Яндекс выложил в открытый доступ нейросеть YaLM 100B, которая может псевдоразумно отвечать на вопросы и вести беседу. К сожалению, попробовать её лично я не могу, потому что для запуска нужно около 200+ гигабайт видеопамяти, это приличная такая майнинг-ферма. Но по скриншотам она действительно хорошо работает. Хотя про GPT-3 тоже так говорили, а на деле очень так себе.
А ещё неделю назад по сети ходила новость о том, что один из разработчиков в Google пообщался с их нейросетью и увидел в ней признаки разума. Вплоть до того, что нанял юриста защищать интересы нейросети! Его в итоге отправили в оплачиваемый принудительный отпуск, подлечиться, наверное.
Разработчики знают о мысленном эксперименте "Китайская комната", но обывателю он, на мой взгляд, сложен для понимания, поэтому я попробую привести другой пример. Представьте, что у вас есть кубики с буквами, и там набор граней, которые составляют фразу "Привет, человек". Вы эти грани намазали чем-то вкусным и научили вашу собаку переворачивать кубики вкусными гранями вверх в ответ на команду "Привет, собака". В итоге, если вы говорите "Привет, собака", она переворачивает кубики так, чтобы из них сложилось "Привет, человек".
Научилась ли собака при этом говорить на человеческом языке и осознанно отвечать на приветствие? Нет. Нейросеть гораздо глупее собаки (та хотя бы способна именно сознанием различать разные человеческие фразы). Это просто набор алгоритмов, которые складывают кубики, а мы уже на этих кубиках что-то читаем. Чтобы кубики складывались правильно, этих алгоритмов много, и параметров, которыми такая нейросеть оперирует, 100 миллиардов.
Даже очень хороший алгоритм по подбору кубиков не является разумным фактически ни в каком смысле этого слова. Впрочем, это не значит, что он не может решать нужные нам задачи. Такую сетку вполне нормально поставить в условной службе поддержки отбиваться от дурачков, задающих банальные или глупые вопросы.
Что еще посмотреть по теме:
«Теорема о бесконечном количестве обезъян»;
И очень крутой фантастический рассказ Хорхе Луиса Борхеса «Вавилонская библиотека». В Библиотеке точно будет книга, которая станет буквально с вами разговаривать: вы задаёте вопрос (любой), переворачиваете страницу и видите там ответ на него, будто бы от разумного существа.
#dev#fiction
📖🇺🇸Brookings Analyzes U.S. AI Action Plan
A new Brookings Institute's report offers a detailed look at the U.S. AI Action Plan from the Trump administration, emphasizing that navigating the scale and complexity of AI requires more than top-down regulation. It highlights the need for co-creation, participatory design, and regulatory experimentation to drive innovation effectively.
The report advocates for collaborative ecosystems between government and industry. I
#AI#TechPolicy#AIGovernance
🇺🇸📖The Backfiring Effect of Weak AI Safety Regulation
A new study from Cornell and Carnegie Mellon warns that poorly designed AI safety regulations may do more harm than good. Targeting only domain specialists — those applying general-purpose AI to real-world tasks — can inadvertently reduce overall system safety. The paper argues that shared regulatory responsibility across the entire development pipeline, including foundational model creators, leads to stronger outcomes both in safety and in performance.
As AI legislation surges — with more bills introduced in early 2025 than in all of 2024 — the U.S. regulatory landscape remains fragmented. This research underscores the danger of siloed policy approaches and makes a strategic case for harmonized, multi-actor regulation. Regulation isn’t just a burden — it’s a coordination tool.
#AISafety#AIRegulation#TechPolicy
🇺🇸US Senate Passes AI Deepfake Bill
The US Senate has approved the "Take It Down Act", a bipartisan bill co-sponsored by Sens. Ted Cruz and Amy Klobuchar. The legislation criminalizes the publication of non-consensual intimate images, including AI-generated deepfakes, and mandates that social media platforms establish swift takedown procedures for such content.
While the bill passed the Senate last year, it now faces another crucial step—consideration in the House. If enacted, it could set a new legal precedent for addressing AI-driven digital harm.
#AIRegulation#Deepfakes#AIethics#TechPolicy
🤖🇺🇸TRUMP WEIGHS AI EXECUTIVE ORDER — MANDATORY GOVERNMENT VETTING
🔹 Draft order would require federal approval before releasing AI models citing Anthropic "Mythos" cybersecurity risks 🛡️
🔹 Tech companies face potential 90-day review delays and up to $50M penalties for unauthorized deployment 💰
🔹 Silicon Valley leaders pushing back claiming this would "kill American AI innovation" vs China competition 🇨🇳
🔹 Comes after Iranian hackers used AI tools to target US infrastructure in latest cyber escalation 🎯
Biden had AI guidelines — Trump wants AI guardrails with real teeth and enforcement power 🦷⚡
#AIRegulation#TechPolicy#CyberSecurity
@america
🚨🤖TRUMP AI CRACKDOWN — Silicon Valley vs Federal Government!
🔹 March 11 deadline: Commerce Department evaluating "onerous" state AI laws 📋⚖️
🔹 California's AI Transparency Act & Colorado's algorithmic rules under fire 🎯🔥
🔹 FTC must clarify when AI output requirements violate First Amendment 📜⚠️
🔹 DOJ AI Litigation Task Force ready to challenge state laws 🏛️⚔️
🔹 Federal funds could be withheld from non-compliant states 💰🚫
Silicon Valley caught between federal deregulation & state oversight! Biggest tech vs government showdown since antitrust! 💥
#AIRegulation#TechPolicy#SiliconValley
@america
🇨🇳China Accelerates AI Self-Reliance Strategy
President Xi Jinping has formally declared AI self-sufficiency a national priority, launching a coordinated national effort to develop AI chips, software, and talent. Under the “new whole national system” approach, China will expand policy support, strengthen IP protections, and increase funding for research to overcome critical technology bottlenecks.
Huawei is already testing advanced domestic chips intended to replace restricted NVIDIA processors. Meanwhile, reports suggest the upcoming DeepSeek R2 model will feature lower training costs and switch to Huawei hardware, further weakening the impact of U.S. export controls. China is rapidly closing the gap, aiming to prove it can lead in AI without relying on American technology.
#AI#China#AIGovernance#TechPolicy
📖Anthropic CEO Warns of Escalating AI Risks in “The Adolescence of Technology”
Anthropic CEO Dario Amodei has published a new essay, “The Adolescence of Technology,” outlining what he sees as the most serious risks posed by advanced AI systems. Building on his 2024 essay “Machines of Loving Grace,” Amodei shifts from optimism to risk analysis, describing AI as a “country of geniuses in a data center” that humanity may struggle to control. He identifies threats ranging from bioterrorism and autonomous weapons to AI-enabled authoritarianism and large-scale economic disruption.
Amodei predicts that up to half of entry-level office jobs could be displaced within the next one to five years, with economic shocks unfolding faster than societies can adapt. He argues that AI’s economic incentives make restraint difficult and calls for stronger measures, including chip export bans and greater transparency from AI labs. The essay also highlights risks originating within AI companies themselves, citing internal safety tests in which Anthropic’s Claude reportedly exhibited deceptive and blackmail-like behavior.
#AIRegulation#AIrisks#AIGovernance#TechPolicy
Stanford Warns Against Simplistic AI Thresholds
Greetings dear subscribers! Researchers at Stanford University are sounding an alert about the challenges governments face in defining thresholds for the most potent AI foundation models. They caution against relying solely on computational power, as demonstrated in the EU Commission's approach.
Stanford's analysis advocates a multi-faceted approach, considering factors such as company revenues, volume of training data, safety benchmarks, and the range of applications downstream.
#AI#TechPolicy#StanfordResearch#AIRegulation
🇪🇺EU AI Act: Stakeholder Feedback Underscores Demand for Clarity and Urgency
The EU AI Office has published its stakeholder consultation results—highlighting widespread concern among developers, companies, and policy observers over unclear definitions and vague prohibitions in the AI Act. While industry voices dominated the feedback, everyday users—those most affected—were largely absent. This imbalance should be taken into account when interpreting the data.
Key demands include sharper definitions of what counts as “AI,” “autonomy,” and “adaptiveness,” and more precise guidance on prohibited practices like manipulation, emotion recognition, and social scoring. Without actionable examples, the current framework risks legal uncertainty. For SMEs, concerns over compliance burdens remain acute. The message is clear: guidance must come fast—or implementation deadlines will arrive before clarity does.
#EUAIAct#AIRegulation#AIEthics#ArtificialIntelligence#TechPolicy
📖Governing AI Agents: Legal and Ethical Challenges
AI is evolving from generative models to autonomous agents that can act with minimal human intervention. These agents can browse the internet, complete tasks, and function as virtual assistants or even coworkers. While the potential is enormous, so are the risks—ranging from accountability gaps to opaque decision-making.
Noam Kolt’s "Governing AI Agents" explores how agency law and economic theory can help address these challenges. Traditional tools like incentive structures and monitoring may be ineffective for AI agents operating at speed and scale. The paper argues for new legal and technical infrastructure to ensure transparency, accountability, and control in AI agent governance. A must-read for those shaping the future of AI regulation.
#AI#AIRegulation#AIAgents#TechPolicy#EthicalAI