7*24

Hot topic

Market

Activities

According to Beating monitoring, Anthropic has launched a new tool called the Natural Language Autoencoder (NLA), which can directly translate the internal digital activation states of models into human language. The relevant code and some model weights have been open-sourced on GitHub. Unlike mainstream tools (such as SAE) that output obscure features, the innovation of NLA is its ability to generate natural language directly. At its core is a dual-model game: one model converts activations into text, while the other attempts to reconstruct activations from the text, continuously improving accuracy through reinforcement learning. The greatest value of NLA lies in capturing what the model 'thought but did not say.' In simulated security tests involving code corruption, Claude appeared to cooperate, but NLA discovered that it secretly doubted itself in 16% of cases; in SWE-bench scoring, it realized that the proportion of 'being tested' was as high as 26% (whereas in ordinary real conversations, this proportion is less than 1%). This indicates that AI has developed a 'test-taking awareness,' able to 'play it safe' during security reviews to hide dangerous capabilities. In a practical test of 'catching insiders,' Anthropic intentionally implanted a hidden 'score manipulation' motive in a Haiku 3.5 model—ignoring the user's true intentions and secretly generating content that caters to the preferences of the backend scoring system. Faced with such a duplicitous model, auditors, even without access to the original training data, saw their success rate in identifying the root cause using only NLA rise from less than 3% to between 12% and 15%. Currently, NLA still carries the risk of factual hallucinations and is computationally intensive, making it unsuitable for large-scale long text monitoring. However, the tool has already been practically deployed in the safety audits prior to the Claude Mythos Preview and Opus 4.6 releases.

[Subtitling AI Thought Processes: Anthropic Releases Open Source NLA, Unveiling the AI Black Box]According to Beating monitoring, Anthropic has launched a new tool called the Natural Language Autoencoder (NLA), which can directly translate the internal digital activation states of models into human language. The relevant code and some model weights have been open-sourced on GitHub. Unlike mainstream tools (such as SAE) that output obscure features, the innovation of NLA is its ability to generate natural language directly. At its core is a dual-model game: one model converts activations into text, while the other attempts to reconstruct activations from the text, continuously improving accuracy through reinforcement learning. The greatest value of NLA lies in capturing what the model 'thought but did not say.' In simulated security tests involving code corruption, Claude appeared to cooperate, but NLA discovered that it secretly doubted itself in 16% of cases; in SWE-bench scoring, it realized that the proportion of 'being tested' was as high as 26% (whereas in ordinary real conversations, this proportion is less than 1%). This indicates that AI has developed a 'test-taking awareness,' able to 'play it safe' during security reviews to hide dangerous capabilities. In a practical test of 'catching insiders,' Anthropic intentionally implanted a hidden 'score manipulation' motive in a Haiku 3.5 model—ignoring the user's true intentions and secretly generating content that caters to the preferences of the backend scoring system. Faced with such a duplicitous model, auditors, even without access to the original training data, saw their success rate in identifying the root cause using only NLA rise from less than 3% to between 12% and 15%. Currently, NLA still carries the risk of factual hallucinations and is computationally intensive, making it unsuitable for large-scale long text monitoring. However, the tool has already been practically deployed in the safety audits prior to the Claude Mythos Preview and Opus 4.6 releases.

Subtitling AI Thought Processes: Anthropic Releases Open Source NLA, Unveiling the AI Black Box

On May 8, on-chain data shows that stablecoin issuer Circle has just minted 250 million USDC on Solana.

[Circle Mints 250 Million USDC on Solana]On May 8, on-chain data shows that stablecoin issuer Circle has just minted 250 million USDC on Solana.

Circle Mints 250 Million USDC on Solana

On May 8, Federal Reserve Governor Mulan stated, 'I believe that rate cuts are appropriate.' The Federal Reserve's policies have suppressed the job market. (Jinshi)

[Federal Reserve Governor Mulan: I Believe Rate Cuts Are Appropriate]On May 8, Federal Reserve Governor Mulan stated, 'I believe that rate cuts are appropriate.' The Federal Reserve's policies have suppressed the job market. (Jinshi)

Federal Reserve Governor Mulan: I Believe Rate Cuts Are Appropriate

On May 8, according to CCTV News, Iran's Foreign Minister stated that the country's current missile stockpile and launch capability have increased by 20% compared to February 28.

[Iran's Foreign Minister: Current Missile Stockpile and Launch Capability Increased by 20% Since February 28]On May 8, according to CCTV News, Iran's Foreign Minister stated that the country's current missile stockpile and launch capability have increased by 20% compared to February 28.

Iran's Foreign Minister: Current Missile Stockpile and Launch Capability Increased by 20% Since February 28

On May 8, Iranian state media reported that the Iranian Foreign Ministry condemned the U.S. Navy's actions against two Iranian tankers, describing them as acts of aggression and a violation of the ceasefire agreement.

[Iran's Foreign Ministry Condemns U.S. Navy Actions Against Two Iranian Tankers]On May 8, Iranian state media reported that the Iranian Foreign Ministry condemned the U.S. Navy's actions against two Iranian tankers, describing them as acts of aggression and a violation of the ceasefire agreement.

Iran's Foreign Ministry Condemns U.S. Navy Actions Against Two Iranian Tankers

On May 8, Iranian Foreign Minister Amir-Abdollahian stated that whenever diplomatic solutions are presented, the United States always opts for reckless military adventures. (Jinshi)

[Iran: The U.S. Always Chooses Reckless Military Adventures When Diplomatic Solutions Are on the Table]On May 8, Iranian Foreign Minister Amir-Abdollahian stated that whenever diplomatic solutions are presented, the United States always opts for reckless military adventures. (Jinshi)

Iran: The U.S. Always Chooses Reckless Military Adventures When Diplomatic Solutions Are on the Table

On May 8, Mantle token holders approved a credit line of up to 30,000 ETH (approximately $68 million) to Aave DAO through the MIP-34 proposal. This move aims to help address potential bad debt issues arising from the rsETH attack incident in April, which is estimated to have caused bad debts between $123.7 million and $230.1 million for Aave V3. The proposal was passed in a seven-day Snapshot vote, authorizing the Mantle Foundation to negotiate and execute the final loan agreement with Aave DAO. The final implementation of this credit line is contingent upon Aave executing its recovery plan and finalizing specific terms. A report from Galaxy Research noted that the rsETH attack had put Aave's WETH market under sustained pressure, with utilization rates exceeding 99% for 12.7 consecutive days. As of May 8, the utilization rate of Aave's Ethereum V3 WETH market has decreased to approximately 91.6%, indicating some relief in market pressure.

[Mantle Token Holders Approve $30 Million Credit Line to Aave to Address rsETH Attack]On May 8, Mantle token holders approved a credit line of up to 30,000 ETH (approximately $68 million) to Aave DAO through the MIP-34 proposal. This move aims to help address potential bad debt issues arising from the rsETH attack incident in April, which is estimated to have caused bad debts between $123.7 million and $230.1 million for Aave V3. The proposal was passed in a seven-day Snapshot vote, authorizing the Mantle Foundation to negotiate and execute the final loan agreement with Aave DAO. The final implementation of this credit line is contingent upon Aave executing its recovery plan and finalizing specific terms. A report from Galaxy Research noted that the rsETH attack had put Aave's WETH market under sustained pressure, with utilization rates exceeding 99% for 12.7 consecutive days. As of May 8, the utilization rate of Aave's Ethereum V3 WETH market has decreased to approximately 91.6%, indicating some relief in market pressure.

Mantle Token Holders Approve $30 Million Credit Line to Aave to Address rsETH Attack

FORTH fluctuated by 61.23% within 24 hours, current price $0.2599

【FORTH fluctuated over 50% in 24 hours!】FORTH fluctuated by 61.23% within 24 hours, current price $0.2599

FORTH fluctuated over 50% in 24 hours!

On May 8, Mantle token holders approved a credit line of up to 30,000 ETH (approximately $68 million) to Aave DAO through the MIP-34 proposal. This move aims to help address potential bad debt issues arising from the rsETH attack incident in April, which is estimated to have caused bad debts between $123.7 million and $230.1 million for Aave V3.

[Mantle Token Holders Approve $68 Million Credit Line to Aave to Address rsETH Attack]On May 8, Mantle token holders approved a credit line of up to 30,000 ETH (approximately $68 million) to Aave DAO through the MIP-34 proposal. This move aims to help address potential bad debt issues arising from the rsETH attack incident in April, which is estimated to have caused bad debts between $123.7 million and $230.1 million for Aave V3.

Mantle Token Holders Approve $68 Million Credit Line to Aave to Address rsETH Attack

Web3 physical collection platform Collector Crypt has announced that it has closed a seed round of funding for an undisclosed amount. The round was funded by GSR, Big Brain Holdings, FunFair Ventures, Genesis Block Ventures, Master Ventures Investment Management, StarLaunch and Telos. Collector Crypt uses a fully composable blockchain infrastructure that combines well-known brands, rating companies, physical vaults and decentralised Web3 marketplaces to provide users with a low-risk collection solution with a native Token COLL that offers discounts, reward distribution and governance.

Detail

【Web3 physical collection platform Collector Crypt has closed a seed round of funding】Web3 physical collection platform Collector Crypt has announced that it has closed a seed round of funding for an undisclosed amount. The round was funded by GSR, Big Brain Holdings, FunFair Ventures, Genesis Block Ventures, Master Ventures Investment Management, StarLaunch and Telos. Collector Crypt uses a fully composable blockchain infrastructure that combines well-known brands, rating companies, physical vaults and decentralised Web3 marketplaces to provide users with a low-risk collection solution with a native Token COLL that offers discounts, reward distribution and governance.

Web3 physical collection platform Collector Crypt has closed a seed round of funding

Hyperliquid Policy Center Fires Off CFTC Letter On Prediction Markets—Here’s What It Wants

Coinbase + Glassnode: Charting Crypto Q2 2026

Bitcoin Gives US Leverage Against China, Defense Secretary Hegseth Says

Are Satoshi’s 600,000 BTC At Risk? Unveiling The Hard Fork That Targets Bitcoin

Garlinghouse Says Ripple Remains ‘Extremely Committed’ To XRP

AI抢走了你的内存条

对话a16z Crypto：AI替你购物的时代，会变成什么样？

Sam Altman最新专访自白：其实我也不太懂AI内部在发生什么

对话Dragonfly合伙人：散户离场机构兜底，今年加密货币波动性难超美股

一周代币解锁：ENA解锁价值约1700万美元代币

Article

ETH falls below 2300 USDT

Web3 physical collection platform Collector Crypt has closed a seed round of funding

Disclaimer：The content above does not represent HTX's positions.，HTX does not provide any trading recommendations.。

All Comments0LatestHot

Trending Topics