Anthropic前员工做「自我改进AI」，3个月估值从10亿涨到50亿美元

Beating AI Insight News Flash. Leiphone reports that over the past six months, multiple domestic large language model (LLM) vendors have begun redoing their pre-training processes, with all issues traced to data quality. One vendor spent nearly a year training a trillion-parameter model, only to have its performance outperformed by a small model with just around 1% of the parameter count. The team ultimately found the root cause lay in the training data: web spam, duplicate corpora, and low-quality annotations had been mixed into the training set, rendering even the largest models ineffective. The case comes from an anonymous source, and the specific company has not been disclosed. Tencent Hunyuan has publicly announced a full redo of its pre-training work. Since February this year, the team has rebuilt its pre-training and reinforcement learning infrastructure. Previous media reports revealed that the old version of Hunyuan had issues including ranking manipulation data being mixed into the training set and inconsistent annotation rules. The new team redefined data standards, cleaned up existing corpora, and Hy3 listed data quality and diversity as key improvement areas. Alibaba and Baidu are also stepping up data governance efforts. Alibaba’s Qwen3 uses the Qwen2.5 series models to clean documents, and has significantly supplemented synthetic math and code data. Baidu’s Ernie 4.5 has added deduplication, low-quality filtering, data mapping, and manual review processes. Neither company has publicly stated that they had to redo work due to dirty data, but both have invested more efforts in data processing for their new-generation models.

The U.S. Treasury is developing a plan that could reshape how short-term funding markets function: deploying a portion of cash held in the Treasury General Account (TGA) into the private repo market on a regular basis, rather than leaving all of it parked at the Federal Reserve. The proposal earned backing from multiple market participants at the New York Fed’s annual U.S. Treasury market conference this week. The TGA, which currently holds just under $1 trillion, serves as the Treasury’s day-to-day operating cash account. The Treasury first put forward this idea back in May, and this week’s discussions have advanced to focus on specific implementation methods. Market observers see this move as a potential new stabilizing tool for short-term funding markets. If the Treasury actively deploys part of its TGA cash into the repo market, it would add a direct channel for traditional fiscal cash flows to enter the money market. Market participants are thus particularly concerned about whether the operations will be predictable.

According to an official announcement, Binance's cross-margin trading now supports the ARB/USDT and ENA/USD1 trading pairs.

Ryan Myher, co-founder of Genius Terminal, announced in a post today that Genius.fun is set to launch its foundation page soon. He added, “I’ve been exploring various new mechanisms lately to better drive narratives on Genius.fun. There’s a really interesting feature that could help users reach their goal of accumulating stocks faster, which is especially exciting for our larger-cap meme coins.” On September 17, Genius Terminal—an on-chain trading terminal backed by YZi Labs and advised by CZ—officially launched its token issuance platform genius.fun on BNB Chain. The platform allows creators to transfer tokenized stocks such as bSTOCKS into a non-custodial foundation and split them, enabling subsequent attempts to exercise equity-related rights off-chain.

According to monitoring by Farside Investors, Bitcoin ETFs posted a net inflow of $714.7 million yesterday, marking their third consecutive day of net inflows exceeding $400 million. BlackRock’s IBIT recorded a net inflow of $350 million, while FBTC saw a net inflow of $257.4 million. Ethereum ETFs had a net inflow of $162.2 million yesterday, with ETHA registering a net inflow of $88.1 million and FETH notching $33.6 million in net inflows.

Huobi HTX Chief Analyst Cloud noted that Bitcoin bottomed out amid dual pressures of interest rate hikes and stalled legislative clarity, then rebounded rapidly, gaining roughly 16% in a week. On September 21, its intraday price rose to $87,307 (Huobi HTX spot price), marking a new high since January. This rally was driven by three factors: U.S. spot Bitcoin ETFs recorded a single-day net inflow of about $1 billion on September 21, the largest such daily inflow this year; short positions worth around $650 million were liquidated within 24 hours; and falling oil prices cooled inflation expectations. Unlike last week’s rebound dominated by passive covering, this week saw spot buying take the lead. Whether the rally can sustain depends on two points: whether ETF inflows can shift from a one-day pulse to a continuous trend, and whether leverage can remain at a non-overheated level. Currently, the funding rate is around 0.01%, sitting in the neutral range, but Bitcoin’s derivatives open interest has rebounded to over $61 billion. If market sentiment weakens, high leverage will amplify drawdowns. Market sentiment has entered the extreme greed zone, a historical precursor to short-term trend reversals. On the technical front, $87,500 is the immediate resistance level, while the first support zone ranges from $84,000 to $85,000. The trend has conditions to continue, but the phase of peak sentiment also brings the highest short-term volatility risk. Note: The content of this article is not investment advice, nor does it constitute an offer, solicitation, or recommendation for any investment product.