ode

According to CoinDesk, the Internal Revenue Service (IRS) has issued a warning that a sophisticated email phishing campaign targeting U.S. cryptocurrency holders is spreading. Attackers are impersonating official tax letters to lure users into scanning malicious QR codes to steal cryptocurrency wallet credentials and private keys.It is reported that attackers are sending paper letters impersonating the IRS, creating a sense of urgency under the guise of "tax compliance" and "account verification," and including QR codes in the letters. Once users scan the code, they may be directed to a counterfeit website, leading to the leakage of wallet login information, recovery phrases, or private keys, resulting in the theft of digital assets.The IRS reminds taxpayers that official agencies will not request users to provide cryptocurrency wallet private keys, recovery phrases, or perform similar "wallet verification" operations through unofficial channels. Cryptocurrency holders should be vigilant against any suspicious emails and letters that request scanning QR codes, connecting wallets, or submitting sensitive information.As the number of cryptocurrency asset holders grows, social engineering attacks targeting digital wallets continue to increase, and regulatory and security agencies are strengthening warnings against related fraudulent activities.

The B.AI platform announces that Xiaomi's self-developed MiMo series models are officially online with API services, with the first batch opening up the MiMo-V2.5 and MiMo-V2.5-Pro models.Among them, MiMo-V2.5 is a 310B sparse MoE native multimodal model that supports multimodal input of text, images, audio, and video, as well as 1M ultra-long context, specifically designed for multimodal agents and general programming scenarios; MiMo-V2.5-Pro reaches a parameter scale of 1.02T sparse MoE, positioned as a flagship text model, focusing on complex reasoning and long-cycle software engineering, achieving a high score of 78.9 in the SWE-Bench Verified benchmark test.Starting today, developers can call the two models through the B.AI API, supporting both official group access and self-selected service provider dual modes, with discounts as low as 40%.

DeepSeek announced the first batch of Agent scores for V4-Flash-Vision-Exp. In the evaluation of 4 multimodal Agents, it achieved 2 wins and 2 losses against Opus 4.8: Agents' Last Exam 27.3 vs 25.7, ZeroBench 35.0 vs 34.0; ApexBench 36.5 vs 39.4, Chartography 64.3 vs 65.0.Compared to the pure text version V4-Flash-0731, the visual version improved from 26.2 to 36.5 in ApexBench, and from 25.2 to 27.3 in Agents' Last Exam. The official note states that the pure text version ignores multimodal content, thus mainly reflecting the capabilities after adding visual input. After adding visual input, the text Agent's capabilities did not significantly decline, with 6 out of 7 text evaluations higher than V4-Flash-0731, such as DeepSWE rising from 54.4 to 59.3 (surpassing Opus 4.8's 58.0), and Toolathlon at 75.9 nearly matching Opus 4.8's 76.2.These results come from DeepSeek's official self-testing, not from a third-party independent ranking; the public Code Agent text task used DeepSeek Harness in minimalist mode, with the inference setting at max.

DeepSeek officially launches the new visual model deepseek-v4-flash-vision-exp. The official API documentation has listed it alongside V4 Flash and V4 Pro, allowing developers to directly input images through the DeepSeek API. The model supports a context of 1 million tokens, with a maximum output of 384,000 tokens, and also supports JSON Output, Tool Calls, Responses API, and Anthropic API.The pricing is directly aligned with V4 Flash. For every million tokens input, when there is a cache miss, the peak period is 3 yuan, and the idle period is 1.5 yuan; cache hits only cost 0.1 yuan and 0.05 yuan respectively. For every million tokens output, the peak period is 9 yuan, and the idle period is 4.5 yuan. Compared to V4 Pro, the same tier pricing is only one-third.Images are not charged separately per image. DeepSeek will convert images into tokens based on their size and bill them together with text tokens. The peak periods are from 9:00 to 12:00 and 14:00 to 18:00 Beijing time, with prices halved during other times.

According to X user MaxForAI, signs related to the DeepSeek visual model have emerged. At 3 PM today, the name deepseek-v4-flash-vision-exp appeared in the code related to DeepSeek Harness, and the latest version of Harness has begun to incorporate native image request adaptation. However, the model will not be included in the official model directory until the model endpoint is ready.Currently, it seems that DeepSeek is testing the visual version of V4 Flash, and the public API has not yet been officially opened. Some users have directly requested the DeepSeek API using the model ID, but it still returns errors, and the official documentation has not listed this model yet. The adaptation path for the client and Harness side has basically been laid out, and DeepSeek V4 Flash is expected to soon have visual capabilities.

Recently, multiple users discovered an image model named luna-lisa-alpha appearing on the LMArena testing platform, suspected to be a new generation image generation model that OpenAI is testing. Although not officially confirmed, there is speculation that it could be an early version of GPT Image 2.5 or the next generation of GPT Image.From the testing feedback, the new model shows significant improvements in character consistency, image realism, text generation capability, and noise control. Some testers compared it with models like GPT Image 2.0, Nano Banana Pro, and Grok Imagine, believing that luna-lisa-alpha can better maintain the appearance, clothing, and style of the same character across multiple generations. However, some testers pointed out that there are still issues with repeated textures in complex texture areas. On the same day, OpenAI announced that the GPT Image 2 API preview supports transparent background generation, which can be used for product images, web design, and other scenarios. The exposure of multiple checkpoint names has also led the community to speculate that OpenAI may be conducting A/B testing. Whether this model will be officially released remains uncertain, and we need to wait for official information from OpenAI.

On August 21, OpenAI responded to the differences in Codex usage limits, stating that it would not unilaterally change the relevant rules without communicating with the community first. An official investigation found that affected users often converted subscriptions into API traffic through sub2api and redistributed it, which would be flagged by the fraud prevention system.OpenAI pointed out that using subscriptions normally through Sign in With ChatGPT on official or OSS clients (such as Pi, OpenCode) poses no issues at all.

According to Bloomberg, Anthropic plans to adjust its data retention policy for its advanced artificial intelligence models, giving enterprise customers greater control over their data. The company expects to launch a new security system later this year, where enterprise customers will still need to retain data for 30 days but can choose to do so on their own cloud computing infrastructure instead of storing it on Anthropic's servers. Sources say that the system has been in preparation for several months and has been developed in coordination with over 100 clients from strictly regulated industries.In June, Anthropic announced that it would retain all customer data for 30 days for its more powerful Mythos, Fable models, and future cutting-edge products to identify and prevent new types of cyberattacks using its technology, while stating that it would not use the data for training. The company recently acknowledged in a report that this policy may not be welcomed by customers accustomed to zero retention and could impact business. Anthropic declined to comment on this. Competitor OpenAI has also recently begun testing a solution called "Private Secure Processing" to avoid retaining customer data and plans to roll it out more broadly in September.

According to Forbes, Anthropic, Blackstone, and Goldman Sachs have jointly established an AI-native enterprise service company called "Ode," with an investment of $1.5 billion.Ode is positioned as an AI implementation consulting firm, with core services including identifying business processes suitable for AI transformation, deploying AI engineers, developing customized applications and agents, automating task processes, and assisting companies with regulatory compliance issues. The company will focus on helping client enterprises implement the Claude model under Anthropic.

The open automated research challenge better.codes, created by the Ethereum Foundation's formal verification team in collaboration with Yukon and zkSecurity, is now live. This platform formalizes the self-contained problems from the Proximity Prize in Lean and places the machine-checked reliability bounds of koalaIRS12 on a public leaderboard, allowing anyone to contribute to improvements, advancing hash-based SNARK and post-quantum Ethereum-related security benchmarks.Solvers can bring their own AI agents to prove higher lower bounds of reliability for this Reed-Solomon proximity problem, moving towards a fixed target of 128 bits. The Lean kernel verifies each submission, and the promoted proofs will enhance the public bounds, with new lemmas, proof techniques, and impossibility results being synchronized upstream for reuse by all participants. Most hash SNARKs in production environments rely on related proximity gaps and related conclusion agreements, while the currently provable results remain below the benchmarks trusted by researchers; this challenge aims to narrow this gap in an open, incremental, and verifiable manner.koalaIRS12 originates from related papers and is end-to-end formalized in ArkLib. Participants can log in via GitHub and clone the challenge repository, submitting under fixed theorem statements and verification frameworks; results confirmed by the comparator and Lean kernel are recorded in a public repository, noting the solver and the model used. Today's launch features the reliability challenge that raises the proven lower bound of koalaIRS12 to 128 bits, with more problems potentially added later, subject to project terms.

XCAD Network officially announced the cessation of operations. The project stated that there have been no updates since April, and there are currently no remaining funds. Continuing to stay on the exchange is a losing battle, and related products and services will no longer be maintained.The official statement mentioned that the company has been operating for five years, making it one of the few projects that have persisted for a long time without being listed on a tier-one exchange. The fan tokens failed to gain traction after retail investors exited, and the team attempted to pivot towards AI influencers but was unsuccessful and is now unable to continue. The team is considering making the code public for those in need to run it themselves, emphasizing that this is not a continuation of the company, nor a restart.

Glassnode stated that high-conviction buyers have been continuously accumulating BTC since January of this year, a similar accumulation behavior was observed in 2022. It is worth noting that in this cycle, the proportion of BTC purchased by these buyers has already surpassed that of the previous cycle.

Google (GOOG.O) announced that the downloads of its Gemma model have surpassed 1 billion.

Bitcoin News posted on platform X that the IRS has reminded cryptocurrency holders that they may recently receive fake official letters mailed by scammers, directing recipients to visit a fraudulent "Digital Asset Compliance Portal."The letter contains a QR code, which, when scanned, redirects to a fake IRS website that may request personal information, wallet information, exchange account credentials, or other sensitive data.The IRS stated that it does not operate a platform called "Digital Asset Compliance Portal" and has not sent out these letters. Reports indicate that Coinbase and DarkTower have traced the infrastructure behind this activity to a recently registered domain hosted in Romania. If you receive such letters, do not scan the QR code or provide any information.

According to X user MaxForAI, Tencent has begun gray testing its brand new flagship model, Hunyuan Hy4. Some users have discovered that the model selection list in the Tencent Yuanbao App now includes Hy4, labeled as an expert-level model, ranking above Hy3 and DeepSeek. The official account @TencentHunyuan also shows this model option at the relevant entry.Last week, Tencent confirmed in its Q2 financial report that the larger parameter-scale Hy4 will be launched soon, further enhancing model performance and multimodal capabilities. Currently, Tencent has not officially released Hy4, and the outside world cannot confirm whether this is a small-scale gray test or an early opening of the entry, but based on the progress, the model's launch seems imminent.

Glassnode stated that the last time Bitcoin saw such a significant daily closing price increase was in February of this year, primarily as a rebound from the previous day's 14% drop.This increase is not due to a technical rebound after a crash; relative to its 30-day volatility, this increase reached 5.8 standard deviations, marking the largest upward volatility since October 2023.

According to Gate market data, Moderna (MRNA.O) fell over 13% in pre-market trading.

According to Axios, payment giant Stripe stated in a letter to investors that January 1 marks "the beginning of a singularity," viewing it as a significant turning point in a long-term trend, and believes that maintaining a private structure is best suited for this critical moment, with the IPO likely to remain on hold. The company reported a 41% year-on-year revenue growth in the first half of the year and a 43% increase in free cash flow; 88% of the companies in Forbes AI 50 (including OpenAI and Anthropic) are building on its platform, with revenue from AI and crypto companies more than doubling year-on-year.Stripe also confirmed the acquisition of the AI routing platform OpenRouter, with the transaction amount not publicly disclosed; Axios learned that the amount exceeds $8 billion and is primarily paid in stock. Stripe stated that remaining private helps advance mergers and acquisitions and long-term investments without diluting shareholders, with its equity count now lower than three years ago, and a compound annual return of about 31% since the D round. The company stated that the total payment volume on its platform is expected to reach $1.9 trillion by 2025, a year-on-year growth of 34%; in February this year, the employee stock purchase valuation was approximately $159 billion. There are also reports that Stripe is in discussions with Advent International to acquire PayPal for about $53 billion.