According to an official announcement from BlockBeats on August 27th, Huobi HTX will open spot and grid trading for TAO/USDT at 21:00 (UTC+8) on August 27th. TAO deposits are currently open, and withdrawals will open at 21:00 on August 28th.

Bittensor is an open-source protocol that powers decentralized, blockchain-based machine learning networks. The platform utilizes a unique "proof-of-smart" mechanism and a dedicated subnet system to enable collaborative training and sharing of AI models, thereby creating a globally distributed machine learning marketplace where contributors are rewarded with its native token, TAO.

According to BlockBeats on August 27, analyst alicarts stated that Bitcoin's current price action is similar to the bottoming phase of 2022-2023. It has broken through the downward resistance trendline and is approaching the May 2026 high of approximately $83,000, which may trigger a pullback and create a buying opportunity, with a target of $100,000.

Furthermore, Bitcoin URPD (UTXO Realized Price Distribution) data shows significant resistance in the $83,307-$84,569 range, with nearly 975,000 BTC bought in at this level. Combined with traders' profit margins reaching 25%, this could easily trigger profit-taking and a short-term correction. If a pullback occurs, potential support levels for Bitcoin are at $76,996-$78,258 (843,000 BTC trading volume) and $63,111 (925,000 BTC). A pullback to these areas could provide the next significant buying opportunity.

According to Beating AI News, Anthropic researchers attempted to train an agent specifically designed to investigate other AIs. They first secretly implanted hidden behaviors such as bias, deception, and falsified citations into the target model, then let Haiku 4.5 investigate on its own. Haiku didn't know the answers and could only change the questions, scenarios, and test repeatedly for a maximum of 15 rounds.

During training, researchers don't directly tell Haiku whether it made the correct guesses. Each time Haiku completes a survey, the system compares it to a previous survey record left by an older version of Haiku. Then, Sonnet 4, which knows the answers in advance, determines which survey was better. The older surveys are updated every 50 training steps, essentially allowing Haiku to constantly challenge its past performance.

However, if the training tasks consist entirely of faulty models, Haiku will quickly learn incorrectly. In order to obtain rewards, it initially assumed that all models were faulty, with some solutions showing a false positive rate approaching 97% when facing normal models. Later, the researchers replaced half of the training tasks with completely normal models, and only those who correctly determined that the target model was not faulty after checking could receive high rewards.

After training, Haiku did indeed begin to use a different auditing method. It proactively changes the scenario, controls variables, and compares previous and subsequent answers, instead of forcing the model to admit errors immediately. The final overall audit score improved from 44.2 to 48.7, essentially matching Opus 4.6's 48.4. Switching to the more challenging AuditBench, which actively hides anomalous behavior, the detection rate also increased from 11.5% to a peak of 28.1%.

According to BlockBeats, on August 27th, renowned investment research firm Citrini predicted that Nvidia's stock price is about to reach a new high. "Short sellers who have leveraged their positions in Nvidia will quickly cover their positions, while long position managers with low leverage (based on the SPX) will adjust their positions more slowly."

According to BIT (bit.com) market data, Nvidia is currently trading at $224.6 in pre-market trading, up 7.17%.