Lookonchain APP

App Store

REA, which claims to "replicate any software", has been put to practical testing: out of 51 public attempts, most successful cases only involved partial functionality.

52 minutes ago

Beating AI News (Dongcha) reports that independent technical author quanglee reviewed 51 public use cases of REA, finding that so-called "replication success" often only means restoring a single function. On October 12, he released his findings: 9 of the cases came from REA project contributors, all achieving their respective goals; among the other 42 independent attempts, 20 were fully completed, 16 partially completed, and 6 failed. The crux lies in the definition of "completion". Among the 20 independent successful cases, 9 targeted small games or mini-programs, 6 only restored one function of a larger software, and 5 did not even disclose the target name. One notable case is the puzzle game Untangle: after Agent analyzed its 167 KB program, it created a playable web version, with 52 out of 54 technical conclusions being fully accurate. However, developers later audited and found that 10 function and field names in the result did not come from REA’s analysis—likely because the model memorized the original’s open-source code. More complex projects require significant effort. In the reverse engineering of *Touhou Imperishable Night*, REA contributors matched 1106 out of 1107 functions to the original machine code, but the process took about two weeks and built on prior manual research. Another attempt to replicate voice input software Wispr Flow analyzed parts of its workflow (recording, transcription, text pasting) but only produced a 64-line Python script. The original’s core features—cloud-based models, context processing, real-time feedback—were not fully replicated. REA can help Agent call tools like Ghidra and IDA to analyze existing programs and rebuild partial functions. However, a commercial software’s capabilities are not limited to its installation package; server-side models, data, and complex business logic cannot be directly restored by decompiling the client.

Relevant content

Castle Securities: U.S. retail investors net $91 billion in U.S. equities last week, marking a 15-week high.

Bloomberg reported Tuesday, citing data from Citadel Securities, that retail investors on the platform net bought $9.1 billion in U.S. stocks last week, hitting a 15-week high. Semiconductors were a key area of capital concentration: retail investors on Citadel Securities' platform net purchased $2.6 billion in semiconductor stocks last week, the third-highest weekly level in the past year, while AI infrastructure stocks also drew around $2.4 billion in buying. Scott Rubner, head of stock and derivatives strategy at Citadel Securities, attributed the surge in buying to improved directional confidence. The platform's retail investors' weekly buying bias reached the 90th percentile, while total trading volume stood at the 42nd percentile over the past year.

9 minutes ago

JPMorgan Chase: Microsoft’s Maia300 delivers 40% higher inference performance, cutting costs for deploying millions of chips.

JPMorgan Chase’s October 12, 2026 research report notes that Microsoft’s Maia 300 delivers 40% more tokens per megawatt than leading GPU systems, with rack-level token throughput at full 96-accelerator configuration reaching 7 times that of Maia 200. Maia 200 offers 1.7 times more tokens per watt than GPU infrastructure. Reuters reports Microsoft plans to deploy over 1 million Maia 300 chips in 2027. Microsoft’s fiscal 2027 capital expenditure is projected at around $210 billion, with calendar 2026 capex at roughly $175 billion. JPMorgan assesses that Maia’s adoption as the primary inference system will boost Copilot performance and reduce per-compute-unit costs, lifting Microsoft Cloud’s gross margin; the bank maintains an Overweight rating on Microsoft with a target price of $625.

9 minutes ago

WeChat Work and WorkBuddy achieve deep integration: AI can read chat and meeting records, and track computing power costs.

Beating AI Express News: Tencent announces deep integration of WeChat Work and WorkBuddy. In the new WeChat Work 5.0.12, WorkBuddy is officially added to the corporate directory, allowing employees to chat with it, assign tasks, or add it to work groups just like contacting colleagues. For example, users can ask it to summarize group chat discussions, organize meeting to-dos, or draft emails, with results returned directly upon completion. WorkBuddy 5.8.0 supports direct login via WeChat Work accounts. Within the scope of enterprise authorization, it can access 10 office capabilities including chat messages, documents, meetings, and WeChat Drive. Employees no longer need to repeatedly upload materials; they can have the AI process tasks by leveraging previous discussions and project files, and generated outputs can be saved to WeChat Drive with one click. Additionally, the WorkBuddy bot can integrate with WeChat Work’s built-in AI assistant "Dayuan" to collaborate on more complex tasks. Administrators can assign WorkBuddy accounts, set data access permissions, and use the "Observability Center" (which launched its public beta on October 13) to monitor departmental point consumption, model usage, task duration, and even track issues in specific execution steps. Currently, the new interoperability features are open to certified WeChat Work enterprises, with limited free trial spots available.

9 minutes ago

Binance Leverage to Add 4 bStocks Collateral Assets Including MDBB on October 14

According to an official announcement, Binance’s Cross Margin, Portfolio Margin, and Portfolio Margin Pro will add four bStocks tokens—MongoDB (MDBB), BlackBerry (BBB), PenguinSolutions (PENGB), and ProShares Ultra VIX ETF (UVXYB)—as eligible collateral assets at 12:00 UTC on October 14, 2026. Leveraged trading for the corresponding bStocks trading pairs will also be opened concurrently.

9 minutes ago

A trader, bullish on the subsequent performance of Google's new model, went long on GOOGL with 20x leverage following the news of its launch.

According to TradingBeats monitoring, wallet address 0x6936…9417 opened a long position in GOOGL following the announcement of Gemini 4 Argon’s launch. As of now, the address holds 2,000 contracts, with the position valued at roughly $710,000 and an average entry price of approximately $355.09. The position uses 20x isolated leverage, with a liquidation price of around $344.45. Google announced on Tuesday that Gemini 4 Argon is now available globally to Google AI Pro and Ultra subscribers, developers, and enterprises. Prior to this, the model was first rolled out on September 30 to cybersecurity partners in the Fairwind program.

9 minutes ago

Fed's Collins: The Federal Reserve still needs to raise interest rates further before the end of this year.

Boston Federal Reserve President Susan Collins said Tuesday she currently believes the U.S. Federal Reserve may need to raise interest rates further before the end of this year to ensure inflation returns to its 2% target in a timely and sustained manner. However, she emphasized that any further policy tightening and the timing of such actions will depend heavily on upcoming economic data and how these data shape her assessment of the economic outlook and balance of risks. Collins warned that artificial intelligence (AI) infrastructure development could trigger an excessively large investment boom. If related demand further fuels household spending and overheats the labor market, inflation could remain above the Fed's target for an extended period.

9 minutes ago

Popular tokens

BitcoinEthereumHyperliquidSolanaTRONBNBTetherAaveXRPPepeFartcoinOndoJupiterUniswapBonkPendleEthenaArbitrumAvalancheLidoChainlinkPolygonDogecoinCardano