ETAM DFDC DTDU 🥺🥺

いつもスマホで推しのあの娘を応援していて、会いに行きたいけど、撮影会に行くにはカメラが無いしなぁ…まだスマホの分割払い終わってないんですけど！ そんな貴方に朗報です！シフォン撮影会は創業以来スマホ撮影大歓迎です！さぁスマホを持って推しに行こう！ chiffon.photo/ dfdc

“Making Delhi dust-free is our environmental duty.” Read this insightful piece by Union Minister Shri @byadavbjp ji on the launch of the ‘Dust Free Delhi Campaign’ (DFDC). The initiative aims to plant over 28 lakh trees and restore the city's “green lungs” the Delhi Ridge–to combat air pollution and urban heat. Collective action is key to a greener, climate-resilient future!

Congrats to the DFDC Class of 2026 on graduation night! Charlie Marks - Salutatorian🏅 Jay Noren - Salutatorian🏅 Hugo Albrecht-Buehler Sam Chaban Jamie Chandler Aurin Dasgupta Seth Frank Reed Galasinski Reid Gandy Scott Harvey Will Lambert Noelan Tossing

dfdC rH twU lMK wmr FddR

Reply from @solsticefi intern gives some extra boost to the community. For real, I don’t know who that person is but pretty sure this 1% allocation should go to this person🤌🙂‍↔️🙂‍↔️. But that aside, so in the last 24 hours, another partnership landed with DeFi Development Corporation. DFDC is a company that just because one of the first Nasdaq listed company to have treasury capital through solstice yield vault. Some people are clearly not seeing beyond the millions on FDV but I can surely tell you that right now solstice wants trillions in that TVL so as to be a leading example in Solana. I also saw the intern reply one post with ‘this is really just the start… flare szn incoming 💯’ could this be a message that we should be prepared for the race ahead?👀 We shall see but right now, all I can bull post about is solstice and I am here for it.

🚨 FREE BTC (SIGNAL) TRADING SIGNALS! 🚨 t.me/ uXJTwRMcgTg4Zjk8 🔥 95% Accuracy Rate 📈 3–6 Profitable Signals Every Day #BTCUSD #Crypto #Bitcoin #BTC💯 #btc #btcsignals #dfdc

re beg g’s zcvfxfzzsxxcsaxxq dfdc -samoa

“quienes creen que el precio por levantar la voz es muy alto, espérense a ver el que pagaremos si nos quedamos callados” DFDC

#Bittensor $TAO #decentralizedAI Claims and Cross Check. 1. SN2 – @omron_ai — “World’s best SSL for AI.” What it is: A zk‑ML / proof‑of‑inference subnet (cryptographic receipts that a specified model really ran on given inputs), not “self‑supervised learning (SSL).” Docs, code and diagrams confirm a ZK proof flow between miners and validators. Evidence for “world’s best”: There’s no neutral benchmark comparing Omron’s prover/verifier throughput, latency or costs vs leading zk‑ML stacks (EZKL, RISC Zero, Modulus, Orion). Third‑party zk‑ML benchmarks show massive variance across frameworks, but don’t include Omron yet. What to publish: End‑to‑end proof time, gas/CPU cost, and verification time on standard nets (ResNet‑18, tiny‑BERT) vs EZKL/RISC‑Zero; reproducible harness traces. Also clarify model‑authenticity guarantees (proof binds to model hash?) and adversarial cases (quantization drift, non‑determinism). 2. SN3 – @tplr_ai — “World’s best distributed decentralized pre‑training.” What it is: A distributed training subnet that already finished a ~1.2B‑parameter run over 20k cycles with ~200 GPUs; roadmap to 8B→70B. Preliminary benchmark deltas vs AdamW are published. Evidence for “world’s best”: We have proof it trains at scale; we don’t have SOTA wins against centralized runs or public loss‑curves at 8B/70B. What to publish: Full training cards (data tokens, optimizer, LR schedule, wall‑clock, hardware mix), eval on HELM/MMLU/ARC‑C/HellaSwag for Templar‑8B; scaling laws vs baselines; reproducible shards checkpoints. 3. SN4, SN51, SN64 – @TargonCompute / @lium_io / @chutes_ai — “World’s cheapest compute / inference / AI deployment.” What they are: The compute trifecta (serverless inference & rentable GPU capacity). Code, docs, and public orgs exist; Targon emphasizes secure/CC execution; Lium exposes a renter UI; Chutes ships API keys on BitTensor auth. Investor analysis claims $20M ARR combined after monetization switch‑on. What to publish: A public price/perf dashboard: $/1M tokens for Llama‑3.1‑8B/70B at fixed latency SLOs, $/GPU‑hour by SKU (A100/H100/L40S), cold‑start p95, and egress. Put it side‑by‑side with Vertex/SageMaker quotes and Runpod on a fixed workload. 4. SN13 – @Data_SN13 Universe (Macrocosmos) — “World’s largest decentralized data scraper.” What it is: A subnet focused on web/social scraping & distribution (e.g., Gravity/MCP), positioned as Bittensor’s “data layer.” Evidence for “largest”: Strong outward momentum, but no audited totals (docs/pages/day, unique domains, total tokens) vs decentralized peers (Common Crawl‑style mirrors, on‑chain data nets). What to publish: Live counters (unique URLs/day, avg delay from publish→ingest, cost/post), provenance metrics, and dedup/robots compliance reports. 5. SN17 – 404‑Gen — “World’s largest collection and generator of 3D models.” What it is: A 3D generation subnet that released a 21.5M‑asset open dataset; a 20k‑asset slice is on Hugging Face; claim is it exceeds Objaverse‑XL (10M ). Evidence for “largest”: Plausible. Objaverse‑XL is ~10M; 404‑Gen reports 21.5M. Multiple third‑party write‑ups repeat the size claim, though peer‑reviewed auditing is thin. What to publish: Independent audit of unique meshes (hash‑based dedup), license metadata coverage, basic quality metrics (mesh watertightness, triangle count distributions) and alignment with Unity/Blender import success rates. 6. SN18 – @zeussubnet & SN57 – Gaia — “World’s most accurate weather forecasting.” What they are: Decentralized ML weather/geophysical subnets. Zeus has analyses on 2‑m temperature & precipitation using ERA5 and mixture‑of‑experts; Gaia targets geomagnetic Dst and soil moisture (Sentinel‑2/SMAP inputs). Evidence for “most accurate”: Internal PDFs show encouraging results, but the global bar is ECMWF/HRES (and now AIFS). There’s no head‑to‑head, station‑verified leaderboard (RMSE/CRPS) across horizons and regions. What to publish: A public forecast‑verification portal vs ECMWF/GFS/ICON with RMSE, MAE, ACC, CRPS by lead time (0–120h), by variable (2‑m T, 10‑m wind, precip), and by WMO station set—plus PIT calibration plots. For Gaia’s Dst/soil moisture: agreed scoring windows and comparisons to published baselines. 6. SN32 – @ai_detection — “World’s best AI text detector.” What it is: An AI‑generated text detector with published benchmarks (RAID, GRID, CUDRT). Their March 2025 report shows SOTA‑level RAID numbers at 5% FPR. Evidence for “best”: Contested but strong. RAID is the right benchmark; however, the live RAID leaderboard and other detectors (e.g., Desklib) also claim top spots at times. The task is adversarial and volatile; external media warn detection is brittle in the wild. What to publish: A live, signed submission to RAID/CUDRT with versioned model hashes; false‑positive studies on student essays & bilingual corpora; robustness to paraphrase/homoglyph attacks; and calibrated thresholds per use case (education, journalism, moderation). 7. SN33 – @ReadyAI_ — “World’s best decentralized text data cleaning/labeling.” What it is: A structured‑data/annotation subnet using LLMs to replace parts of human labeling pipelines. Public material compares itself conceptually to Scale AI. Evidence for “best”: No audited cost/quality comparisons vs commercial LLM‑labeling stacks. What to publish: Benchmarked quality (Krippendorff’s α, task F1) and cost/1k labels across NER, sentiment, multi‑label topics—against Scale AI/Labelbox/Mechanical Turk with blinded human adjudication. 8. SN34 – @bitmind — “World’s best deepfake detection.” What it is: An adversarial detector–generator subnet (GAS) for deepfakes. Active repo and explainer content exist. Evidence for “best”: Needs standardized evaluation on FaceForensics , DFDC, DeeperForensics with cross‑model generalization and adversarial evasion. What to publish: ROC‑AUC/EER by generator family (latent diffusion, face‑swap), robustness to compression & frame rate, and live red‑team reports. 9. SN44 — “World’s best decentralized computer vision.” What it is: A decentralized video CV stack (starting with soccer), with live tasks and a public site/repo. Evidence for “best”: Great to see leaderboards, but no COCO/AVA/TrackingNet style numbers; sports‑specific tasks are hard to compare. What to publish: mAP for detection/pose, MOTA/MOTP for tracking, event detection F1, and cost/video‑minute vs centralized pipelines. 10. SN50 – @SynthdataCo — “World’s best price path projection forecasting.” What it is: Probabilistic crypto price‑path simulation (distributions, not point forecasts) with Monte‑Carlo style ensembles and open whitepaper/code. Evidence for “best”: promising framing. The right way to judge this is CRPS, calibration (PIT), and tail risk capture, not directional accuracy. Community analyses praise the methodology, but we still need neutral backtests. What to publish: Rolling CRPS vs implied vols and naive baselines; calibration curves; PnL of risk‑managed strategies that only read the distribution (no oracle peeking). 11. SN56 – @gradients_ai — “World’s cheapest 1‑click AutoML.” What it is: A competitive fine‑tuning platform (text/image; Instruct/DPO/GRPO) with v1.0 and claims of $100–$500 runs vs “$10k ” on big‑cloud. Evidence for “cheapest”: Cloud prices vary; public docs don’t show matched case studies vs Vertex/SageMaker for the same dataset/model/SLO. What to publish: A basket of open fine‑tune jobs (e.g., Llama‑3‑8B FAQ bot, 5–20k examples): quality deltas, wall‑clock, and fully loaded cost vs Vertex/SageMaker open infra. 12. SN62 – @ridges_ai — “World’s best open‑source coding assistant.” What it is: A code‑agent subnet targeting CI regression/code‑gen; roadmap promises SWE‑Bench style benchmarks and a public API/leaderboard. Evidence for “best”: Today’s gold standards are SWE‑bench Verified and now SWE‑bench Pro; we need apples‑to‑apples scores. What to publish: End‑to‑end SWE‑bench Verified results (solve‑rate, time, human‑intervention), patch acceptance in real repos, and diffs signed by miners. 13. SN68 – @metanova_labs — “World’s best novel drug discovery AI.” What it is: A drug‑discovery subnet with evolving NOVA protocol papers (adversarial, model‑agnostic search). Evidence for “best”: Research‑stage. No prospective wet‑lab validations or blinded external challenge wins are public. What to publish: Prospective hit‑finding on public targets (DUD‑E/LIT‑PCBA) with docking selectivity; ADMET predictions vs pharma baselines; even a small wet‑lab validation would be huge. 14. SN75 – @hippius_subnet — “World’s cheapest cloud storage.” What it is: A storage subnet launched in March 2025 with community hub and docs. Evidence for “cheapest”: Needs $/GB‑month, durability, egress, and retrieval latency compared to S3, Cloudflare R2, Storj, Filecoin. What to publish: A pricing & durability SLA: $/GB‑mo, 11‑nines target, multi‑region replication, and public chaos tests (corruption, partial loss). 15. SN85 – @vidaio_ — “World’s cheapest video compression & upscaling AI.” What it is: A video processing subnet with an active site and repo/media explaining upscaling/compression services. Evidence for “cheapest”: We need VMAF/SSIM/PSNR vs Topaz/DaVinci cost/minute at fixed output specs (1080p→4k, p95 latency). What to publish: Full benchmark suite: input qualities, GPU class, throughput, and cost curves; ablation on temporal consistency. 16. SN93 – @Bitcast_network — “World’s largest decentralized video generation agency.” What it is: A UGC mining subnet where creators publish to briefs (YouTube, X) and validators score via OAuth’d analytics; on‑chain economics are tracked (dTAO/Alpha). Evidence for “largest”: Architecture and token flows are clear; “largest agency” needs counts of active creators, videos/month, and aggregated watch‑time vs any other decentralized UGC marketplace. What to publish: Transparency board: active miners, briefs filled, views/watch‑hours verified per cycle, payout dispersion, and anti‑sybil/traffic‑quality stats.

DFDCは設計ツールではなく、解析ツール。設計法はワスしか発表していない。DFDCではダクテッドファンのブレードの「妥当な解」として付け根から先端まで一定の循環を与え、これを実現するブレードを出力する。ワスの方法はこの循環分布を最適化できる。

MITのダクテッドファン設計ツールのDFDCは決定的に間違っている。翼端渦が壁面を移動するときは外部流の丁度半分の速度になる。DFDCでは外部流の速度で移動する。これの効果が大きくて計算結果は信じられない。権威があるから有名だからとなんでも信じる癖はやめよう。

DFDC frequently can be expanded as De-dedicated FDC.

Para los “Ingenieros expertos” en sistemas. Modelo CNN-LSTM para detección de deepfakes de audio en AWS SageMaker, usando instancias EC2 g5.xlarge (NVIDIA A10G). Datasets: ASVspoof 2019, ADD 2021, Facebook AI DFDC.

Hola Gabriel Te felicito por tu análisis. Yo sé que no es fácil aplicar este análisis, porque siempre habrá algún cuestionamiento. Me gustaría hacer una observación: Un modelo CNN-LSTM entrenado con el conjunto de datos del Deepfake Detection Challenge no sería lo mejor opción. El dataset del DFDC fue recopilado entre 2018 y 2019, se enfoca en manipulaciones visuales y contiene en su mayoría habla en inglés. Como el conjunto fue creado antes de los sistemas modernos (2025) de síntesis de voz, podría un modelo basado en este conjunto reconocer los sutiles artefactos producidos por las nuevas técnicas de generación de voz con IA? Generalizaría adecuadamente entre idiomas? Saludos!

Fuentes: Facebook AI, AWS, Microsoft & Partnership on AI. (2020). DeepFake Detection Challenge (DFDC) Wu, Z., et al. (2015). ASVspoof 2015: The First Automatic Speaker Yi, J., et al. (2021). ADD 2021: Audio Deepfake Detection Challenge. Villalba, J., & Dehak, N. (2020).

Para lograr la máxima precisión, mi modelo de detección de anomalías se entrenó rigurosamente en el DeepFake Detection Challenge (DFDC) dataset, un conjunto masivo de videos y audios falsos y reales proporcionado por Meta (Facebook AI).

Kidslympic bakal bentuk skuad bawah 4 tahun hasil kolaborasi bersama DFDC! ⚽🔥💪🏼 #NadiWeekend #AstroArena

Una mascota muy perversa , DFdC . Saqueador del México al puro estilo PRIAN .

I am Emmet. You are cool //THANKSSS du fhdfjh Hu dfdc jh