ONEЦена Harmony

U.S. spot Bitcoin and Ethereum ETFs saw a net inflow of $2.6 billion for the week, marking the highest since October 2025, according to SoSoValue. Bitcoin ETFs attracted $1.9 billion while Ethereum funds drew $697.2 million, both setting 2026 records. This reversed a $392 million outflow from the prior week. Bitcoin ETF trading volume surged over 219% to $22.1 billion, and assets under management (AUM) grew 25.4% to $96.1 billion. Inflows were concentrated around a rally on August 19-20, with BlackRock's IBIT fund alone accounting for $503 million of the $606.3 million inflow on Thursday. Ethereum ETFs saw their strongest week since October 2025, with trading volume jumping 259.4% to $6.9 billion and AUM rising 35.9% to $14.3 billion. Despite the strong week, Bitcoin ETFs still show a net outflow of approximately $2.9 billion year-to-date for 2026, and Ethereum funds are down about $191.8 million. The recent inflows reduced the combined year-to-date deficit for both categories from $5.7 billion to $3.1 billion.

The most prominent Chinese crypto miner Jiang Zhuo'er, founder of BTC.TOP, has revised his earlier bearish outlook, now expressing 90% confidence that the bear market is over. He predicts Ethereum (ETH) may outperform Bitcoin (BTC) in the current cycle. Jiang detailed his recent trading activity, repurchasing ETH between $1,738 and $1,931 and setting a stop-loss at $2,100, then selling 50% of his spot ETH at $2,525 with a stop-loss at $2,550. He plans to allocate 20-30% of his remaining capital to buy ETH, using the rest if Bitcoin drops to the $67,000-$72,000 range. Otherwise, he intends to purchase ETH at market price by the end of October. Jiang notes the market's strong upward trend makes precise peak prediction difficult.

The Grayscale Zcash Trust has filed its fifth amended registration statement with the SEC, aiming to convert into the first U.S. spot ETF directly holding a privacy coin (ticker ZCSH). Key terms include a 2.5% annual fee, Bank of NY Mellon as administrator, and Coinbase for custody and brokerage. The trust held ~388,700 ZEC (~$155M) as of June 30; its AUM surged to ~$260M by August 21, driven by ZEC's price rally from $250 to over $800. The ETF represents a major identity shift for Zcash, a coin previously delisted by many exchanges and flagged by regulators for its anonymity. However, the ETF strips away Zcash's core privacy feature. To ensure compliance, all trust-held ZEC will be stored in transparent addresses, making transactions auditable and traceable. Zcash's optional privacy—unlike Monero's mandatory privacy—is the key compromise enabling this Wall Street entry. Market pricing reflects aggressive optimism for institutional access, but challenges remain. Only ~22% of circulating ZEC is in shielded pools; if ETF demand siphons transparent coins, Zcash's privacy narrative weakens. Competition from programmable privacy on Ethereum L2s could also dilute its value proposition. With prices already up over 200% in four months, the rally may have prematurely priced in the ETF approval, leaving limited safety margin. Grayscale's move highlights traditional finance's capacity to repackage regulatory-edge assets into compliant products, raising questions about the future boundaries of crypto ETF adoption.

Altcoins are surging, with the "Altcoin Season Index" rising from 25 in May to 43. Tokens like PUMP and ENA have seen weekly gains of nearly 100%, while Zcash has hit a new all-time high. However, most major altcoins remain significantly below their previous peaks, with some still down over 90% from their highs. This selective rally contrasts with Bitcoin trading near $77,400, close to its record high from October 2025. Despite some traders on social media declaring "Altseason 3.0" is here, Google Trends data shows global search interest for "altcoin" remains low at 26 out of 100, far from the peak levels seen before Bitcoin's last ATH. The index is still well below the 75 threshold needed to officially confirm an altcoin season, suggesting current excitement may be premature and the broader market mania has not yet materialized.

NVIDIA's general-purpose coding agent AVO has achieved a perfect 100.00 RHAE score on the ARC-AGI-3 benchmark. It solved all 183 levels across 25 game environments in just 6,624 steps. ARC-AGI-3 is a notoriously difficult test where agents are placed into unfamiliar games with no instructions, forcing them to deduce rules and goals through trial and observation. Top models like Claude Opus 5 typically score only around 30% when acting alone. AVO's breakthrough comes not from improving the underlying AI model (Claude Opus 5), but from adding an intelligent external framework or "harness" around it. This framework addresses three key failure modes: misunderstanding global rules, misapplying familiar mechanics, and failing to learn correctly from success. Two core mechanisms drive AVO's performance: 1. **Persistent Memory:** It stores past attempts, compiler outputs, and reasoning across tasks, preventing the model from resetting and re-exploring dead ends when its context window fills. 2. **Supervisor Agent:** A separate module monitors the main agent's progress, intervening to redirect strategy when it gets stuck or repeats unproductive actions. Notably, AVO processed the visual game environments using a pure 64x64 text grid representation, without any image tokens. Originally developed for GPU kernel optimization, AVO autonomously evolved CUDA code for 7 days, producing attention kernels that outperformed NVIDIA's own cuDNN and the leading open-source implementation, FlashAttention-4. The same core architecture was then successfully applied to ARC-AGI-3, demonstrating its generalizability for long-horizon, autonomous problem-solving. The project features a predominantly Chinese team, including Distinguished Engineer Bing Xu (co-author of the original GAN paper and creator of MXNet) and Tianqi Chen (creator of TVM and XGBoost). NVIDIA's strategy highlights that while AI models are crucial, the surrounding system—managing memory, tools, and feedback—is equally vital for creating capable, long-term autonomous agents. This approach aligns with NVIDIA's focus on providing the essential hardware and software infrastructure for the next generation of AI workloads.