ПродвижениеChatgpt/Claude/Grok 会员24h在线自助充值 安全快速 稳定保障
奥创(Ultra)商城会员充值中心
ПродвижениеChatgpt/Claude/Grok 会员24h在线自助充值 安全快速 稳定保障
奥创(Ultra)商城会员充值中心
Продвижение本页公开展示实时样本、模型覆盖与响应时间。先用 5 元按量跑自己的 Claude Code/Codex 任务,成功后再决定是否长期使用。
ACS Gateway
Продвижение【腾讯云】WorkBuddy 全场景 AI 办公工作台,团队旗舰版低至198元/月
腾讯云
Продвижение人人可用的AI应用创作平台 自然语言交互,全栈应用搭建,伟大创意快速实现;个人版新用户注册送5000积分
阿里云Старший вице-президент Google DeepMind Корай Кавукчуоглу говорит, что прогресс в AGI нельзя определить по одному тесту; важна и надежность моделей в повседневном использовании.
Google DeepMind опубликовала фрагмент беседы старшего вице-президента компании Корая Кавукчуоглу с Логаном Килпатриком. Кавукчуоглу заявил, что не существует теста или бенчмарка, который однозначно подтвердил бы достижение искусственного общего интеллекта (AGI).
По его словам, прогресс не сводится к результатам одной проверки: модели должны становиться достаточно надежными, чтобы люди могли пользоваться ими и доверять им в повседневных ситуациях. В беседе также затрагивались внедрение систем и обратная связь пользователей.
Участники не предложили формального порога или срока достижения AGI. Разговор подчеркивает, что оценка прогресса требует наблюдения и за возможностями моделей, и за их надежностью.
Проверка источников не гарантирует вечную точность. Учитывайте первоисточник, время наблюдения и последующие исправления.
We don’t have a clear benchmark, but I think trend extrapolation suggests that in the next one or two years, we’ll start to see more and more open problems being solved in other domains too. This will start with domains that are done mainly virtually and have verifiable answers, such as theoretical physics, computer science, and chemical and biological modelling. A single invention like nuclear weapons can change the course of history, so the importance of these capabilities should not be underestimated. ### What’s the case against impressive AI progress by 2030? Here’s the strongest case against in my mind: [...] Later, I’ll discuss empirical evidence that we’re seeing rapid progress in harder-to-verify domains as well as verifiable ones, just from a lower base. ### 3. Increasing how long models think If you could only think about a problem for a minute, you probably wouldn’t get far. If you could think for a month, you’ll make a lot more progress — even though your raw intellig
For instance, models are quickly approaching human-level performance on challenging evaluations like GPQA (Rein et al., 2023) for expert-level knowledge, SWE-bench (Jimenez et al., 2024) for software engineering, and FrontierMath (Glazer et al., 2024) for advanced mathematics. This saturation highlights the need for benchmarks that measure true generalization, such as ARC-AGI (Chollet, 2019) and successors, or the development of private benchmarks and continuous adversarial evaluation to accurately track progress. Two big open research challenges that stand out in that field are: one, designing ASI benchmark methodologies that measure general capabilities and do not saturate at human level and can be produced and run with no or very little human input, and two, measuring how capabilities [...] Current models lack a mechanism to discover the concepts of ’force’ or ’causality’ from scratch. They inherit these by successfully ingesting large amounts of data generated by an intelligence (u
Another issue is that AI systems have different strengths and weaknesses from humans, so even if we define AGI as “AI that can match humans at most tasks,” we can debate which tasks really count, and which humans set the standard. Direct comparisons are difficult. “We’re building alien beings,” says Geoffrey Hinton, a professor emeritus at the University of Toronto who won a Nobel Prize for his work on AI. Undaunted researchers are busy designing and proposing tests that might lend some insight into our future. But a question remains: Can these tests tell us if we’ve achieved the long-sought goal of AGI? ## Why It’s So Hard to Test for Intelligence [...] But this year, researchers reported that when people conversed with both another person and OpenAI’s GPT-4.5 for 5 minutes and then had to guess which one was human, they picked the AI 73 percent of the time. Meanwhile, top language models frequently make mistakes that few people ever would, like miscounting the number of times the l
Each test measures a slice, and the slices disagree. In the week Astra launched, Epoch AI ranked it first while Artificial Analysis rated it no better than its predecessor. Same model, same week, opposite verdicts. The people who build the tests expect them to keep changing. DeepMind argues an AGI benchmark “should be a living benchmark.” Chollet says benchmarking “is a continual process that co-evolves with the models.” Every test that falls gets replaced by a harder one, further down the field. ## Not one test, not one moment. Just a zone. The conclusion I’ve reached: AGI is a zone we're wading through, with goalposts that keep moving alongside our expectations. There's no single line to step over, and no moment when everyone agrees we've crossed it. [...] That’s four declarations in under eighteen months, mostly from people who build these systems or sell the chips they run on, so some of it is talking their book. However, the researchers can’t agree either. In February, four UC
alliekmiller The CTO of Google DeepMind has a simple tell for how close we are to AGI, and it has little to do with benchmarks. Count the agents in your day. When I sat down with Koray Kavukcuoglu, he told me the first signal is surface area, meaning agents handle more of your work because their decisions finally got good enough to trust. The second signal comes after that, when a model can take on a problem nobody trained it for and learn from it on its own. That second signal is the one to watch. In our full conversation we also got into whether the chip shortage is real, how self-learning AI already works without retraining, and the one daily habit he says will carry you through 2026. Follow me + comment DEEPMIND and I’ll DM you the link.
### Transcript [0:00] There's always a lot of discussion about is there a test for AGI? There's no test for AGI. [0:05] I don't think anyone has a test that they can say, okay, if you get this much on this test, that means you are at AGI. [0:12] And I don't think that is— - If you do have that benchmark, send it to us, please, because we want to— we'll use it, we'll run our models on it. [0:24] Hey, everyone, welcome back to Release Notes. My name is Logan Kilpatrick. I'm on the Google DeepMind team. Today we're joined by Koray Kavukcuoglu. Koray, I'm super excited for this conversation. [...] [8:47] That's how we have something general that people like interacting with, that people rely on, that people trust. [8:53] - Yeah, I think last time we were talking you had mentioned this co-building AGI with our customers framing, which I think has just continued to like it makes so much sense. [9:02] It resonates. I think it makes it clear why the deployment matters so much, why this feedbac