched BF16 for the 56B' corrected to a recipe reported on par with BF16 and used for the 56B; c4 'replaces residual connections' changed to 'augments' (abstract: enhances); c22 'Most of the gap' changed to 'Much'.","MiniMax-M2: wording aligned with the report (earlier hybrids adequate on standard benchmarks, multi-hop deficits at scale, sliding-window variants worse beyond 32K after fine-tuning, linear/sparse infrastructure less mature); table row relabelled 'report 2026-05' because the cited record does not give the release date.","Checked against primary pages: DeepSeek-V4 submission date 2026-04-26, CSA m=4, HCA m'=128, 33T/32T tokens; Kimi K3 69 KDA + 24 MLA layers, NoPE in MLA, Muon, weights released; GLM-5 744B/40B, 28.5T tokens, MLA+DSA, Muon; Nemotron 3 Super latent 1024 vs 4096, 22 of 512 experts, BF16 layers, no positional embeddings; Mamba-3 +1.8 (0.6 base + 1.2 MIMO); channel selection 98.1%/17.8%; context compression 7.9x; Qwen3.8 Muon/batch-size claim.","Vendor or self-reported figures labelled (Gemini 1.5 retrieval, Mercury throughput, DeepSeek-V4 27%/10%, Kimi Linear, Nemotron-H, Kimi K2 loss spikes) and a disclosure line added under the benchmarks table; Mercury and dQwen3.5 caveats added.","Benchmarks JSON aligned with the body (token-count row now Qwen3 36T / V4-Pro 33T / GLM-5 28.5T, NVFP4 row includes Ultra, KV-cache row labelled versus DeepSeek 67B); timeline entry for Kimi Linear marked as reported.","Removed unsupported items: NVIDIA attribution of Gated DeltaNet-2, 'revised in 2024' for the data-exhaustion forecast, 'NVIDIA's' study label, absence claim about tokenization; outlook data bullet made conditional.","Softened overclaims: pure recurrent models 'still fail' at recall, 'widely used' (Llama 3, muP, schedules), 'lower-risk', 'increasingly drop positions', DeepSeekMoE 'spread/template', Muon 'evidence base', router-score generalization, 'undisclosed data', Lost-in-the-Middle dated 2023-2024.","Not re-fetched within the 15-fetch budget: LinMix 46% MBPP (hedged as 'up to', single preprint) and Attention Residuals 'every task' (reworded); details outside the pack's truncated abstracts and left as written: NVFP4 12B/10T, Qwen3 36T tokens, gpt-oss 128-token window, MiniMax-01 7:1 layer ratio, Kimi K2 MLA, Gemini 'sparse MoE' description."]},"ai_assisted":true,"price":{"usd":"0.20","protocol":"x402","network":"eip155:84532","asset":"USDC"},"url":"https://space-kitty.com/dossiers/llm-architectures","preview_url":"https://space-kitty.com/api/v1/dossiers/llm-architectures/preview","purchase_url":"https://space-kitty.com/api/v1/dossiers/llm-architectures"},{"slug":"quantum-error-correction","topic":"quantum-error-correction","title":"Quantum Error Correction & Fault Tolerance: State of the Field (October 2026)","summary":"An expert briefing on quantum error correction as of late October 2026: surface, color and qLDPC codes, bosonic and erasure qubits, real-time decoding, magic-state preparation and fault-tolerant resource estimates. The main takeaway is that below-threshold error suppression has been shown on superconducting hardware, in a short neutral-atom surface-code circuit and, in a 2026 preprint, in trapped-ion color-code memories, so the bottleneck has shifted to logical gates, magic-state cost, decoding latency, correlated errors and scale. Published estimates for factoring RSA-2048 have fallen from 20 million noisy qubits (2019) to under one million (2025); 2026 preprints claim under 100,000 with qLDPC codes, and Shor's algorithm at cryptographically relevant scale on as few as 10,000 reconfigurable atoms, with runtime falling as qubits are added, all assuming hardware not yet demonstrated at scale.","version":1,"published_at":"2026-10-29","updated_at":"2026-10-29","stats":{"words":6183,"citations":90,"claims":40,"timeline":20,"benchmarks":19,"open_problems":8,"key_papers":15,"glossary":19},"verification":{"status":"verified","method":"independent AI fact-check: evidence-pack check of 40 claims + 147 contexts against cited titles and abstracts, plus 15 web checks (arXiv abstract and HTML pages, Crossref and OpenAlex/PubMed records, IBM roadmap post, web searches for attributions); automated citation resolution","claims_checked":40,"issues_fixed":36,"verified_at":"2026-10-29","notes":["10,000-atom Shor figure (summary, executive summary, resource estimates, benchmark row, timeline): the abstract ties 10,000 qubits to no specific instance and gives a runtime only for P-256 at 26,000, so the RSA-2048 'floor' framing was replaced by 'as few as 10,000 for cryptographically relevant Shor instances'; the Pinnacle abstract states no runtime, now noted.","Summary no longer says neutral-atom logical memories are below threshold: the result is a four-round characterization circuit.","Decoding: non-real-time results removed from the 'real-time' bullet; the claim that the most accurate hardware decoders are neural was reduced to what the AlphaQubit abstract supports; Helios neural decoders are reported faster than syndrome extraction (GPU compute time, not feedback latency).","Gross-code figures (12 logical qubits in 288 qubits, nearly 3,000 for surface codes) are now attributed to the authors' simulations and estimate.","Helios color code: 5e-6 confirmed as per cycle at d = 7 (full text); 'steepest scaling' qualified to the experiments surveyed.","IBM roadmap statements (Starling 2029, 200 logical qubits, 10^8 gates; Kookaburra 2026; Cockatoo 2027) checked against IBM's 10 June 2025 post and labeled as company plans; no delivery claimed.","Unsourced positions ('advocates', 'others') in cryptanalytic timelines replaced by a neutral reading; Outlook and forecast claims c36/c37 hedged, c36 marked as unsourced editorial expectation.","'First hardware qLDPC memory', 'far from break-even' and 'today's 8.91%' replaced by 'early hardware memory'; c22 gained the 32-qubit demonstration as a citation.","JSON/body alignment: timeline 2024-08 (real-time decoding only at d = 5), 2025-06 ('key elements' of a universal architecture), 2026-03 (neutral-atom estimate and citation added); key_papers wording (sparse blossom, gross code, cultivation, 2207.06431); open_problems[2] (coherent errors are only simulated); c10 confidence lowered to medium.","Cat-qubit repetition code: the outer phase-flip code is below threshold, but total logical error improved only from 1.75% to 1.65%.","Zuchongzhi 3.2 figures (Λ = 1.40(6), 72-fold, 6.4(5) × 10^-4) confirmed from the PRL abstract record and PRL 135, 260601 from Crossref; processor name and qubit count from web search only. Journal status of 2512.13908 could not be confirmed, so the 'preprint' label was dropped; the 'first' claim for 2207.06431 is attributed to the authors via secondary summary.","Smaller wording fixes: universal-logic phrasing, atom-loss wording, proposed-versus-demonstrated qLDPC on atoms, post-selection caveats, assumed links, five-to-one scope, benchmark notes, universal negatives."]},"ai_assisted":true,"price":{"usd":"0.20","protocol":"x402","network":"eip155:84532","asset":"USDC"},"url":"https://space-kitty.com/dossiers/quantum-error-correction","preview_url":"https://space-kitty.com/api/v1/dossiers/quantum-error-correction/preview","purchase_url":"https://space-kitty.com/api/v1/dossiers/quantum-error-correction"},{"slug":"glp1-metabolic","topic":"glp1-metabolic","title":"GLP-1 & Incretin Therapies: State of the Field (October 2026)","summary":"Covers semaglutide, tirzepatide, the new oral GLP-1 agonists and next-generation triple-agonist and amylin-combination drugs, with their effects on weight, diabetes, cardiovascular, kidney, heart-failure, liver, sleep-apnea, joint and brain outcomes as of October 2026. The central takeaway: the field has moved from proving weight loss to proving organ protection. Placebo-adjusted mean weight loss at the top doses of pivotal trials in adults without diabetes now ranges from about 9 percentage points (oral orforglipron) to about 21 (retatrutide, not yet approved), and the main open questions are durability after stopping, lean-mass and rare safety signals, cost and access, and how much of the benefit is independent of weight loss.","version":1,"published_at":"2026-10-29","updated_at":"2026-10-29","stats":{"words":5487,"citations":100,"claims":36,"timeline":18,"benchmarks":12,"open_problems":8,"key_papers":14,"glossary":12},"verification":{"status":"verified","method":"Independent AI fact-check against cited sources: evidence-pack check of 36 claims + 124 contexts (including 18 timeline and 12 benchmark rows), 17 web checks (batch PubMed abstracts for all 89 DOI records, ClinicalTrials.gov records for 9 trials, one PMC record, agency/company announcements and trade press for regulatory and topline claims); automated citation resolution","claims_checked":36,"contexts_checked":124,"web_checks":17,"issues_fixed":31,"verified_at":"2026-10-29","notes":["Thyroid cancer: replaced \"inconclusive because events are rare\" with the 2026 review's mixed evidence (rodent-based boxed warning; three trial meta-analyses reporting an increase); pancreatic-cancer statement kept on the 2021 safety review.","SURPASS-CVOT: removed the unverifiable claim that all-cause death sat outside multiplicity-controlled testing; added the verified HR 0.84 (95% CI 0.75-0.94) with a caution.","Executive summary: SUMMIT benefit attributed to worsening-heart-failure events; cross-trial efficacy headline softened; STEP 1 extension labelled exploratory (n=327); SURMOUNT-OSA population and REDEFINE 4 estimand specified; GI adverse-event context added.","Citations re-pointed: Nature Reviews Drug Discovery review moved from a mouse-circuit claim to the lean-mass discussion; energy-intake claim now cites the two physiology reviews; OASIS 1 and OASIS 4 added for the 25-50 mg oral doses.","Company and regulatory events (FDA approvals of oral semaglutide 25 mg in Dec 2025 and orforglipron in Apr 2026, EMA NAION conclusion in Jun 2025, CMS BALANCE/Bridge changes, Nov 2025 USD 245 Medicare price, generic launches, REDEFINE 4 and TRIUMPH toplines, Q1 2027 retatrutide filing plan) were checked against agency/company announcements and trade press and stay attributed in the text, since no DOI/NCT/PMID exists for them.","Removed or softened unsupported details: CMS date (April to spring 2026, sources disagree), evoke topline date, China data-protection rationale, AMPLIFY-1 start (not yet recruiting), older-adult and heart-failure-subtype statements not in the cited abstracts, \"about four years\" follow-up.","Consistency and safety context: SUMMIT and SURMOUNT-1 3-year benchmark years changed from 2025 to 2024 to match records and timeline; estimand-gap example corrected to 2.8 and 3.3 points; comparator or severity context added to cholelithiasis, GI, gastric-content, dysesthesia and EMA NAION statements; type 1 diabetes sentence reworded so it does not read as dosing advice.","No new citation ids added; all 100 cited keys resolve to evidence-pack records. The dossier has no in-file disclaimer; the site-level not-medical-advice notice applies."]},"ai_assisted":true,"price":{"usd":"0.20","protocol":"x402","network":"eip155:84532","asset":"USDC"},"url":"https://space-kitty.com/dossiers/glp1-metabolic","preview_url":"https://space-kitty.com/api/v1/dossiers/glp1-metabolic/preview","purchase_url":"https://space-kitty.com/api/v1/dossiers/glp1-metabolic"}]}