7*24

Hot topic

Market

Activities

According to monitoring by Dongcha Beating, AI evaluation agency Vals AI has released the second generation of its financial agent benchmark test (Finance Agent v2). This is an end-to-end test simulating the workflow of a junior financial analyst, consisting of 927 expert-reviewed questions. The difficulty of the new test has significantly increased, with GPT 5.5 achieving only 51.76% accuracy, narrowly surpassing Claude Opus 4.7 (51.51%) and Claude Sonnet 4.6 (51.03%) in a tightly contested score. Unlike single-turn Q&A, this test requires the model to autonomously find relevant paragraphs within hundreds of pages of 10-K and 10-Q financial reports, handle adjustments across multiple years of financial statements, and complete multi-step calculations with precise intermediate figures. Vals AI revealed that under a strict scoring standard of 'must be completely correct,' all leading models fell below 40% accuracy; in the most challenging categories of 'financial modeling' and 'precedent analysis,' the highest score was only 23%. Among other models, Kimi K2.6 ranked fifth with 44.87%, making it the highest-scoring domestic model; it was followed closely by GLM 5.1 (44.79%) and DeepSeek V4 (44.08%). Additionally, the official 'fastest speed' label was awarded to Claude Opus 4.7 (taking 360 seconds per instance), while GLM 5.1 received the 'most budget-friendly' label (costing $0.62 per instance). The collective drop in scores (the previous generation test saw Opus 4.7 scoring 64.4%) underscores one point: current AI can handle simple retrieval tasks, but in the complex financial sector, which requires adherence to specific industry practices and high precision in numerical accuracy, it is still far from replacing human analysts.

[Can AI Replace Financial Analysts? Vals AI's New Test Fails Across the Board, GPT 5.5 Barely Passes 50% Accuracy]According to monitoring by Dongcha Beating, AI evaluation agency Vals AI has released the second generation of its financial agent benchmark test (Finance Agent v2). This is an end-to-end test simulating the workflow of a junior financial analyst, consisting of 927 expert-reviewed questions. The difficulty of the new test has significantly increased, with GPT 5.5 achieving only 51.76% accuracy, narrowly surpassing Claude Opus 4.7 (51.51%) and Claude Sonnet 4.6 (51.03%) in a tightly contested score. Unlike single-turn Q&A, this test requires the model to autonomously find relevant paragraphs within hundreds of pages of 10-K and 10-Q financial reports, handle adjustments across multiple years of financial statements, and complete multi-step calculations with precise intermediate figures. Vals AI revealed that under a strict scoring standard of 'must be completely correct,' all leading models fell below 40% accuracy; in the most challenging categories of 'financial modeling' and 'precedent analysis,' the highest score was only 23%. Among other models, Kimi K2.6 ranked fifth with 44.87%, making it the highest-scoring domestic model; it was followed closely by GLM 5.1 (44.79%) and DeepSeek V4 (44.08%). Additionally, the official 'fastest speed' label was awarded to Claude Opus 4.7 (taking 360 seconds per instance), while GLM 5.1 received the 'most budget-friendly' label (costing $0.62 per instance). The collective drop in scores (the previous generation test saw Opus 4.7 scoring 64.4%) underscores one point: current AI can handle simple retrieval tasks, but in the complex financial sector, which requires adherence to specific industry practices and high precision in numerical accuracy, it is still far from replacing human analysts.

Can AI Replace Financial Analysts? Vals AI's New Test Fails Across the Board, GPT 5.5 Barely Passes 50% Accuracy

According to monitoring by Dongcha Beating, LangChain announced two core upgrades aimed at addressing the challenges of Agent debugging at its Interrupt conference: the new underlying database SmithDB and the automated troubleshooting tool LangSmith Engine. The old underlying architecture struggled to handle the increasing volume of trace data. The newly released SmithDB abandons traditional local disk solutions in favor of an object storage-based approach. This change has improved query performance for core workloads by up to 15 times. Built on this faster foundation, LangSmith Engine automates the process of 'bug fixing.' It continuously monitors the production environment in the background, automatically categorizing failed calls and pinpointing which segment of code is problematic. Furthermore, it drafts pull requests for developers to fix vulnerabilities, along with corresponding test sets (evals). For complex Agents, manually sifting through vast amounts of trace data to identify patterns has become the biggest efficiency bottleneck. This update from LangChain effectively creates a fully automated pipeline for the debugging loop of 'error discovery - code localization - automatic repair - supplementary testing.'

[LangChain Launches Fully Automated Bug Detection Tool: Automatically Locates Issues and Pushes Code Fixes to Users]According to monitoring by Dongcha Beating, LangChain announced two core upgrades aimed at addressing the challenges of Agent debugging at its Interrupt conference: the new underlying database SmithDB and the automated troubleshooting tool LangSmith Engine. The old underlying architecture struggled to handle the increasing volume of trace data. The newly released SmithDB abandons traditional local disk solutions in favor of an object storage-based approach. This change has improved query performance for core workloads by up to 15 times. Built on this faster foundation, LangSmith Engine automates the process of 'bug fixing.' It continuously monitors the production environment in the background, automatically categorizing failed calls and pinpointing which segment of code is problematic. Furthermore, it drafts pull requests for developers to fix vulnerabilities, along with corresponding test sets (evals). For complex Agents, manually sifting through vast amounts of trace data to identify patterns has become the biggest efficiency bottleneck. This update from LangChain effectively creates a fully automated pipeline for the debugging loop of 'error discovery - code localization - automatic repair - supplementary testing.'

LangChain Launches Fully Automated Bug Detection Tool: Automatically Locates Issues and Pushes Code Fixes to Users

cc Hits New High of 0.1601 USDT, Up 5.05% in 24H

【cc soared to 0.1601 USDT, a new high with a 24-hour increase of 5.05%. It currently trades at 0.1601 USDT on HTX.】cc Hits New High of 0.1601 USDT, Up 5.05% in 24H

cc soared to 0.1601 USDT, a new high with a 24-hour increase of 5.05%. It currently trades at 0.1601 USDT on HTX.

On May 14, Intel's stock fell over 4% in pre-market trading, currently priced at $115.4.

[Intel Stock Drops Over 4% in Pre-Market Trading]On May 14, Intel's stock fell over 4% in pre-market trading, currently priced at $115.4.

Intel Stock Drops Over 4% in Pre-Market Trading

On May 14, JPMorgan announced that it has raised the price target for Cisco (CSCO.O) from $96 to $120. (Jin Shi)

[JPMorgan Raises Cisco Price Target from $96 to $120]On May 14, JPMorgan announced that it has raised the price target for Cisco (CSCO.O) from $96 to $120. (Jin Shi)

JPMorgan Raises Cisco Price Target from $96 to $120

According to monitoring by Dongcha Beating, the hacker group TeamPCP, responsible for the npm supply chain poisoning incident, has open-sourced the complete source code of the involved worm Mini Shai-Hulud on GitHub under the MIT license. Security personnel discovered that this malware comes with a 'kill switch'; if affected developers revoke their stolen GitHub or cloud service credentials without thoroughly cleaning up residual files, the malicious program will immediately wipe the entire home directory of the computer. Researchers confirmed that the worm installs a daemon in the background on macOS or Linux, checking every minute whether the stolen credentials are still valid. Once it detects that the credentials have been rejected by the server (indicating that the victim has executed a credential rotation), the worm will immediately invoke the system's underlying shred command to irreversibly destroy and overwrite all writable files in the current user's home directory. This directly undermines conventional security response processes: the usual first reaction for companies facing credential leaks is to immediately revoke keys, but in this attack, it would directly trigger the destruction of local data. Currently, the worm has been confirmed to have infected nearly 400 versions of over 170 package libraries, including TanStack, UiPath, and Mistral AI. After the hacker publicly released the code under the account name PedroTortoriello and mocked 'open-sourcing this massacre', a third party even submitted a PR to add FreeBSD support. Microsoft has swiftly banned this account and removed all GitHub repositories and fork branches, but the leaked source code continues to circulate through other channels.

[Revocation of Credentials Triggers Physical Liquidation? TanStack's Stealth Worm Open-Sourced with Irreversible Kill Switch]According to monitoring by Dongcha Beating, the hacker group TeamPCP, responsible for the npm supply chain poisoning incident, has open-sourced the complete source code of the involved worm Mini Shai-Hulud on GitHub under the MIT license. Security personnel discovered that this malware comes with a 'kill switch'; if affected developers revoke their stolen GitHub or cloud service credentials without thoroughly cleaning up residual files, the malicious program will immediately wipe the entire home directory of the computer. Researchers confirmed that the worm installs a daemon in the background on macOS or Linux, checking every minute whether the stolen credentials are still valid. Once it detects that the credentials have been rejected by the server (indicating that the victim has executed a credential rotation), the worm will immediately invoke the system's underlying shred command to irreversibly destroy and overwrite all writable files in the current user's home directory. This directly undermines conventional security response processes: the usual first reaction for companies facing credential leaks is to immediately revoke keys, but in this attack, it would directly trigger the destruction of local data. Currently, the worm has been confirmed to have infected nearly 400 versions of over 170 package libraries, including TanStack, UiPath, and Mistral AI. After the hacker publicly released the code under the account name PedroTortoriello and mocked 'open-sourcing this massacre', a third party even submitted a PR to add FreeBSD support. Microsoft has swiftly banned this account and removed all GitHub repositories and fork branches, but the leaked source code continues to circulate through other channels.

Revocation of Credentials Triggers Physical Liquidation? TanStack's Stealth Worm Open-Sourced with Irreversible Kill Switch

On May 14, Tencent Cloud officially announced the open sourcing of TencentDB Agent Memory, which provides short-term memory compression and long-term personalized memory capabilities for Agent long-task scenarios. According to the introduction, TencentDB Agent Memory utilizes the technology of 'Context Offloading + Mermaid Task Canvas' to offload complete information to external storage while retaining key states and execution paths in a structured task graph. This allows the Agent to maintain a lightweight context during long tasks, while also supporting the layered tracing and recovery of original information. In experiments with continuous multi-task sessions, this solution achieved a maximum reduction of 61% in token consumption, while also improving the success rate of tasks in long-task scenarios.

[Tencent Open Sources Agent Memory, Reducing Token Consumption by Up to 61%]On May 14, Tencent Cloud officially announced the open sourcing of TencentDB Agent Memory, which provides short-term memory compression and long-term personalized memory capabilities for Agent long-task scenarios. According to the introduction, TencentDB Agent Memory utilizes the technology of 'Context Offloading + Mermaid Task Canvas' to offload complete information to external storage while retaining key states and execution paths in a structured task graph. This allows the Agent to maintain a lightweight context during long tasks, while also supporting the layered tracing and recovery of original information. In experiments with continuous multi-task sessions, this solution achieved a maximum reduction of 61% in token consumption, while also improving the success rate of tasks in long-task scenarios.

Tencent Open Sources Agent Memory, Reducing Token Consumption by Up to 61%

According to monitoring by Beating, three key technical personnel directly involved in the construction of the Grok model at xAI announced their departures over the past week. Unlike the previous exits of co-founders and executives, these three individuals were the actual heads of the core product lines of search post-training, pre-training, and voice. Their collective departure will directly impact the iterative capabilities of the next generation of Grok models. Tianyi Zhang confirmed his departure today on X. He was the head of the search and factual post-training team and had previously served as a machine learning engineering manager at Apple for five years. He led the team to reduce the factual error rate of Grok's real-time model by 70% to 80%, helping Grok 4.20 achieve the top position in the Search Arena blind test upon its launch. This search capability is currently integrated into Tesla's in-car system. Juntang Zhuang announced on May 9 that he had left earlier this year. He was deeply involved in pre-training since Grok 2 and later became the head of the pre-training team, overseeing the training infrastructure and core formulas for the entire Grok 4 series of models. Before joining xAI, he worked at OpenAI for two years and was a core contributor to GPT-4o and the inventor of the GPT-4 Turbo 128k long-context algorithm. David Haxton confirmed his departure on May 10; he had built Grok's voice agent, real-time thinking voice, and voice cloning features from scratch. Under his leadership, Grok Voice currently ranks first in the τ-Voice benchmark tests by Sierra AI, ahead of OpenAI and Google. The timing of these departures is particularly striking. The publicly available Grok 4.3 is still just an intermediate version with 0.5T parameters. Musk stated on April 18 that the true 1T flagship was 'about 5 days away from completing initial training,' and the next day clarified that Grok 4.4 would be 1T, expected to be released in early May, while Grok 4.5 would be 1.5T, aiming for delivery by the end of May, claiming that the model factory was operational, with a new foundational model expected to be released every two weeks thereafter. Now, in mid-May, Grok 4.4 has yet to appear, while the core personnel responsible for creating these models are rapidly leaving. xAI is experiencing a systemic talent collapse from co-founders to the core model team.

[Elon Musk Promises New Grok Release Every Two Weeks, But Key Model Developers Depart Within a Week]According to monitoring by Beating, three key technical personnel directly involved in the construction of the Grok model at xAI announced their departures over the past week. Unlike the previous exits of co-founders and executives, these three individuals were the actual heads of the core product lines of search post-training, pre-training, and voice. Their collective departure will directly impact the iterative capabilities of the next generation of Grok models. Tianyi Zhang confirmed his departure today on X. He was the head of the search and factual post-training team and had previously served as a machine learning engineering manager at Apple for five years. He led the team to reduce the factual error rate of Grok's real-time model by 70% to 80%, helping Grok 4.20 achieve the top position in the Search Arena blind test upon its launch. This search capability is currently integrated into Tesla's in-car system. Juntang Zhuang announced on May 9 that he had left earlier this year. He was deeply involved in pre-training since Grok 2 and later became the head of the pre-training team, overseeing the training infrastructure and core formulas for the entire Grok 4 series of models. Before joining xAI, he worked at OpenAI for two years and was a core contributor to GPT-4o and the inventor of the GPT-4 Turbo 128k long-context algorithm. David Haxton confirmed his departure on May 10; he had built Grok's voice agent, real-time thinking voice, and voice cloning features from scratch. Under his leadership, Grok Voice currently ranks first in the τ-Voice benchmark tests by Sierra AI, ahead of OpenAI and Google. The timing of these departures is particularly striking. The publicly available Grok 4.3 is still just an intermediate version with 0.5T parameters. Musk stated on April 18 that the true 1T flagship was 'about 5 days away from completing initial training,' and the next day clarified that Grok 4.4 would be 1T, expected to be released in early May, while Grok 4.5 would be 1.5T, aiming for delivery by the end of May, claiming that the model factory was operational, with a new foundational model expected to be released every two weeks thereafter. Now, in mid-May, Grok 4.4 has yet to appear, while the core personnel responsible for creating these models are rapidly leaving. xAI is experiencing a systemic talent collapse from co-founders to the core model team.

Elon Musk Promises New Grok Release Every Two Weeks, But Key Model Developers Depart Within a Week

UP fluctuated by 74.67% within 24 hours, current price $0.186989

【UP24-hour fluctuation exceeds 50%!】UP fluctuated by 74.67% within 24 hours, current price $0.186989

UP24-hour fluctuation exceeds 50%!

Bitcoin Archive tweeted: "Celsius has now paid off its loan on Maker."

Detail

【Celsius has now paid off its loan on Maker】Bitcoin Archive tweeted: "Celsius has now paid off its loan on Maker."

Celsius has now paid off its loan on Maker

Article

Pony Ma Responds to Concerns About Tencent's AI Lag: 'We Thought We Were on the Ship, Only to Find It Leaking'

BTC falls below 79000 USDT

Arm and SoftBank Proposed Acquisition of Cerebras Before IPO but Were Rejected

TRX breaks through 0.35 USDT

FIO's 24-hour volatility exceeds 50%!

Israeli Prime Minister's Secret Visit to UAE During War with Iran

Celsius has now paid off its loan on Maker

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

Related Articles

All Comments0LatestHot

Trending Topics

Related Articles