2024 年人工智能现状:辉煌、戏谑和“牛市”

2024 年人工智能现状:辉煌、戏谑和“牛市”

2024 年人工智能现状:辉煌、戏谑和“牛市” 2024 年人工智能现状:辉煌、戏谑和“牛市” 中文 探索人工智能的过山车之旅,充满着《让子弹飞》(嗯,中文世界的非物质文化遗产)中那些越品越有的梗,过度解读?低估还是唱衰唯有时间这条长河可以给予我们终极答案…… 太长不看! 但如果有空!还是请看我顶着高烧,在抖包袱又表面平和、不断搜寻脑海中的词汇、尝试用滑稽的幽默来博君一笑努力吧!again,以下内容 anti AI!人工撰写 by 🦄ChaX • OpenAI 的 o1 模型主导:OpenAI 最新推出的 o1 模型正在重新定义 AI 在数学、科学和推理方面的极限,使竞争对手困惑甚至“破产”。 • 中国的 AI 崛起:无视制裁,中国的模型凭借坚韧和战略智慧正在“屠榜”,证明他们也仍然在牌桌之上…… • 生成式 AI 的数十亿繁荣:AI 初创公司正赚得盆满钵满,但可持续性仍然像一个能理解细微差别的聊天机器人一样难以捉摸。 报告链接: The State of AI 2024 (友情提示,报告 212 页)| 报告的中文翻译版 英文解读链接: The State of AI 2024: Brilliance, Banter, and Bull Runs 写在前面的:欢迎来到……2024 年人工智能现状 亲爱的读者们,热烈欢迎踏入这篇关于 2024 年人工智能现状的奇妙文章——《AI 现状报告》第七版!这里不仅展示了 AI 领域的辉煌成就和艰巨挑战,还穿插了一些让人捧腹大笑的荒诞画面。这份报告就像是一碗热辣辣的四川火锅,麻辣鲜香,让人欲罢不能。 研究篇:边缘竞赛(及以后) • OpenAI 的 o1 模型:这个新来的“壮汉”正在疯狂推动深度强化学习的推理,代价嘛,就是和现在全球都在降本增效的“降本”格格不入。它解决问题的能力一流,但每个 token 的成本,都让你自我怀疑,是的“世界上只有一种病,就是穷病”而“日渐见底的荷包,显然是你我的问题,不是它的问题……”毕竟一个一心进取、不断自我突破的 OpenAI 有什么坏心眼呢,是吧!🙂 • Meta 的 Llama 3:Meta 这次可是拼了,不顾一切(甚至忘了 Chinchilla 最优限制),用 15 万亿 token 训练出了 Llama 3.1 模型。Llama 的力量现在是大白于天下,但“开放”的程度嘛,得看你怎么定义“开放”🙂 • 数据污染闹剧:是的!一些模型的出色表现可能是因为……作弊,而且方法既不先进也不高端——训练集被基准数据污染。因此目前的通用测试对它们来说,就是明晃晃的开卷考试。看起来 AI 正在追随“每个有天赋的孩子”的脚步,通过已经给了答案的内定考试?! 行业篇:NVIDIA、AI 独角兽和 3 万亿美元俱乐部 • NVIDIA 的霸主地位:NVIDIA 以 3 万亿美元的估值和无人能敌的 GPU 统治力,稳坐 AI 计算界的头把交椅。不过,这种权力集中可是引起了监管机构的“关注”,他们就像那些爱八卦的只在春节时相见的“长辈们”,时刻盯着你的一举一动。 • 生成式 AI 的金钱盛宴:其实,我作为一个移动互联网时代的活化石,成熟的生成式 AI 公司赚得盆满钵满,初创公司也跃跃欲试。结果呢?AI 公开市场迎来了一波大牛市,总价值高达 9 万亿美元。不过,房间里的大象(长期盈利能力的问题)依然挥之不去。 政治篇:监管进展,算是吧 • 全球治理的尴尬局面:全球在 AI 治理上的合作就像一场尴尬的华丽晚宴,大家都不知道该说些什么——承诺满天飞,实际行动却寥寥无几。🙂 • 国家和地区法规的角力:疫情之后,越发魔幻的世界让我产生了世界在玩一种很新的“闭关锁国”,美国和欧盟等国家或地区正在通过有争议的国家层面立法,仿佛在说:“看,我们比你们更懂得搞官僚主义!”科技公司在这场与监管巨头的博弈中,还需面对自身可持续性目标的挑战——发展 AI 的排放量简直就像是一场全国范围的 SUV 自驾游,环保得让人哭笑不得。 安全篇:调子在改变 • 从恐慌到繁荣:AI 安全曾经是个热门话题,但现在公司们更关心的是如何加速赚钱而非灭绝风险——毕竟,没有什么比快速赚钱更能代表“安全的未来”了。而“落后就要挨打”这达摩克利斯之剑横亘在头顶上!让全世界大国小国、商业或政府亦或是平民各个层级的参与者都不得不先补一张这条通向未来船票再说😄 • 越狱问题的持续挑战:越狱 AI 系统的技巧已经成为一个强大的对手。研究人员推出的每一个“补丁”都被巧妙地规避,这让安全领域的专家们忧心忡忡。这就像是一场打地鼠游戏,地鼠变得越来越快,而你手里还拿着个巨大的泡沫锤。而中国古话“道高一尺魔高一丈”,奈何魔会自己进化…… 未来预测 • 更多资金,更多问题:预计明年会有团队花费超过 10 亿美元来训练单个大型模型。生成式 AI 的热潮不会消退,只会变得更加“奢华”。😱 • 计算压力的挑战:政府和大型科技公司将继续承受计算需求的压力,这些需求已经逼近电网的极限。这就像一个健身网红的体重增长超过了突破了牛顿定理、超过了人类健身房的承重能力——总有一天,天花板可能会真的塌下来。😱 • AI 介入选举:虽然预期的 AI 对选举和就业的影响尚未成真,但这并不意味着我们可以掉以轻心。但请记住,AI 的影响就像潘多拉魔盒:一旦打开,将会在接下来的几年里、几十年甚至百年亦或是人类此后的历史中都无处不在。而众所周知,诚如百年前的黑格尔所预言的那样,我们人类啊!从历史中获得唯一的教训就是……我们没有从历史中获得任何教训!🤔 总而言之,言而总之,人工智能领域依然是一个充满惊喜、伦理挑战和很多真金白银的奇妙世界。在这个舞台上,OpenAI、Meta 和来自东方的神秘力量等各方势力竞相角逐,2024 年的 AI 领域更像是一场激动人心的 UFC(无限制综合格斗冠军赛)比赛而非平静的学术研讨会。所以,请准备好你的爆米花和 GPU 吧——新的一年里,AI 将继续突破界限、挑战极限,甚至可能重新定义一些我们熟悉的词汇! 英文 Exploring the rollercoaster ride of artificial intelligence with flair and unpredictability Key Takeaways • OpenAI's o1 Model Dominates: OpenAI’s o1 is redefining the limits of AI in math, science, and reasoning, leaving competitors bewildered and broke. • China's AI Surge: Defying sanctions, Chinese models are scaling leaderboards with sheer resilience and strategic brilliance, proving they’re here to dominate. • Generative AI's Billion Dollar Boom: AI start ups are cashing in big, but sustainability remains as elusive as a chatbot that can grasp nuance. Report Link: The State of AI 2024 Introduction: Welcome to the State of AI 2024 The seventh installment of the State of AI Report gives us a panorama of triumphs, trials, and a sprinkle of the absurd in the AI landscape. It covers everything from research breakthroughs and regulatory drama to the fine balance between AI’s glorious potential and its lurking risks. And yes, as always, it’s as much about understanding what’s happening now as it is about predicting what’s going to slap us in the face next year. Research: The Race to the Edge (and Beyond) • OpenAI's o1 Model: Say hello to o1, the model that’s bringing deep reinforcement learning to inference in a big way—and at a cost that makes your fancy coffee habit look downright frugal. It’s great at solving complex problems, but the bill per token makes it a bit like the AI equivalent of a luxury yacht: impressive, but you better have deep pockets. • Meta’s Llama 3: Meta decided to throw caution (and Chinchilla optimal limits) to the wind, training the Llama 3.1 model with 15 trillion tokens and managing to level the playing field with closed frontier models. The Llama's power is now in the open, yet how open it really is depends on how you define "open." • Data Contamination Drama: Researchers are realizing that some models' stellar performances may be due to good ol’ fashioned cheating—with training sets contaminated by benchmark data. Looks like AI is following in the footsteps of “every gifted kid” by acing exams through familiarity. Industry: NVIDIA, AI Unicorns, and a 3 Trillion Dollar Club • NVIDIA Dominates: NVIDIA has secured its place as the overlord of AI compute, with its $3 trillion valuation and unchallenged GPU supremacy. But this concentration of power has regulators peeking around the corner—they’re the kind of neighbor that watches your every move through the blinds. • GenAI's Cash Splash: Established GenAI companies are raking in billions, and there’s no shortage of start ups trying to cash in. The result? A bull run in AI public markets with a combined value of $9 trillion. However, there’s a small elephant in the room: long term profitability is still a giant shrug emoji. Politics: Regulation Advances, Sort Of • Global Governance Is Stalling: Global cooperation on AI governance seems like a polite dinner party where nobody knows what to say—voluntary commitments abound, but there’s no real action. • National and Regional Regulations: Meanwhile, countries like the US and EU are passing controversial legislation at the national level, just to remind everyone that they can play with red tape better than anyone else. As tech companies wrestle with regulatory giants, they also face the not so small challenge of their own sustainability targets—ironic, given that scaling AI is about as emissions friendly as an SUV road trip across the country. Safety: A Changing Tune • From Doom to Boom: AI Safety was all the rage until companies realized fear wasn’t selling as many enterprise subscriptions as they’d hoped. Now, these same players are talking more about acceleration than extinction—because nothing says "safe future" like pivoting to make a quick buck. • The Jailbreaking Problem Persists: The art of jailbreaking AI systems has proven to be quite the adversary. Every “fix” introduced by researchers has been circumvented, which only makes those in the safety field more concerned. It’s like a game of whack a mole where the moles have gotten faster and you’re still holding a giant foam hammer. Predictions for the Future • More Money, More Problems: Expect a group to spend over $1 billion training a single large scale model next year. The GenAI craze isn’t going anywhere; it’s just getting a pricier wardrobe. • Compute Squeeze: Governments and Big Tech alike will continue to feel the crunch of compute requirements, which are outgrowing the power grid’s ability to support them. It’s like a bodybuilder bulking up faster than their gym can handle—at some point, the ceiling might just collapse. • AI Enters Elections: Anticipated AI effects on elections and employment haven’t materialized yet, but that doesn’t mean we’re in the clear. Remember, AI’s influence is like glitter: once it’s out, you’ll be finding it in everything for years to come. Artificial intelligence continues to be a mixed bag of thrilling breakthroughs, unanswered ethical questions, and oh so many trillion dollar bets. With OpenAI leading in some areas, Meta in others, and Chinese labs rising undeterred by sanctions, the stage for AI in 2024 feels more like a tightly packed boxing ring than a peaceful conference room. So grab your popcorn and keep those GPUs handy—we’re in for another year of AI pushing boundaries, testing limits, and possibly rewriting a few definitions along the way. The State of AI 2024 报告的中文翻译版 The State of AI 2024: Brilliance, Banter, and Bull Runs The State of AI 2024 中文 探索人工智能的过山车之旅,充满着《让子弹飞》(嗯,中文世界的非物质文化遗产)中那些越品越有的梗,过度解读?低估还是唱衰唯有时间这条长河可以给予我们终极答案…… 太长不看! 但如果有空!还是请看我顶着高烧,在抖包袱又表面平和、不断搜寻脑海中的词汇、尝试用滑稽的幽默来博君一笑努力吧!again,以下内容 anti AI!人工撰写 by 🦄ChaX • OpenAI 的 o1 模型主导:OpenAI 最新推出的 o1 模型正在重新定义 AI 在数学、科学和推理方面的极限,使竞争对手困惑甚至“破产”。 • 中国的 AI 崛起:无视制裁,中国的模型凭借坚韧和战略智慧正在“屠榜”,证明他们也仍然在牌桌之上…… • 生成式 AI 的数十亿繁荣:AI 初创公司正赚得盆满钵满,但可持续性仍然像一个能理解细微差别的聊天机器人一样难以捉摸。 报告链接: The State of AI 2024 (友情提示,报告 212 页)| 报告的中文翻译版 The State of AI 2024 报告的中文翻译版 英文解读链接: The State of AI 2024: Brilliance, Banter, and Bull Runs The State of AI 2024: Brilliance, Banter, and Bull Runs 写在前面的:欢迎来到……2024 年人工智能现状 亲爱的读者们,热烈欢迎踏入这篇关于 2024 年人工智能现状的奇妙文章——《AI 现状报告》第七版!这里不仅展示了 AI 领域的辉煌成就和艰巨挑战,还穿插了一些让人捧腹大笑的荒诞画面。这份报告就像是一碗热辣辣的四川火锅,麻辣鲜香,让人欲罢不能。 研究篇:边缘竞赛(及以后) • OpenAI 的 o1 模型:这个新来的“壮汉”正在疯狂推动深度强化学习的推理,代价嘛,就是和现在全球都在降本增效的“降本”格格不入。它解决问题的能力一流,但每个 token 的成本,都让你自我怀疑,是的“世界上只有一种病,就是穷病”而“日渐见底的荷包,显然是你我的问题,不是它的问题……”毕竟一个一心进取、不断自我突破的 OpenAI 有什么坏心眼呢,是吧!🙂 • Meta 的 Llama 3:Meta 这次可是拼了,不顾一切(甚至忘了 Chinchilla 最优限制),用 15 万亿 token 训练出了 Llama 3.1 模型。Llama 的力量现在是大白于天下,但“开放”的程度嘛,得看你怎么定义“开放”🙂 • 数据污染闹剧:是的!一些模型的出色表现可能是因为……作弊,而且方法既不先进也不高端——训练集被基准数据污染。因此目前的通用测试对它们来说,就是明晃晃的开卷考试。看起来 AI 正在追随“每个有天赋的孩子”的脚步,通过已经给了答案的内定考试?! 行业篇:NVIDIA、AI 独角兽和 3 万亿美元俱乐部 • NVIDIA 的霸主地位:NVIDIA 以 3 万亿美元的估值和无人能敌的 GPU 统治力,稳坐 AI 计算界的头把交椅。不过,这种权力集中可是引起了监管机构的“关注”,他们就像那些爱八卦的只在春节时相见的“长辈们”,时刻盯着你的一举一动。 • 生成式 AI 的金钱盛宴:其实,我作为一个移动互联网时代的活化石,成熟的生成式 AI 公司赚得盆满钵满,初创公司也跃跃欲试。结果呢?AI 公开市场迎来了一波大牛市,总价值高达 9 万亿美元。不过,房间里的大象(长期盈利能力的问题)依然挥之不去。 政治篇:监管进展,算是吧 • 全球治理的尴尬局面:全球在 AI 治理上的合作就像一场尴尬的华丽晚宴,大家都不知道该说些什么——承诺满天飞,实际行动却寥寥无几。🙂 • 国家和地区法规的角力:疫情之后,越发魔幻的世界让我产生了世界在玩一种很新的“闭关锁国”,美国和欧盟等国家或地区正在通过有争议的国家层面立法,仿佛在说:“看,我们比你们更懂得搞官僚主义!”科技公司在这场与监管巨头的博弈中,还需面对自身可持续性目标的挑战——发展 AI 的排放量简直就像是一场全国范围的 SUV 自驾游,环保得让人哭笑不得。 安全篇:调子在改变 • 从恐慌到繁荣:AI 安全曾经是个热门话题,但现在公司们更关心的是如何加速赚钱而非灭绝风险——毕竟,没有什么比快速赚钱更能代表“安全的未来”了。而“落后就要挨打”这达摩克利斯之剑横亘在头顶上!让全世界大国小国、商业或政府亦或是平民各个层级的参与者都不得不先补一张这条通向未来船票再说😄 • 越狱问题的持续挑战:越狱 AI 系统的技巧已经成为一个强大的对手。研究人员推出的每一个“补丁”都被巧妙地规避,这让安全领域的专家们忧心忡忡。这就像是一场打地鼠游戏,地鼠变得越来越快,而你手里还拿着个巨大的泡沫锤。而中国古话“道高一尺魔高一丈”,奈何魔会自己进化…… 未来预测 • 更多资金,更多问题:预计明年会有团队花费超过 10 亿美元来训练单个大型模型。生成式 AI 的热潮不会消退,只会变得更加“奢华”。😱 • 计算压力的挑战:政府和大型科技公司将继续承受计算需求的压力,这些需求已经逼近电网的极限。这就像一个健身网红的体重增长超过了突破了牛顿定理、超过了人类健身房的承重能力——总有一天,天花板可能会真的塌下来。😱 • AI 介入选举:虽然预期的 AI 对选举和就业的影响尚未成真,但这并不意味着我们可以掉以轻心。但请记住,AI 的影响就像潘多拉魔盒:一旦打开,将会在接下来的几年里、几十年甚至百年亦或是人类此后的历史中都无处不在。而众所周知,诚如百年前的黑格尔所预言的那样,我们人类啊!从历史中获得唯一的教训就是……我们没有从历史中获得任何教训!🤔 总而言之,言而总之,人工智能领域依然是一个充满惊喜、伦理挑战和很多真金白银的奇妙世界。在这个舞台上,OpenAI、Meta 和来自东方的神秘力量等各方势力竞相角逐,2024 年的 AI 领域更像是一场激动人心的 UFC(无限制综合格斗冠军赛)比赛而非平静的学术研讨会。所以,请准备好你的爆米花和 GPU 吧——新的一年里,AI 将继续突破界限、挑战极限,甚至可能重新定义一些我们熟悉的词汇! 英文 Exploring the rollercoaster ride of artificial intelligence with flair and unpredictability Key Takeaways • OpenAI's o1 Model Dominates: OpenAI’s o1 is redefining the limits of AI in math, science, and reasoning, leaving competitors bewildered and broke. • China's AI Surge: Defying sanctions, Chinese models are scaling leaderboards with sheer resilience and strategic brilliance, proving they’re here to dominate. • Generative AI's Billion Dollar Boom: AI start ups are cashing in big, but sustainability remains as elusive as a chatbot that can grasp nuance. Report Link: The State of AI 2024 The State of AI 2024 Introduction: Welcome to the State of AI 2024 The seventh installment of the State of AI Report gives us a panorama of triumphs, trials, and a sprinkle of the absurd in the AI landscape. It covers everything from research breakthroughs and regulatory drama to the fine balance between AI’s glorious potential and its lurking risks. And yes, as always, it’s as much about understanding what’s happening now as it is about predicting what’s going to slap us in the face next year. Research: The Race to the Edge (and Beyond) • OpenAI's o1 Model: Say hello to o1, the model that’s bringing deep reinforcement learning to inference in a big way—and at a cost that makes your fancy coffee habit look downright frugal. It’s great at solving complex problems, but the bill per token makes it a bit like the AI equivalent of a luxury yacht: impressive, but you better have deep p

在 小宇宙note 阅读完整内容