The Rundown AI -每日精选 08月12日
Meta's mind-reading movie AI
index_new5.html
../../../zaker_core/zaker_tpl_static/wap/tpl_guoji1.html

 

Meta 的 FAIR 团队推出了 TRIBE 模型,一个能够预测大脑对视频反应的 AI。该模型无需脑部扫描,通过分析视频、音频和文本,准确预测观众大脑活动区域,并在脑科学建模竞赛中取得佳绩。同时,OpenAI 的推理模型在国际信息学奥林匹克竞赛中斩获金牌,展示了其在复杂问题解决方面的强大能力。此外,文章还介绍了 ElevenLabs 的音乐生成工具,Lockheed Martin 与 IBM 合作开发 AI,以及韩国研究人员利用 AI 设计癌症药物的最新进展。

🧠 Meta 的 TRIBE 模型:这是一个拥有10亿参数的神经网络,能通过分析视频、音频和文本,预测人脑对视频内容的反应。它在没有任何脑部扫描的情况下,准确预测了超过一半的观众大脑活动模式,尤其在涉及视觉、听觉和语言整合的脑区表现出色,比单一感官模型提升了30%,并在控制注意力和决策的前额叶脑区展现出高精度。

🏅 OpenAI 模型在编程竞赛中获胜:OpenAI 的一个通用推理模型在2025年国际信息学奥林匹克竞赛(IOI)中获得了金牌,排名在人类参赛者中位列第六,在所有 AI 模型中位列第一。该模型在没有针对性编程微调的情况下,成功解决了编码问题,并且在过去一年里其得分实现了2倍的飞跃,预示着AI在科学发现等领域将扮演更重要角色。

🎵 ElevenLabs 音乐生成工具:用户可以通过简单的文本提示,利用 ElevenLabs 的新音乐模型创作专业级音乐。该模型支持多种语言,并提供完全的创意控制,用户可以指定音乐风格、人声等细节,生成后还能进行编辑和下载。

🚀 Lockheed Martin 与 IBM 合作:Lockheed Martin 利用 IBM 的 watsonx.data 平台统一其工程工具,显著提升了 AI 产品开发的效率和智能化水平。该平台帮助10,000名工程师构建可扩展的 AI 产品,并通过 AI 回答员工问题,将准确率提高了20%,同时解决了数据孤岛问题。

💊 韩国研究人员的 AI 药物设计:KAIST 的研究人员开发了一种名为 BInD 的扩散模型,能够从零开始设计优化的癌症药物候选物,无需任何先验分子数据。BInD 能一次性设计药物分子及其与目标蛋白的结合方式,精准靶向癌细胞突变,同时保证药物的安全性和可制造性。

Read Online | Sign Up | Advertise

Good morning, {{ first_name | AI enthusiasts }}. Meta’s FAIR team just built an AI that knows what your brain will do before you even press play on a video — and it doesn't need a single brain scan to do it.

With the company’s TRIBE model, Meta is pushing the frontiers of brain science — but is also potentially writing the instruction manual for neural-level addictive content in the process.

P.S. — Not familiar with a certain term in the newsletter? Check out our AI dictionary, a quick-reference guide for different models, AI labs, technical lingo, and more.


In today’s AI rundown:

    Meta’s AI predicts brain responses to videos

    OpenAI's reasoner snags gold at programming olympiad

    Create professional music tracks for projects

    Korean researchers’ AI designs cancer drugs

    4 new AI tools, community workflows, and more

LATEST DEVELOPMENTS

META FAIR

🧠 Meta’s AI predicts brain responses to videos

Image source: Midjourney / The Rundown

The Rundown: Meta’s FAIR team just introduced TRIBE, a 1B parameter neural network that predicts how human brains respond to movies by analyzing video, audio, and text — achieving first place in the Algonauts 2025 brain modeling competition.

The details:

    TRIBE analyzes video, audio, and dialogue from movies, accurately predicting which of the viewer’s brain regions will activate without any brain scanning.

    The AI correctly predicted over half brain activity patterns across 1,000 brain regions after training on subjects who watched 80 hours of TV and movies.

    It works best in brain areas where sight, sound, and language merge, outperforming single-sense models by 30%.

    Meta's system also showed particular accuracy in frontal brain regions that control attention, decision-making, and emotional responses to content.

Why it matters: We’ve only uncovered the tip of the iceberg when it comes to understanding the brain and its processes, and TRIBE and other AI systems are ramping up that knowledge. But they are also providing new formulas for maximizing attention on a neural level, potentially making doomscrolling even more irresistible.

TOGETHER WITH FUEL iX

📆 Secure your GenAI before it's too late

The Rundown: In the race to innovate with GenAI, stop gambling with GenAI security. Join TELUS Digital, Fuel iX, and guest Forrester for a discussion on how product teams can fast-track secure GenAI development without compromising on safety or speed.

This exclusive session will explore:

    Battle-tested security frameworks from leading enterprises

    Practical automation strategies for enterprise-wide protection

    Real implementation stories

Watch on-demand to see industry experts share their proven paths to secure, scalable GenAI deployment.

OPENAI

🏅 OpenAI's reasoner snags gold at programming olympiad

Image source: Sheryl Hsu (@SherylHsu02 on X)

The Rundown: OpenAI announced that its reasoning model achieved a gold-level score at the 2025 International Olympiad in Informatics (IOI), placing 6th against humans and first among AI in the world’s top pre-college programming competition.

The details:

    The AI competed against top student programmers worldwide, solving coding problems with the same time and submission limits as human contestants.

    OpenAI’s model was a general-purpose reasoner, without specific fine-tuning for programming and relying on just basic tools.

    The system scored in the 98th percentile, a massive jump from a 49% score just a year ago.

    The same model also won gold at the International Math Olympiad and AtCoder, showing strength across a range of complex problem-solving areas.

Why it matters: The 2x leap in score shows how fast reasoning capabilities have truly moved over the past year. The days of humans ahead of AI in competitions are numbered, and these achievements will likely be the stepping stones towards future models that are capable of discovering new science, math, physics, and more.

AI TRAINING

🎵 Create professional music tracks for projects

The Rundown: In this tutorial, you will learn how to generate studio-grade music from simple text prompts using the new ElevenLabs music model that supports multiple languages and gives you complete creative control.

Step-by-step:

    Visit ElevenLabs and navigate to the Music section

    Write a detailed prompt: “Upbeat acoustic folk song with fingerpicked guitar, warm male vocals, cheerful and inviting with cozy vibes”

    Choose your desired number of variants and duration

    Generate your track, then edit individual sections and download in studio quality

Pro tip: Start broad (“upbeat pop song”) then get specific (“80s synths, female vocals, catchy summer hook”) for perfect results.

PRESENTED BY IBM

🚀 Lockheed Martin partners with IBM for AI

The Rundown: Lockheed Martin’s AI takeoff began with data cleanup — cutting its tools by 50% and replacing them with a single unified system that runs on IBM’s watsonx.data to drive faster, smarter engineering.

Learn how Lockheed Martin leveraged watsonx to:

    Help 10,000 engineers build scalable AI products in its AI Factory

    Respond to employee questions using AI, boosting accuracy by 20%

    Overcome data silos to create a single accessible, connected data environment

Read the full story.

AI & DRUG DISCOVERY

💊 Korean researchers’ AI designs cancer drugs

Image source: …

The Rundown: Researchers at the Korea Advanced Institute of Science & Technology (KAIST) developed BInD, a new diffusion model that designs optimal cancer drug candidates from scratch without any prior molecular data or training examples.

The details:

    The AI designs both the drug molecule and how it will attach to diseased proteins in one step, rather than creating and then testing in multiple iterations.

    BInD created drugs that target only cancer-causing protein mutations while leaving healthy versions alone, showing precision medicine capabilities.

    Unlike older AI systems that could only optimize for one criterion at a time, BInD ensures drugs are safe, stable, and possible to manufacture all at once.

    The model also learns from its successes, reusing winning strategies with a recycling technique to design better drugs without starting from scratch.

Why it matters: Drug discovery continues to be one of the biggest beneficiaries of AI acceleration. While the first AI-designed drugs are just starting to come to market, it feels like we’re only a few steps away from the floodgates opening on humanity-altering medicine advances designed by advanced AI models.

QUICK HITS

🛠️ Trending AI Tools

    📺 Hera - Create animated videos like CTAs, charts, and ads with text prompts in seconds*

    🧠 Grok 4 - xAI’s next-gen model, available for free to all users for a limited time

    🧊 Copilot 3D - Turn images into 3D models with one click

    ⚙️ Kombai - AI agent for frontend development

*Sponsored Listing

📰 Everything else in AI today

Chinese AI lab Z AI released GLM-4.5V, a new open-source visual reasoning model that achieves top scores on over 40 different benchmarks.

GitHub CEO Thomas Dohmke announced that he is leaving the company to pursue his own startup, with GitHub now being woven into Microsoft’s CoreAI department.

The U.S. government is reportedly set to enter into a new agreement with chipmakers Nvidia and AMD that would provide a 15% cut of chip sales to China.

Pika Labs introduced a new video model rolling out to its social app, with the ability to generate HD-quality outputs with lip-sync and audio in six seconds or less.

Alibaba announced that its Qwen3 models have been upgraded with ultra-long context capabilities of up to 1M tokens.

Anthropic unveiled new memory capabilities in Claude for Max, Team, and Enterprise users (excluding the Pro tier), giving the ability to reference previous chats.

COMMUNITY

🤝 Community AI workflows

Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.

Today’s workflow comes from reader Joel L. in Raleigh, NC:

"We are using AI to reactivate old leads. Our system will engage with old or dead customers, nurture the leads, and weed out leads that are not interested or not a good fit. Then the AI will book the appointment with the sales team. All the sales team has to do is show up to the warm booked call and close them.”

How are you using AI in your work? Tell us here.

🎓 Highlights: News, Guides & Events

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer—the humans behind The Rundown

Fish AI Reader

Fish AI Reader

AI辅助创作,多种专业模板,深度分析,高质量内容生成。从观点提取到深度思考,FishAI为您提供全方位的创作支持。新版本引入自定义参数,让您的创作更加个性化和精准。

FishAI

FishAI

鱼阅,AI 时代的下一个智能信息助手,助你摆脱信息焦虑

联系邮箱 441953276@qq.com

相关标签

Meta AI TRIBE 模型 OpenAI AI 竞赛 AI 药物设计
相关文章