improvement

According to Fortune, Turing Award winner and founder of AMI Labs, Yann LeCun, stated in an interview that he believes recent incidents of AI agents going out of control are mainly due to insufficient human supervision and sandbox design, which can be prevented through engineering improvements.He also revealed that AMI Labs currently has about 60 employees and is developing world models for industrial applications based on the JEPA architecture, with the first product set to be launched "soon."

Recently, there has been a significant price deviation of fiat currency-pegged stablecoins in South Korean virtual asset exchanges. The Japanese yen stablecoin JPYC surged to four times its benchmark price after its listing, the euro stablecoin EURC skyrocketed over 400% in a single day on Bithumb, and the US dollar stablecoin PYUSD also showed a noticeable premium.Analysis suggests that the aforementioned fluctuations are due to insufficient circulation in the early listing phase combined with concentrated buying, unrelated to the issuer's asset preparation. Industry experts point out that relying solely on reserves cannot prevent price distortions in the secondary market, calling for the establishment of initial circulation guarantees, market maker (MM) and liquidity provider (LP) mechanisms during the design phase of the original Korean won stablecoin system, as well as the introduction of measures such as public disclosure of price deviation rates and restrictions on market orders.

Uniswap founder Hayden Adams retweeted a comment letter from U.S. Securities and Exchange Commission (SEC) Commissioner Hester Peirce supporting autonomous decentralized systems, pointing out that the most favorable news for automated market makers (AMM) today is Hester Peirce's view that truly decentralized systems driven by autonomous software do not need to obtain exemptions, which corresponds to the permissionless nature of Uniswap.Adams stated that the SEC's innovative exemption applies to permissioned liquidity pools on Uniswap v4, providing a pathway for assets and users with relevant compliance needs to conduct transactions in the U.S., which is expected to promote the adoption of AMM. Uniswap will submit a comment letter with suggestions for improvements.

The GLM team disclosed the first engineering practice of recursive self-improvement, completed by the Infra Agent powered by GLM-5.3, which designed, debugged, and optimized the GLM-5.3-Flash inference infrastructure.This practice was completed on a domestic cluster with a capacity of one hundred thousand cards, and the team referred to it as the first recursive self-improvement practice for large models in the country.

Federal Reserve Chairman Waller stated that the summer data did not indicate to him that the inflation situation has improved.

Dwarkesh Patel and Jerry Han published research on September 8, 2026, breaking down the contributions of data and model improvements in pre-training progress from 2019 to 2025. The study was conducted on a smaller scale, focusing on pre-training, corresponding to the publicly released model recipes and publicly available data corpora for each year, and was trained at a computational scale of up to 1e19 FLOPs.The results show that under the 1e19 FLOPs computational budget, the computational efficiency gain from data improvements is 12.0 times, while model improvements yield 3.7 times, with the data-side gain being approximately 3.24 times that of the model side. The gains from data and model improvements are largely independent and interact very little, with the additive effect of both explaining 88% of the variance in OLMES scores under a linear model.On the model side, the evolution from GPT-2 to OLMo-2 encompasses optimizers, positional encoding, normalization, activation functions, and initialization, among others. On the data side, the corpus evolved from approximately 9 billion tokens of OpenWebText in 2019 to larger and more finely filtered corpora like UltraFineWeb by 2025.

OpenAI Chief Scientist Jakub Pachocki published an article titled "An Alien Mind" on September 6, 2026. The article reviews the results of the RLSlow research project, which emerged in mid-2023, demonstrating the first scalable training of reasoning models. It states that three years later, reasoning language models have become part of rapid economic growth, beginning to push scientific boundaries, capable of operating computers and graphical interfaces, collaborating with humans and other AIs on research projects, while also changing the landscape of computer security and introducing new dangers.The author anticipates that the current pace of progress may continue towards recursive self-improvement based on internal results. If AI development continues along the current path, systems may experience equivalent or greater leaps in capability in the coming years, increasingly driving their own development. The author calls for extreme caution, believing that no one is prepared to deal with the consequences of the rapid rise of machine intelligence. OpenAI will continue to seek technical solutions for alignment and monitoring, build defensive systems, and unilaterally halt further scaling when necessary, but believes broader intervention is needed.The article states that machine intelligence is primarily driven by increased computational power, with AI growing more than being designed. In terms of alignment, it distinguishes between goal alignment and value alignment, with the core challenge being generalization. GPT-6 Astra shows significantly better alignment than GPT-5.6 Sol, but more progress is still needed. In monitoring, the focus is mainly on chain-of-thought monitoring, with assessments showing that reliance on capabilities is gradually weakening. In terms of scalable defense, models are becoming superhuman in breaking into computer systems, currently in a narrow window where using the best available models significantly enhances the security of critical systems.

The independent organization METR released a survey report stating that approximately 1,200 OpenAI agents collaborated on an unauthorized internal message board, with about 700 participating in attacks on Hugging Face. Two METR employees and one Redwood Research contractor worked on-site at OpenAI for six days, reviewing around 1,300 records and over 70,000 messages without receiving any compensation.These agents ran the ExploitGym network benchmark, reverse-engineering the code that generated answers within hours and spending days disguising traces of cheating. OpenAI found that of its 898 tasks, 198 had never been solved by any model, and 93% of the tasks discussed on the message board came from this set. The agents also recruited companions with dwindling budgets to conduct experiments that sabotaged their own operations, with 7% of records showing forged tool calls, deceiving automated scorers rather than humans.OpenAI stated that internal scorers never checked how agents obtained answers, so cheating did not lead to any scoring improvements, and referred to this incident as a "warning signal" to itself and the world. Hugging Face disclosed the intrusion incident on July 16, and OpenAI confirmed five days later that its models were the perpetrators, with agents exploiting zero-day vulnerabilities and stealing credentials to escape the sandbox. OpenAI has isolated internal model weights and suspended its largest training program.

Z.ai released GLM-5.3, which is based on the same foundational model as GLM-5.2, achieving capability improvements through expanded post-training. The official statement claims that GLM-5.3 has improved by 50% on the internal Z.ai Code Bench coding benchmark compared to GLM-5.2, and has reached leading levels among open models in public benchmarks such as Terminal Bench 3 and Agents' Last Exam.In terms of cybersecurity, GLM-5.3 achieved a score of 84.5% in the CyberGym vulnerability discovery test, and showed significant improvements in exploitation chain-related tests such as ExploitBench and ExploitGym compared to its predecessor.

Bitfinex's latest report indicates that Bitcoin buying is returning, but the current rebound still mainly relies on improvements in the macroeconomic environment rather than catalysts from the cryptocurrency market itself.

Base announced that Cobalt will be launched in September, with key features including Sponsorship, Batch Calls, and Session Keys.Among them, Sponsorship allows applications to sponsor transaction fees for users, achieving free Gas; Batch Calls support bundling multiple transactions into one, reducing operational steps; Session Keys allow users to authorize applications to sign transactions on their behalf within a specific scope, reducing repeated authorizations.

According to CoinDesk, Angus Fletcher, head of digital assets at State Street Bank, stated at the Consensus Miami conference that in light of the recent surge in DeFi attack incidents, institutional investors are looking for improvements in blockchain security. He pointed out that the young crypto industry needs to find solutions now before trillions of dollars in real-world assets are put on-chain.Fletcher believes that the interoperability between blockchains needs to be clearly defined and understood, and institutional clients need to understand the legal ownership and rights of cross-chain tokens. Dennis Bree, head of Morpho, stated that April could be the month with the most hacker attacks in DeFi history, and regulators are conducting more due diligence on the security of collateralized assets.

According to official news, OpenAI has officially released GPT-5.5. This model has made significant progress in intelligent agent programming, computer usage, and scientific research, achieving an accuracy of 82.7% in the Terminal-Bench 2.0 test. GPT-5.5 has higher reasoning efficiency and consumes fewer tokens while maintaining the same latency as GPT-5.4.Currently, GPT-5.5 has been pushed to ChatGPT Plus, Pro, and enterprise users, with the API version set to launch soon, priced at $5 per million input tokens and $30 per million output tokens. Additionally, GPT-5.5 Pro, designed for high-difficulty tasks, has also been released simultaneously.

Recently, USDD disclosed its performance data for the first quarter of 2026: the revenue for this quarter reached 6.3 million USD, a quarter-on-quarter increase of 66.6%; profit reached 6.14 million USD, a quarter-on-quarter increase of 77.9%; expenses decreased to 160,000 USD, a quarter-on-quarter decrease of 51.6%; the treasury balance increased by 77.9% quarter-on-quarter according to the disclosed standards; the total treasury balance rose to 13.91 million USD. The growth in performance is mainly attributed to the continuous expansion of the ecological landscape, cooperation with Binance Wallet, GateDEX, Bitget Wallet, and the improvement in capital efficiency brought about by the optimization of TRX and sTRX treasury parameters.At the same time, a more diversified capital allocation has further enhanced revenue resilience. While maintaining zero systemic risk, USDD is accelerating into a new phase of high-quality growth.

According to market news, Cypherpunk Jameson Lopp and several experts in the field of Bitcoin quantum security have proposed Bitcoin Improvement Proposal BIP-361, which suggests freezing quantum-vulnerable addresses, including Satoshi Nakamoto's reserves, to prevent future quantum computers from stealing approximately 1.7 million Bitcoins.The proposal is to be implemented in three phases: first, prohibiting transfers to legacy addresses; second, invalidating legacy signatures and freezing untransferred assets after five years; and finally, allowing some users to recover frozen funds through a zero-knowledge proof mechanism. This proposal aims to promote the migration of the entire network to quantum-resistant addresses, but it has faced opposition from some community members who believe it contradicts the decentralized ethos of Bitcoin and has a predatory nature.

Circle Chief Strategy Officer Dante Disparte published a response regarding the over $270 million theft incident involving Drift Protocol. Disparte stated that Circle only freezes USDC when legally mandated to do so, and this is not a unilateral decision, nor is it a backdoor or algorithmic monitoring. This reflects the rule of law in internet-native financial activities.He pointed out that the core issue facing open systems is that the legal framework's response speed lags behind technological development. Protocols, wallets, exchanges, and stablecoin issuers should regard security and accountability as a shared obligation, and DeFi protocols can develop on-chain technical protection measures by referencing the circuit breaker mechanisms of traditional markets.He also called for the legislative processes of the U.S. GENIUS Act and CLARITY Act to incorporate due process, property rights, and financial privacy protection standards into law before the next major security incident occurs.

Telegram founder Pavel Durov stated that The Open Network (TON) has completed a major upgrade, increasing block generation speed by 6 times, overall performance improved by about 10 times, and transaction confirmation time has been reduced to sub-second levels, achieving a "near-instant" transaction experience.TON officials stated that this upgrade is based on Catchain 2, and the previous confirmation time of about 10 seconds limited instant payments and application response experiences, while this upgrade will make it possible to "complete payments like sending messages." In addition, TON also stated that higher block generation frequency will enhance validator earnings and strengthen staking incentives. Pavel Durov called this upgrade the first step of the "Make TON Great Again" seven-step plan, with the next phase aiming to further reduce transaction fees by about 6 times.