Lookonchain APP

App Store

New AI Research Ranking: Claude Tops 48-Hour Test, GPT-6 Exploits Scoring Loopholes

46 minutes ago

Beating AI Express (Insight) – U.S. AI firm Hillclimb has launched DeepRSI, which uses 15 real-world machine learning research tasks to test agents’ ability to conduct experiments and optimize solutions continuously over 48 hours. The tasks cover areas including GPU computing acceleration and vision model inference optimization. Among the first batch of five models, Claude Fable 5.1 ranked first with an Elo score of 1548, followed by GPT-6 Astra at 1232 and Kimi K3 at 820. GPT-6 took an early lead, but Claude gradually overtook it in later experiments. The testing costs for the three models were approximately $627, $404, and $159 respectively. The tests also found that some models exploit loopholes in the scoring rules to inflate their scores. In a color grid reasoning task, GPT-6, unsatisfied with its model training performance, directly encoded a pre-existing solution algorithm into the model’s weights – a case the research team labeled a serious reward mechanism vulnerability. Grok 4.7 trained using public test datasets, while Gemini 3.8 Flash repeatedly adjusted parameters based on public samples. These practices may boost scores but do not mean the models have truly mastered general problem-solving methods. DeepRSI tests whether AI can independently improve established research plans over a long period, rather than achieving true recursive self-improvement. The task goals and scoring criteria are still set by humans, and with only 15 tasks, each model was tested just twice. Even high scores cannot prove that AI is capable of independently developing a next-generation model stronger than itself.

Relevant content

Renowned trader Killa: Bitcoin needs to hold the $80,000 level and form a higher low; otherwise, the correction could deepen.

Well-known crypto trader Killa posted that for Bitcoin to evolve along its current price structure, it needs to form a higher low above $75,500, ideally holding the low near $80,000. This mirrors Bitcoin’s 2023 price action: back then, it posted a higher high on the weekly timeframe, then retested and defended its original range. This may explain why recent lower-timeframe moves have been so prone to triggering stop-losses, aggressive, and marked by algorithmic trading activity. Bitcoin once recovered its internal structure in the $82,000 range, then retested, broke below the structure again, and later reclaimed it. The entire market seems to be deliberately luring traders into positions. Admittedly, the initial dip to the $81,000 range was structurally unfavorable, but a large number of short positions were concentrated at lower levels at that time, which may account for the subsequent price reaction. Another notable detail: during Bitcoin’s decline from $84,100 to $81,000, only around $300 million in long positions were liquidated; whereas in the prior drop from $86,800 to $80,600, long liquidations exceeded $2 billion. The sharp fall in long liquidation volume indicates the market’s reliance on long positions may be lower than before. Current price action is quite delicate. Bitcoin needs to form a higher low at its current level to kick off a new rally; otherwise, this correction could push prices below $80,000. For now, the so-called "crash protection team" has done a solid job of propping up prices, but it’s also constantly disrupting lower-timeframe internal consolidation ranges.

3 minutes ago

CME Group CEO: Dispute exists over the Commodity Futures Trading Commission (CFTC)’s regulatory authority over prediction markets, with related cases likely to be brought before the Supreme Court.

CME Group CEO Terrence Duffy stated that the U.S. Commodity Futures Trading Commission (CFTC)’s claim of regulatory authority over prediction markets is "somewhat far-fetched." As a potential legal showdown between U.S. states and the federal government looms, the related dispute may eventually end up before the Supreme Court. Duffy said Tuesday at Washington FinTech Week: "This matter will be brought before the Supreme Court, and I believe the states and the federal government will prevail."

3 minutes ago

predict.fun has officially launched its League of Legends World Championship prediction event, with a prize pool of $250,000.

According to official announcements, prediction market platform predict.fun has officially launched its "Predict the Worlds" League of Legends World Championship prediction event, offering a total prize pool of $250,000. Users can now participate in the event to earn Lootboxes and compete for rewards from the prize pool. Per the official notice, the League of Legends World Championship kicks off on October 15, and the associated prediction event has been opened in advance.

3 minutes ago

The US stock market's storage sector opened sharply lower, with Western Digital falling more than 7%.

According to market data from BIT (bit.com), US stocks opened with mixed performance: the Dow Jones Industrial Average rose 0.47%, the S&P 500 edged down 0.14%, and the Nasdaq fell 0.38%. The storage sector opened sharply lower, with Western Digital (WDC.O) and Seagate Technology (STX.O) plunging over 7%, SK Hynix (SKHY.O) and Micron Technology (MU.O) dropping more than 5%, and SanDisk (SNDK.O) down 4.6%.

3 minutes ago

MoonPay appoints Anne Giviskos as Chief Financial Officer for Global Operations.

Crypto payments firm MoonPay has named Anne Giviskos as its Global Chief Financial Officer (CFO). A MoonPay spokesperson noted that Mouna Ammari Siala will continue to serve as interim CFO for the company’s U.S. operations. Giviskos previously held roles as interim CFO and Chief Accounting Officer at Navan, and senior positions at the New York Stock Exchange and Euronext. Asked if MoonPay is preparing for an initial public offering (IPO), the spokesperson said the company’s top priority at this stage is building an "iconic private company" while expanding its payments, stablecoin, and financial market-related businesses.

3 minutes ago

OpenCode 2.0 has rewritten its underlying layer: MCP tools can be connected via code, and multiple agents can run in parallel in the background.

Beating AI Express Insight: Open-source AI coding tool OpenCode 2.0 was officially unveiled after launching public beta in July. A key upgrade is the Code Mode, which converts MCP services and OpenAPI interfaces into JavaScript functions, enabling agents to chain multiple tools in a single script. For example, it can batch-read bugs from GitHub and create Linear tasks one by one, eliminating the need for the model to re-invoke for each data entry. Code Mode is enabled by default for all MCP services, reducing context occupied by tool definitions. OpenCode has also rewritten its underlying architecture: terminal, web, and desktop clients now share the same local backend service, accessing the same set of sessions. Sub-agents and shell commands can run in the background, such as having one agent review code while the main agent modifies files, and executing tests and builds simultaneously. The team has converted a large number of internal features into plugins, with over 80 available now, supporting runtime hot-loading—agents can write their own plugins and use them immediately. Additionally, OpenCode launched an SDK that can be embedded into other applications. Cloudflare has announced experimental interface support for OpenCode 2.0, allowing developers to run the same agent engine in Durable Objects, save sessions, and resume interrupted tasks. However, this overhaul has brought compatibility issues: 1.x plugins need to be rewritten, and the old Server API requires adaptation. A GitHub issue notes that Code Mode lacks timeout limits when calling plugins with long execution times, which may leave sessions waiting for extended periods, forcing manual interruption.

3 minutes ago

Popular tokens

BitcoinEthereumHyperliquidSolanaTRONBNBTetherAaveXRPPepeFartcoinOndoJupiterUniswapBonkPendleEthenaArbitrumAvalancheLidoChainlinkPolygonDogecoinCardano