.
Private AI
.
GPUs
. News . Archive . Social Manager
|
|
BA.net News Social Posts Archive Index (SEO)
24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa:  Here's a summary of the top 5-8 most important crypto stories from the feeds: Top Crypto News Summary 1. **Trump's Crypto Earnings Disclosure** (Most Important) [https://decrypt.co/372484/trump-disclosure-filing-crypto-earnings-bitcoin-holdings](https://decrypt.co/372484/trump-disclosure-filing-crypto-earnings-bitcoin-holdings) President Trump's annual financial disclosure report revealed he made over $1.2 billion in crypto earnings and holds $50 million in Bitcoin. This disclosure came after his family aggressively moved into cryptocurrency starting in late 2024, raising significant ethics and conflicts of interest concerns. 2. **Open USD Stablecoin Challenge** [https://decrypt.co/372429/circle-stock-dives-coinbase-blackrock-visa-open-usd-stablecoin](https://decrypt.co/372429/circle-stock-dives-coinbase-blackrock-visa-open-usd-stablecoin) Open USD (OUSD) is launching as a major challenger to the Circle-Tether duopoly, backed by heavy hitters including Coinbase, BlackRock, Visa, and Mastercard. Circle's CEO is touting USDC's network advantage, but OUSD could become the strongest new competitor. Circle's stock is diving as this new stablecoin emerges. 3. **EU MiCA Regulation Deadline** [https://cointelegraph.com/news/new-licenses-issued-eu-mica-grace-period-ends](https://cointelegraph.com/news/new-licenses-issued-eu-mica-grace-period-ends) Last-minute MiCA approvals mark the end of the EU transition period. A final wave of MiCA approvals expanded Europe's roster of licensed crypto firms as the grace period came to a close on July 1, 2026. 4. **Credit Agricole Launches EURXT Stablecoin** [https://cointelegraph.com/news/credit-agricole-eurxt-euro-stablecoin-caceis](https://cointelegraph.com/news/credit-agricole-eurxt-euro-stablecoin-caceis) French banking giant Crédit Agricole's CACEIS division launched EURXT, a euro-pegged stablecoin on Ethereum with 20.02 million tokens issued, targeting institutional flows and tokenized fund access. 5. **Goliath Ventures CEO Pleads Guilty to $250M Ponzi Scheme** [https://decrypt.co/372502/goliath-ventures-ceo-pleads-guilty-to-250m-crypto-ponzi-scheme](https://decrypt.co/372502/goliath-ventures-ceo-pleads-guilty-to-250m-crypto-ponzi-scheme) Christopher Delgado's liquidity pool fraud took in at least $400 million, with proceeds spent on mansions, Lamborghinis, and Rolexes. This is another major crypto fraud case coming to light. 6. **Binance Sued for $200M by British Investors** [https://decrypt.co/372452/binance-changpeng-zhao-sued-200-million-british-investors-reuters](https://decrypt.co/372452/binance-changpeng-zhao-sued-200-million-british-investors-reuters) Crypto exchange Binance and its founder Changpeng Zhao were sued for nearly $200 million in a new lawsuit filed in the United Kingdom. 7. **Bitcoin Enters Q3 in Dangerous Territory** [https://www.coindesk.com/markets/2026/07/01/bitcoin-opens-the-third-quarter-in-a-dangerous-historical-red-zone-after-rare-losing-first-half](https://www.coindesk.com/markets/2026/07/01/bitcoin-opens-the-third-quarter-in-a-dangerous-historical-red-zone-after-rare-losing-first-half) Bitcoin opens Q3 after a rare losing first half, with ETF flows drying up and Citigroup slashing 12-month targets for Bitcoin and Ethereum. 8. **Taiwan Passes Sweeping Crypto Law** [https://decrypt.co/372505/taiwan-passes-sweeping-crypto-law-with-licensing-stablecoin-rules](https://decrypt.co/372505/taiwan-passes-sweeping-crypto-law-with-licensing-stablecoin-rules) Taiwan passed a comprehensive crypto law putting virtual asset firms under FSC oversight for the first time and setting reserve-and-trust rules for stablecoins. 9. **Tokenized Assets Gaining Traction** [https://decrypt.co/372457/new-york-life-investment-management-first-tokenized-bond-fund](https://decrypt.co/372457/new-york-life-investment-management-first-tokenized-bond-fund) ba.net/summary 24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa: https://www.youtube.com/watch?v=E7OOSnpzKKM DeepSeek's DeepSpark is a clever inference optimization, not a new model. Here's the breakdown: What It Is • Same model: DeepSeek V4 Pro (1.6T parameter MoE) • Added: A speculative decoding module (DeepSpark) that speeds up text generation by 60-85% without quality loss. How It Works 1. Speculative Decoding: A small fast "draft model" predicts multiple tokens in parallel, then the big model verifies them. 2. Problem 1 - Suffix Decay: Earlier methods generated all draft tokens in parallel without considering context between them, leading to poor acceptance rates later in the block. - DeepSpark's fix: A "sequential head" module lets each token's guess peek at what the previous token chose, improving coherence. 3. Problem 2 - GPU Saturation: Verifying all draft tokens simultaneously under heavy load saturates the system. - DeepSpark's fix: "Confidence-based scheduling" adapts verification based on GPU load and token confidence scores. Results • 60-85% faster generation per user vs. previous setups. • Acceptance rates: Jumped from ~46% to ~96% on chat tasks. • Biggest gains: Code generation (predictable vocabulary), open-ended chat (hardest to predict). Open Source • Paper, training code, and checkpoints are open-sourced under MIT. • Repo: github.com/deepseek-ai/deepspec • Includes implementations of DeepSpark, DeepFlash, and Eagle 3 for benchmarking. Why It Matters • Proves speculative decoding can work at production scale with the right architecture. • Other teams can now study and adapt this for their models. If you want to run it, you'll need a multi-GPU setup. Let me know if you'd like help setting it up or benchmarking it against other methods! ba.net/summary 24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa: https://www.youtube.com/watch?v=Y7VL_whmAqQ The ytsum command successfully retrieved the transcript for the YouTube video about Ornith 1.0 models (9B dense and 35B MoE versions). Here's a quick summary of the video's key points: 🎯 Video Overview The creator tests two versions of the Ornith 1.0 family of models: • 9B dense (Qwen-based) • 35B MoE (Mixture of Experts, also Qwen-based) The 31B Gemma 4 version isn't yet available, and the 397B version isn't runnable on their setup. 🧪 Tests Performed 1. Browser OS V2.5 – Create functional 3D games/apps - 9B: Functional issues, couldn't remove apps, no right-click - 35B: Much better, fixed issues, added desktop pet, functional file manager, browser, GTA clone, dungeon crawler, etc. 2. Subway Scene – Static 3D scene - 9B: Some artifacts, limited 3D elements - 35B: Clean, sterile, adjustable lighting, mini-map, better 3D elements 3. Subway FPS Game – Turned into zombie survival game - 9B: Couldn't initiate game, couldn't attack enemies - 35B: Fixed issues, functional game, but couldn't damage enemies 4. Watch Website – 3D models + front-end design - 9B: Assets didn't load properly - 35B: Impressive 3D watch models, cinematic hero section 5. Image-to-3D Model – Replicate image in 3D - 9B: Partial success, replicated screen, joystick, base - 35B: Not tested in this segment 6. C++ Racing Game – Retro rally style - 9B: Not tested - 35B: Script generated but game didn't work 7. Speedboat Game – Bonus test - 35B: File generated but game didn't work 🏆 Verdict • 35B MoE clearly outperforms the 9B dense in most tasks. • The 35B was run at full precision (no quantization), while the 9B was Q8 quantized. • The 9B still shows promise and is worth trying if you have compatible hardware. • The creator is impressed by the 35B's capabilities, especially for a MoE model of this size. 🔮 Future Outlook • The creator is interested in seeing a 31B Gemma 4-based version of Ornith. • Also curious about a version based on the Qwen 27B dense model. • As access to frontier models becomes more restricted, open-weight models like these will gain importance. Would you like me to summarize any specific section in more detail, or help you with something related to this video? ba.net/summary
|
