Виталик Бутерин раскритиковал погоню за «чрезмерно агентными» ИИ-моделями

cryptonews.ruОпубліковано о 2025-04-10Востаннє оновлено о 2025-08-11

  • Виталик Бутерин отметил, что агентские ИИ-модели — не лучший путь для расширения ИИ.
  • Сооснователь Ethereum считает, что искусственный интеллект должен больше слушать человека, а не действовать автономно.
  • Кроме того, он отметил роль открытых весов и редактирования, а также развитие BCI для будущего ИИ.

Сооснователь Ethereum Виталик Бутерин выразил разочарование тем, что развитие искусственного интеллекта (ИИ) слишком сосредоточено на создании «максимально агентных систем», вместо того чтобы обеспечить больше возможностей для человеческого вмешательства.

Echoing something @karpathy recently said, it does frustrate me how a lot of AI development is trying to be as "agentic" as possible, when actually creating *more* paths for human input both creates a better output (now for quite a while going forward) and is better for safety.

— vitalik.eth (@VitalikButerin) August 11, 2025

По его мнению, такой подход позволяет получать лучшие результаты и повышает безопасность.

«Меня раздражает, что многие разработки в сфере ИИ пытаются быть как можно более “агентными”, хотя на самом деле создание большего количества путей для человеческого вклада и сейчас, и в будущем дает лучший результат и является более безопасным», — заявил Бутерин.

Он отметил, что его больше захватывают ИИ-модели с открытыми весами и развитыми возможностями редактирования, чем те, которые ориентированы исключительно на создание контента с нуля.

В среднесрочной перспективе Бутерин видит развитие интерфейсов «мозг-компьютер» (BCI), которые при генерации контента в реальном времени будут считывать эмоциональную реакцию пользователя на каждый элемент и соответственно корректировать результат.

Между тем идею развития более эффективного взаимодействия ИИ и человека поддержал и основатель и CEO OpenServAI Тим, отметив:

«Новые архитектуры, которые превращают намерения пользователя в ценные действия, — единственный возможный путь развития, по моему мнению. Это то, над чем мы работаем в OpenServAI, применяя непосредственно к агентам для координации сложных задач».

Напомним, что во время онлайн-ивента Incrypted Online Marathon 2025 команда имела возможность пообщаться с CISO по брокерским и финансовым направлениям компании Robinhood Дэвидом Шведом, который поделился своим видением агентных ИИ. Об этом можно почитать в отдельном материале редакции:

Пов'язані матеріали

U.S. Treasury Steps Up Long-Term Bond Buybacks, Why Are Gold and Bitcoin Rising?

The U.S. Treasury's announcement to double its long-term Treasury bond repurchase program has fueled market fears over potential dollar depreciation, leading to significant rallies in both Bitcoin and gold. Analysts interpret the larger buybacks as a form of easing, which could exacerbate inflation and erode the dollar's purchasing power, driving investors toward perceived hard assets like Bitcoin and gold. Bitcoin saw a sharp surge, further amplified by short-covering and positive regulatory developments. Gold also broke above a key technical level amid dollar weakness. However, the bond market's reaction was muted, with long-term yields remaining elevated. Skepticism persists over whether these repurchases can address the fundamental pressures driving the bond sell-off, especially against a backdrop of soaring U.S. government debt and interest costs.

华尔街日报1 год тому

U.S. Treasury Steps Up Long-Term Bond Buybacks, Why Are Gold and Bitcoin Rising?

华尔街日报1 год тому

Mysterious "Ox Alpha" Large Model Goes Viral with Limited-Time Free Access

A mysterious anonymous AI model named "Ox Alpha," nicknamed "Cow is Coming" by Chinese netizens, has appeared on OpenRouter, sparking widespread speculation. The model offers a 1 million token context, supports text, image, and video inputs, can call tools, and is currently free. Its standout feature is strong coding ability. Initial tests on the DeepSWE benchmark, which evaluates real-world software engineering tasks, showed an 80% pass rate on a subset of tasks, reportedly nearing top-tier code models. However, follow-up tests yielded a 63% score, with variations attributed to different task sets and configurations. The model's true developer is a major topic of debate. The prevailing theory points to Zhipu AI's unreleased GLM-5.3 Flash or its multimodal variant. Evidence cited includes identical visual token consumption patterns with GLM-5V-Turbo for videos, a consistent offset in text token counts compared to GLM-5.3, and similar behavioral traits like refusing audio processing. Zhipu has a precedent of anonymous testing. Simultaneously, another anonymous model, "korrine," appeared on Code Arena, with guesses ranging from Moonshot's Kimi K3.1 to models from Qwen or MiMo, adding to the industry's guessing game. This trend of anonymous "undercover" testing allows for unbiased performance evaluation in platforms like Arena and provides real-world, high-pressure testing through tools like OpenRouter before official release. It also serves as an effective marketing tactic, prolonging discussion through suspense. If Ox Alpha is indeed a "Flash" model, its performance raises expectations for the full-scale version's potential.

marsbit3 год тому

Mysterious "Ox Alpha" Large Model Goes Viral with Limited-Time Free Access

marsbit3 год тому

Bitwise CIO Matt Hougan: If You Have 0% Crypto Allocation Now, It's Equivalent to Actively Shorting the Market Outlook

Summary: In a podcast interview, Bitwise CIO Matt Hougan argues that a 0% allocation to crypto is not a neutral stance but an active bet against its future, despite the market being down 50% from its highs. He observes that major Wall Street institutions like BlackRock, Wells Fargo, UBS, and Morgan Stanley are proceeding with crypto and tokenization initiatives regardless of short-term price or regulatory delays like the CLARITY Act, which he sees as perpetually stalled. Hougan states that Bitcoin's resilience to recent bad news (AI stock volatility, large BTC sales, fading regulatory odds) signals the bear market is ending, as "bear markets die in apathy." He recommends a 5% crypto allocation for most portfolios, calling it a "magic number" that boosts returns without significantly increasing overall volatility. For absolute returns, he favors buying now over dollar-cost averaging, given compressed volatility and high upside potential. He remains bullish long-term, citing the vast growth potential of tokenizing real-world assets and setting an $8000 price target for Ethereum.

marsbit3 год тому

Bitwise CIO Matt Hougan: If You Have 0% Crypto Allocation Now, It's Equivalent to Actively Shorting the Market Outlook

marsbit3 год тому

Breaking News: DeepSeek Announces All-Day Off-Peak Pricing on Weekends, Making Weekend Work More Cost-Effective?

DeepSeek has announced a significant change to its API pricing model, effective August 23. The new policy removes peak/off-peak distinctions on weekends (Saturdays and Sundays), charging the lower off-peak rate for the entire two-day period. This adjustment has sparked mixed reactions within the developer and professional communities. For developers and businesses heavily reliant on DeepSeek's V4-Flash and V4-Pro APIs, this is welcome news. It allows them to schedule bulk processing tasks on weekends without the higher peak-hour costs, potentially halving their API bills for such workloads. Some users have celebrated the move for making weekend work more cost-effective. However, the announcement has also raised concerns among employees. There is apprehension that companies, particularly in cost-sensitive sectors like AI-powered short drama production, might reorganize work schedules to align with these new cost incentives. Instances are already emerging where teams schedule high-token tasks during cheaper nighttime hours or adjust staff shifts. This has led to worries about a potential shift towards weekend workdays and weekday time-off, prioritizing cost savings over traditional work-life balance. Debate has ensued regarding the practicality of such schedule changes, with questions about increased communication overhead and overall efficiency. Speculation about DeepSeek's motives for the change includes theories that peak pricing correlates with internal model training schedules, though others counter that training is largely automated. The new pricing structure is now in effect, prompting users to reconsider their task scheduling strategies.

marsbit4 год тому

Breaking News: DeepSeek Announces All-Day Off-Peak Pricing on Weekends, Making Weekend Work More Cost-Effective?

marsbit4 год тому

Just Now, The World's First Human vs. Robot Tennis Match Begins, Robot's Desperate Save Leaves Zheng Jie Astonished

Just now, the world's first human vs. robot tennis match began, featuring stunning robotic saves that left tennis star Zheng Jie in awe. This historic event, part of the second World Humanoid Robot Games and broadcast live globally by China Media Group, marked a pivotal moment in Chinese technological innovation and embodied artificial intelligence. The match featured both mixed human-robot doubles and a groundbreaking singles match between Zheng Jie and the "Galaxy Xingzai" humanoid robot developed by Galaxy General. The robot demonstrated impressive skills including serving, forehands, backhands, and strategic court movement, with serves exceeding 100 km/h. It exhibited remarkable adaptability, recovering from a fall to continue play and handling slices and spins. The doubles match highlighted its ability to coordinate dynamically with a human partner. The event's significance extends far beyond a novelty match. Tennis represents an ultimate pressure test for embodied AI, demanding real-time integration of perception, decision-making, full-body motion control, and live博弈 within fractions of a second—a stark contrast to the discrete, contemplative environment of board games like Go mastered by AlphaGo. It directly confronts Moravec's paradox, showcasing AI's move from digital cognition to physical execution. This capability is powered by Galaxy General's proprietary "Galaxy Star Brain" (AstraBrain) model. Its key innovation is a unified architecture that integrates high-level task understanding/tactical decision-making ("brain") and dynamic whole-body motion control ("cerebellum") into a single model, eliminating latency and information loss between separate modules. The model was trained using a two-step process via the "Galaxy Star Workshop" platform. First, it learned foundational movement priors from "imperfect" human motion data (both amateur and professional). Second, it underwent massive-scale evolution in a virtual tennis simulator where multiple AI agents played millions of games against each other. Through this adversarial training, skills like极限救球and recovery from falls emerged autonomously without explicit programming, before being transferred to the physical robot. This "AstraTennis" moment symbolizes a major leap: a decade after AlphaGo conquered the digital world, embodied AI from China has now demonstrated it can operate under the extreme, unpredictable physical pressures of real-world competition.

marsbit4 год тому

Just Now, The World's First Human vs. Robot Tennis Match Begins, Robot's Desperate Save Leaves Zheng Jie Astonished

marsbit4 год тому

Торгівля

Спот
活动图片