Lookonchain APP

App Store

OpenAI has proposed a cutting-edge AI training "safety validation" framework, recommending that relevant parties complete safety documentation before initiating AI training runs.

41 minutes ago

动察 Beating AI 快讯,OpenAI 官方发布长文《Towards safety cases for frontier AI training》(面向前沿 AI 训练的安全论证)。文中提及,前沿强化学习训练在继续推进前,应先完成结构化安全文档,理想形态对标航空、核电行业的「安全论证」。公司承认,因能力跃迁后的涌现复杂性,AI 很难做到与传统高危行业同等严谨,目前该框架仍在内部落地并面向外界征求反馈。OpenAI 明确只覆盖前沿 RL 训练,内部及对外部署还需另看更广的对齐属性。技术侧,OpenAI 将安全论证拆成对齐训练、隔离与监控三层。对齐方面包括自动/人工审查 RL 环境、调校评分器以防奖励黑客、离线对齐评测与回测,以及禁止自动评分器读取思维链,避免模型学会躲避监控。隔离方面强调多层沙箱与基础设施加固、用训练检查点做红队、限制样本间高带宽通信,并用一次写入存储固化对话记录。监控方面要求模型保持可监测性,对历史事故高召回,新增风险及时补评测,高优先级告警须在约定时限内人工响应或自动暂停训练。运营侧建议包括:由其他团队写异议预演、高管(研究负责人、安全负责人、首席科学家等)可一票否决开跑、训练负责人对安全论证和事故响应问责、安全论证失效时按手册暂停、向内部监督委员会公开、给审计方足够核验权限、设错位严重度分级并允许值班人员呼叫 CEO。系统应默认失败关闭,无法在无监控状态下开跑;同时要能回滚被错位模型污染的下游数据与评分。事故调查方面,OpenAI 提出对标航空调查实践:调查期间对内滚动通报、用消融和重采样追溯训练动力学、做运营与文化复盘、开发不直接拟合事故样本的检测评测,并以事故衍生评测做回归测试。调查结论、复盘与流程改动应在结束后对外披露,受影响第三方应尽快获知。

Relevant content

Tesla’s next-generation supercar delayed again, JPMorgan cuts price target, driving shares down nearly 4%.

Tesla has postponed the unveiling event for its next-generation Roadster from October 1 to October 15. The company cited forecasted severe local weather and the event’s required outdoor setting as reasons for the delay. The event is still expected to take place at SpaceX’s test facility in McGregor, Texas. Amid the potential impact, Tesla closed today at $357.45, down 3.94%. On Hyperliquid, TSLA was temporarily trading at $357.8 during the overnight session. Separately, per TradingBeats’ monitoring, during regular U.S. stock trading hours on September 28, the estimated trading volume of TSLA perpetual contracts on Hyperliquid via Trade[XYZ] was around $17.52 million, down roughly 27% from the same period the previous trading day. In the same period, JPMorgan lowered Tesla’s price target from $445 to $415, a roughly 6.7% cut, while maintaining a "neutral" rating. The market also remains cautious about Tesla’s upcoming Q3 production and delivery figures; Cantor Fitzgerald forecasts these data may come in weaker, and the current decline may be driven by a combination of the delayed product event, the price target cut, and delivery expectation pressure.

10 minutes ago

Whale loses $450K on 5x long ZEC position as token drops 18% from recent high

ZEC has dropped more than 18% from its recent high. Whale 0x362a opened a 5x long on 4,135 $ZEC($5.8M) on @Aster_DEX and is now sitting on over $450K in unrealized losses.

10 minutes ago

A crypto whale opened a 5x long position on ZEC, currently facing an unrealized loss of over $450,000.

According to Lookonchain monitoring, ZEC has fallen more than 18% from its recent peak. Whale address 0x362a opened a 5x leveraged long position on Aster DEX, purchasing 4,135 ZEC tokens valued at around $5.8 million. The address’s ZEC long position has now incurred an unrealized loss exceeding $450,000.

10 minutes ago

Whale 0x362a opened a 5x long on 4,135 $ZEC($5.8M) on @Aster_DEX and is now sitting on over $450K in unrealized losses.

ZEC has dropped more than 18% from its recent high. Whale 0x362a opened a 5x long on 4,135 $ZEC($5.8M) on @Aster_DEX and is now sitting on over $450K in unrealized losses.

10 minutes ago

Cloudflare overhauls its Command-Line Interface (CLI) for AI use cases, with agents now accounting for 48% of its usage.

Beating AI Express: Cloudflare has significantly expanded its `cf` CLI, which it launched in April this year. Initially supporting only a small number of products, the tool now covers all over 3,000 of Cloudflare’s API operations and is positioned to replace the existing Wrangler, which currently has around 280 command paths. Many of `cf`’s designs are tailored for coding agents: command outputs default to JSON, enabling direct reading by agents like Claude Code and Codex. The new `cf cli search` feature lets agents describe tasks in natural language to automatically find corresponding commands, no prior knowledge of Cloudflare’s APIs required. Cloudflare disclosed that agents now account for nearly half of Wrangler’s usage: the proportion stood at ~25% in March this year, rising to 48% in the latest week. Agents use nearly twice as many distinct commands on average as humans, and are nearly four times more likely to execute six or more commands. The configuration system has also been revised: `cf` now uses a TypeScript-based `cloudflare.config.ts`, allowing agents to leverage type checking and autocompletion to identify configuration errors directly. Existing projects can migrate via `cf migrate`, though some JavaScript projects dependent on Wrangler’s build workflow, plus Rust and Python Workers, will remain on Wrangler for now. Cloudflare plans to release Wrangler’s final major version after `cf` stabilizes, followed by 18 months of maintenance.

10 minutes ago

SlowMist’s Yu Xian: The Chainflip bridge is working to block North Korean hackers from money laundering, but its on-chain transfer tracking lags behind.

SlowMist founder Yu Xian stated in a post that after hours of tracking the Bitget hack, the Chainflip bridge has been making significant efforts to block North Korean hackers from money laundering, but anti-money laundering (AML) and know-your-transaction (KYT) systems are lagging behind the hackers' laundering pace. The laundering ring automates the dispersal of large volumes of funds, cross-transferring them across different chains via various bridges. If the funds are blocked or returned due to risk control measures, the gang immediately switches to another laundering path, ultimately converting all funds to Bitcoin (BTC), then further mixing them via CoinJoin—a Bitcoin-based mixing method—on the Bitcoin network. This is an evolving, sophisticated money laundering operation.

10 minutes ago

Popular tokens

BitcoinEthereumHyperliquidSolanaTRONBNBTetherAaveXRPPepeFartcoinOndoJupiterUniswapBonkPendleEthenaArbitrumAvalancheLidoChainlinkPolygonDogecoinCardano