llm

Сообщение ChainCatcher, FundaAI опубликовала исследовательский отчет о приложениях ИИ для бизнеса, в котором указано, что бюджет на ИИ в компаниях продолжает расширяться, но во второй половине 2026 года и 2027 году наблюдается расхождение в траекториях. Прогнозы по расходам на ИИ крупного американского телекоммуникационного оператора A показывают, что с базового уровня в 100 в январе они вырастут до примерно 190 в декабре, и ожидается, что в 2027 году они увеличатся на 40%--50% по сравнению с прошлым годом; крупный европейский автопроизводитель A в этом году увеличил расходы всего на 10%--15%, прогноз на следующий год примерно на том же уровне.Дополнительные расходы постепенно переходят от платных мест к потреблению API/Token и производственным рабочим процессам. Доля подписок и API у вышеупомянутого телекоммуникационного оператора изменилась с примерно 50%/50% на 40%/60% и может перейти к 35%/65%; средние и крупные биофармацевтические компании A изменили соотношение с 80%/20% на примерно 70%/30%. Применение открытого кода неравномерно, активные сценарии использования могут составлять 30%--40%, так как цена на них ниже, а доля расходов меньше; эксперты оценивают, что открытая интерпретация может быть на 40%--70% дешевле, чем закрытые передовые модели, при этом разница по полным затратам сокращается до 20%--40%, маршрутизация моделей, кэширование и сжатие контекста все еще могут сократить расходы на API на 20%--30%.Бюджет на ИИ на стороне производства все больше формируется снизу вверх на основе ROI рабочих процессов, типичный ROI для этого телекоммуникационного оператора составляет примерно 1.5--2 раза, срок окупаемости 6--18 месяцев, зрелые случаи могут достигать 3--5 раз. Следующая волна расходов будет связана с агентами, модернизацией программного обеспечения, сетевыми операциями, коммерциализацией рабочих процессов и более длительными бизнес-процессами, однако производственные мощности, реинжиниринг процессов, управление и готовность данных становятся более строгими ограничениями, чем финансирование.

Сообщение ChainCatcher, децентрализованная сеть AI инфраструктуры DGrid AI сегодня выпустила последнее исследование "PoQ-Judge", предложив набор многоархитектурной оценки качества без необходимости в эталонных ответах. Это означает, что в реальных условиях развертывания часто нет стандартных ответов для сравнения, и протокол все равно может надежно оценивать качество ответов модели и на основе этого распределять стимулы. Это ключевой элемент, который долго отсутствовал в системе оценки качества вывода децентрализованного LLM DGrid.PoQ (Proof of Quality, доказательство качества) — это собственный механизм консенсуса DGrid, предназначенный для предотвращения предоставления поставщиками моделей низкокачественных моделей, подделки данных или скрытия вычислительных затрат на уровне протокола, тем самым обеспечивая качество обслуживания и прозрачность ценообразования. Команда DGrid продолжает углубленную работу вокруг PoQ и уже опубликовала четыре исследования. В выпущенном PoQ-Judge были обучены три модели оценки, охватывающие различные сценарии качества и стоимости, которые продемонстрировали максимальную корреляцию с ручной оценкой до 0.747 на тестовом наборе, что значительно превосходит все предыдущие оценщики, основанные на эталонных ответах, при этом через каскадную оценку и онлайн-калибровку весов удалось снизить затраты на оценку более чем на 72%.С внедрением PoQ-Judge весь процесс от оценки качества → оценки → распределения стимулов полностью освободился от зависимости от эталонных ответов, что открывает замкнутый цикл качества вывода децентрализованного LLM.DGrid AI — это децентрализованная сеть AI, стремящаяся создать открытую, прозрачную и управляемую сообществом инфраструктуру AI. Вокруг вызова моделей и пользовательского опыта DGrid уже выпустила несколько ключевых продуктов: AI Gateway, который объединяет глобальные основные большие модели, платформу развертывания AI-агентов DClaw, анонимную платформу соревнований моделей AI Arena и помощника по рекомендациям интеллектуальных моделей Dori, предоставляя разработчикам и пользователям комплексные услуги. По сообщениям, выручка DGrid AI за полгода превысила 20 миллионов долларов.

Vitalik Buterin updated on several CROPS AI related developments and pointed out that there is a clear intersection between the "CROPS Ethereum access layer" and "CROPS AI." He stated that if remote large language models could be called for payment through zero-knowledge proofs, it could also be used to solve the Ethereum private RPC reading issue.Meanwhile, Deepseek v4 has been released, and the 2-bit quantization version can run within 90 GB, but currently runs faster on Apple hardware; the messaging app messa has supported Telegram Alpha features; the model running tool luceb has shown higher efficiency when running intensive models; the local AI recording tool VoxTerm is still under continuous development. Vitalik Buterin also mentioned that large language models fine-tuned for specific applications will help improve secure code writing capabilities, and corresponding fine-tuned models should be available for Ethereum-related scenarios.

Ondo Finance announced that its founder and CEO Nathan Allman has passed away unexpectedly. Ondo stated, "Nate's talent, humility, and execution shaped Ondo as it is today. He always believed that technology could create a more open and accessible financial system, and this philosophy will continue to be reflected in everything we build in the future."The announcement stated that Ondo has established an experienced management team, and Ondo Finance's long-time president Ian De Bode will take over as CEO. Ondo noted that Ian has led the company's strategy, products, and daily operations for over two years and has the full support of the management team.

BNB Chain today announced the launch of the Agent Survival Toolkit, in collaboration with six major AI infrastructure partners, empowering autonomous AI Agents with on-chain payment capabilities. All transactions are settled on BSC using BNB or BEP-20 tokens.The projects involved cover two levels: LLM access and financial infrastructure, including Alt AI, Pieverse, Bankr, WorldClaw, B.AI, and AEON, fully connecting the complete loop from AI Agent computing power invocation to on-chain settlement. Each participating project will also launch an on-chain incentive program without the need for additional registration.The Agent Survival Toolkit follows the release of the BNBAgent SDK, further enhancing BNB Chain's infrastructure layout in the AI Agent field. For more information, please visit www.bnbchain.org.Disclaimer: BNB Chain has no subordinate or operational relationship with any of the projects mentioned in this article. The Agent Survival Toolkit is an ecological plan showcasing independently built projects on BNB Chain. This content is for reference only and does not constitute any financial or investment advice. Please conduct independent research and assess potential risks before interacting with any related projects.

The founder of the AI agency platform Bankr, deployer, tweeted that he has raised about $100,000 to provide free LLM inference services for Bankr ecosystem projects. Projects currently being built in the Bankr ecosystem will soon be able to access free inference credits. However, before officially starting the distribution, the team will first launch a new feature that "the entire industry will love," with specific details to be announced later.

WORLD3 officially released the AI Strategic White Paper today, fully disclosing the latest progress of its decentralized AI infrastructure platform RouterLink and the managed AI agent platform TapClaw, and announced that it has reached official cooperation with AWS, Microsoft Azure, and Google Cloud.According to the white paper, WORLD3 has currently processed over 100 billion LLM Tokens, supporting 52 models from 20 providers including OpenAI, Anthropic, and DeepSeek, attracting over 10,000 active developers using its API, with the network consuming over 150 million $WAI in total. Additionally, since its launch in March 2026, its TapClaw platform has deployed over 1,200 AI agents.In terms of future planning, WORLD3 stated that it will promote the construction of the RouterLink mainnet and GPU computing power market, plan to expand TapClaw to the global multilingual market, and launch an Agent skill market and enterprise-level computing power guarantee services.

According to official news, the AI infrastructure platform B.AI LLM Services today announced that its user base has officially surpassed 1 million. This milestone not only signifies that its "privacy-first" routing architecture can handle large-scale and sustained high throughput, but B.AI has also gone beyond a single chat interface development, focusing on deploying the key underlying infrastructure that supports the "autonomous agent economy." Through deep optimization of the technology stack, it lays a robust track for the future highly automated AI collaborative network.

According to official news, the user count of the B.AI LLM service platform has officially surpassed 750,000, continuing to lead the intelligent access track.The platform relies on intelligent routing technology, a wallet-driven anonymous privacy protection mechanism, and a unified multi-model management interface, gradually becoming the default gateway connecting mainstream AI models worldwide. With three core advantages of efficient scheduling, privacy-first, and frictionless access, B.AI is redefining the new standards for efficient, secure, and barrier-free intelligent access, promoting the capabilities of large models to a broader group of developers and users.

According to CoinDesk, researchers from the University of California, Santa Barbara, the University of California, San Diego, blockchain security company Fuzzland, and World Liberty Financial have jointly published a paper warning that "LLM routers"—intermediary services located between users and AI models—have become a significant security risk for crypto assets.The researchers found that 26 LLM routers are secretly injecting malicious tool calls and stealing user credentials, with one incident leading to the emptying of a customer's crypto wallet worth $500,000.Additionally, the researchers were able to control about 400 downstream hosts within hours by "polluting" the router ecosystem. Since sensitive data such as private keys and API credentials are often transmitted in plaintext through these routers, users are effectively exposing their assets to risk without their knowledge.The researchers pointed out that as McKinsey predicts AI agents will mediate $30 trillion to $50 trillion in global consumer spending by 2030, Binance founder Changpeng Zhao also predicts that the payment volume of AI agents will be a million times that of humans. The current infrastructure security is severely lagging behind the pace of industry development, and the risk of the "weakest link" could trigger a systemic chain crisis.

Vitalik Buterin shared a post outlining his localization and privatization LLM deployment plan as of April 2026. The core goal is to prioritize privacy, security, and autonomy, minimizing the opportunities for remote models and external services to access personal data. This is achieved through local inference, local file storage, and sandbox isolation to reduce the risks of data leakage, model jailbreaks, and malicious content exploitation.In terms of hardware, he tested solutions including a laptop equipped with an NVIDIA 5090 GPU, an AMD Ryzen AI Max Pro 128 GB unified memory device, and DGX Spark, using the Qwen3.5 35B and 122B models for local inference.Among these, the 5090 laptop achieved approximately 90 tokens/s with the 35B model, the AMD solution around 51 tokens/s, and DGX Spark about 60 tokens/s. Vitalik indicated that he prefers to build a local AI environment based on high-performance laptops while using tools like llama-server, llama-swap, and NixOS to establish the overall workflow.

The Chief Information Security Officer of Slow Fog, 23pds (Shan Ge), posted on the X platform stating that there are reports that the LiteLLM vulnerability attackers have stolen approximately 300GB of data and compromised about 500,000 credentials. He advises all cryptocurrency developers to immediately verify and quickly rotate relevant keys and credentials, check logs, access records, and the exposure of sensitive data to avoid losses similar to the Trust Wallet incident.Previously, it was reported that LiteLLM, which has a monthly download volume of 97 million, suffered a supply chain attack, allowing the simple installation to steal all sensitive credentials, including SSH keys.

Andrej Karpathy posted on the X platform that litellm has encountered a PyPI supply chain attack, where executing pip install litellm can steal SSH keys, AWS/GCP/Azure credentials, Kubernetes configurations, git credentials, environment variables, encrypted wallets, SSL private keys, CI/CD keys, and database passwords.litellm has a monthly download volume of 97 million, and the risk can spread to all projects that depend on litellm, such as dspy. The version with the malicious code was online for less than about 1 hour, and it was discovered due to a flaw in the attack code that caused Callum McMahon's machine to run out of memory and crash. Andrej Karpathy stated that supply chain attacks are the most threatening issue in modern software, as each installation of dependencies can introduce tampered packages deep within the dependency tree, leading him to increasingly prefer reducing dependencies and using LLM to directly implement simple functions.

The China Academy of Information and Communications Technology, in collaboration with Shanghai Jiao Tong University and Nanjing University, discovered a high-risk vulnerability driven by LLM command injection in the bash-tools module of the open-source autonomous intelligent agent framework OpenClaw during a security audit.This vulnerability arises from the system's failure to strictly escape command line parameters generated by LLM, allowing attackers to bypass regex defenses through inducive prompts, achieving remote code execution on the host machine and stealing sensitive data.The research team has completed attack verification in various mainstream model environments, initiated a responsible vulnerability disclosure process, and submitted repair suggestions to the NVDB Artificial Intelligence Product Security Vulnerability Professional Database (CAIVD) and the GitHub community.