Lookonchain APP

App Store

Flowith 3.0 Launches: 3D Director's Station Added to Its Canvas, Supports Direct Claude Code Execution

26 minutes ago

Beating AI News Flash: Flowith has released version 3.0, expanding its original AI conversation canvas into a video production and programming workbench. The new 3D director station allows users to directly place characters, adjust camera angles and motion trajectories, generate a simple 3D preview, and then pass it to Seedance 2.5 for video generation. For instance, to shoot a scene of two people walking towards each other with a side tracking shot, users can now set up movement paths and camera positions directly, instead of relying solely on text prompts and hoping for the best. Flowith also integrated Claude Code and Codex into the canvas: these tools can read images, videos, and documents on the canvas, write and run code in the cloud workspace, and return outputs like web pages and files to the canvas.

Relevant content

OpenAI rolls out its 24-hour AI assistant Dots to the $20 ChatGPT Plus tier, launches a limited beta test.

Beating AI Express News: OpenAI has begun rolling out its personal AI agent "dots" to selected ChatGPT Plus users. Previously, the feature was primarily available to high-tier subscribers including Pro (starting at $100/month), Business Premium, and Enterprise. Now, $20/month Plus users can also join the test. Launched at the end of September this year, dots’ key feature is proactively identifying valuable tasks without requiring new user instructions. It comes with its own cloud computer and browser, supports connections to apps such as email and calendar, and retains users’ long-term goals and work habits.

5 minutes ago

Anthropic's record-breaking IPO generates $20 billion in proceeds for Abu Dhabi's MGX.

Market news: Anthropic's record-breaking IPO has generated $20 billion in returns for Abu Dhabi-based investment group MGX.

5 minutes ago

OKX Adds Support for 8 Tokenized Stocks to Its Simple Earn and Flexible Loan Products

According to official news, OKX’s Simple Earn and Flexible Lending services now support tokenized stocks, with the first batch covering 8 assets: xCRCL, xMSTR, xSPCX, xMU, xQQQ, xGOOGL, xINTC, and xNVDA. Eligible users can use the designated tokenized stocks to subscribe to Simple Earn flexible products to earn interest, or borrow tokenized stocks via Flexible Lending to meet various trading strategies and fund management needs. Simple Earn flexible products have no fixed term and support flexible redemption, while Flexible Lending requires meeting corresponding collateral requirements and paying borrowing interest. For specific supported assets, interest rates, and borrowable limits, please refer to the product page.

5 minutes ago

U.S. House of Representatives releases the public export license list for Nvidia and AMD's H200 chips to China, with ByteDance and Alibaba's applications totaling nearly 2 million units.

Beating AI Insight News: U.S. House Foreign Affairs Committee Chairman Brian Mast has, under the Export Control Reform Act, released detailed information on export license applications for AI chips (primarily the H200 model) submitted by NVIDIA and AMD to China. Data shows the companies together applied for roughly 3.44 million chips, with the Bureau of Industry and Security (BIS) approving over 879,000 units. ByteDance led all applicants with 1.015 million requests, followed by Alibaba with 949,000, bringing the two firms’ total applications to nearly 2 million. Mihoyo surprisingly secured approval for around 57,800 units, outpacing Baidu and JD.com. Tencent, Didi, Kuaishou and other companies also appeared on the application roster. Most approved chips remain in an "undelivered" status. The documents further reveal some licenses cover exports to firms designated by the U.S. Department of Defense as Chinese military enterprises, and that NVIDIA once sought to facilitate sales to China via potential smuggling channels. This marks the most detailed public disclosure since the act’s implementation.

5 minutes ago

US stocks were broadly lower in pre-market trading, AI stocks extended their decline, and the sell-off in Anthropic may not be over yet.

According to data from the BIT Stock Platform, U.S. pre-market trading saw a broad decline in optical communication stocks: Coherent fell around 4%, Credo dropped about 4%, Lumentum down roughly 3%, Marvell Technology down ~3%, and Corning down ~4%. Semiconductor stocks also traded broadly lower: AMD down ~3%, Micron Technology ~3%, Seagate Technology ~3%, Intel ~3%, and Nvidia down ~1%. Anthropic, which completed its IPO yesterday, fell around 6% in pre-market trading, after rallying over 17% on its debut day. Earlier reports noted that Anthropic’s seven founders cashed out nearly $1.95 billion during the IPO. Crypto-related stocks tracked the broader market lower: CRCL down 2.13%, MSTR down 1.58%, and COIN down 1.94%.

5 minutes ago

Yann LeCun Teaches World Models to Review Their Actions: Are Robot Failures Due to Wrong Moves or Bad Luck?

Beating AI News Brief: AMI Labs, founded by Yann LeCun, and researchers from Brown University have published a new paper titled *Counterfactual World Models*. The work enables JEPA world models to, based on an action’s actual outcome, infer whether a different choice at the time would have prevented failure — a capability called "counterfactual reasoning". For example, a robot tasked with delivering cargo to a target is blown off course by a gust of wind. How can it tell if it chose the wrong direction, or if excessive wind caused the task failure? Standard world models can predict typical outcomes of different actions, but cannot determine the wind’s specific impact in this instance. The new method uses the gap between actual results and prior predictions to estimate external disturbances’ impact, then calculates how a different action would have unfolded. The process only requires calling the existing predictor twice, with no extra counterfactual model training needed. This lets robots retrospectively assess whether an alternative action would have completed the goal. In four physical simulation environments, the team reported their model identified task-completing alternative actions in 35%–69% of cases, versus 14%–23% for baseline methods. Even when baselines knew the simulation’s true physical rules, they still underperformed if unaware of the actual disturbances that occurred. That said, the method requires action and disturbance impacts to be roughly additive. If collisions occur, or motor failures cause uneven impacts on different commands, predictions may be inaccurate. The research remains limited to simulations, with no validation on real-world robot control yet.

5 minutes ago

Popular tokens

BitcoinEthereumHyperliquidSolanaTRONBNBTetherAaveXRPPepeFartcoinOndoJupiterUniswapBonkPendleEthenaArbitrumAvalancheLidoChainlinkPolygonDogecoinCardano