Meta’s powerful Llama 4 debut

Good morning, AI enthusiasts. Meta’s hotly-anticipated Llama 4 family is here — with a surprise weekend release debuting recent open-weights models with massive context windows and benchmark-beating performances.

With a 2T “Behemoth” still in training and claims of outperforming GPT-4.5, is that this release a real next-gen step forward? Or will user experience tell a unique story?

In today’s AI rundown:

• Meta launches Llama 4 model family

• Copilot’s recent personalization upgrades

• Unlock the ability of AI across your apps

• ‘AI 2027’ forecasts existential risks of ASI

• 4 recent AI tools & 4 job opportunities

LATEST DEVELOPMENTS

META

🦙 Meta launches Llama 4 model family

The Rundown: Meta just announced its Llama 4 family with multimodal capabilities and industry-leading context length—introducing recent open-weights Scout and Maverick models and previewing a 2T parameter Behemoth model still in training.

The main points:

• The 109B parameter Scout contains a 10M token context window and might run on a single H100 GPU, surpassing Gemma 3 and Mistral 3 on benchmarks.

• The 400B Maverick brings a 1M token context window and beats each GPT-4o and Gemini 2.0 Flash on key benchmarks while being more cost-efficient.

• Meta also previewed Llama 4 Behemoth, a 2T-parameter teacher model still in training that reportedly outperforms GPT-4.5, Claude 3.7, and Gemini 2.0 Pro.

• All models use a mixture-of-experts (MoE) architecture, where specific experts activate for every token, reducing computation needs and inference costs.

• Scout and Maverick can be found for immediate download and can be accessed via Meta AI in WhatsApp, Messenger, and Instagram.

Why it matters: After DeepSeek R1 disrupted the open-source market earlier this yr, Meta needed a powerful response. Llama 4 represents a step up in efficiency, massive context handling, and multimodal capabilities — but questions remain on whether the models truly feel next-level despite the strong benchmarks.

TOGETHER WITH CONCIERGE

🔑 Unlock the ability of AI across your apps

The Rundown: Wish your AI assistant could understand your context and take actions for you? Concierge solves this by connecting seamlessly to your apps (including Gmail, HubSpot, Jira, Linear, Slack, Attio, Sheets, Airtable, and more).

With Concierge, you possibly can:

• Immediately connect your favorite apps and get real-time answers based in your data

• Use top AI models (GPT, Claude, DeepSeek, etc.) without juggling multiple subscriptions

• Draft emails, update CRMs, create tickets, summarize notes, or do deep research — all with a single prompt

Try Concierge today and get AI to really give you the results you want.

MICROSOFT

🤖 Copilot’s recent personalization upgrades

The Rundown: Microsoft rolled out a significant upgrade to Copilot, with recent memory capabilities, web browsing actions, vision features, and a bunch of latest tools designed to integrate more deeply into users’ digital lives.

The main points:

• Copilot can now remember conversations and private details, creating individual profiles that learn preferences, routines, and necessary info.

• “Actions” enable Copilot to perform web tasks like booking reservations and buying tickets through partnerships with major retailers and services.

• Copilot Vision brings real-time camera integration to mobile devices, while a native Windows app also can now analyze on-screen content across apps.

• Other recent productivity features include Pages for organizing research and content, an AI podcast creator, and Deep Research for complex research tasks.

Why it matters: Microsoft is taking Copilot in an identical direction to other competing assistants, pushing for a more proactive and personalized platform. But much of the updates have a consumer focus — and it’s not clear that users will flock to Microsoft over Google, OpenAI, Meta, and others for those sorts of non-work experiences.

AI TRAINING

💻 Implement the newest coding solutions with Claude

The Rundown: On this tutorial, you’ll learn find out how to use Claude’s recent web search feature to seek out current coding solutions, library documentation, and best practices to unravel programming problems with the most recent information.

Step-by-step:

1. Head over to Claude and be certain web search is activated in your settings.

2. Describe your coding challenge clearly, including any specific requirements (e.g., “I would like to implement secure password hashing in Python that meets 2025 standards”).

3. Ask Claude to research and compare the several solutions found with pros and cons on your use case.

4. Request implementation help with code examples based on essentially the most current best practices discovered throughout the search.

Pro tip: When working with open-source libraries, you may also ask Claude to go looking for any known security vulnerabilities or deprecation notices before implementing the answer in your production code.

PRESENTED BY MONKS

🗣️ Uninterested in AI conversations that go nowhere?

The Rundown: Modeled after Monk’s newly appointed Chief AI Officer, Wesley ter Haar, WesleyBot is an AI-powered chatbot designed to assist technologists explore recent possibilities for advancing AI innovation.

Chat with WesleyBot to:

• Connect with a team driving real-world AI innovation

• Explore opportunities to collaborate with top talent across creative, data, media, and technology

• Help shape how AI transforms marketing, technology, and beyond

Chat with WesleyBot now.

AI RESEARCH

🔮 ‘AI 2027’ forecasts existential risks of ASI

The Rundown: Former OpenAI researcher Daniel Kokotajlo and the AI Futures Project published “AI 2027,” predicting advancement to superhuman AI inside two years, potentially triggering an intelligence explosion with consequences for humanity.

The main points:

• The report outlines a timeline starting with increasingly capable AI agents in 2025, evolving into superhuman coding systems after which full AGI by 2027.

• The paper details two scenarios: one where nations push ahead despite safety concerns, and one other where a slowdown enables higher safety measures.

• The authors project that superintelligence will achieve years of technological progress each week, resulting in domination of the worldwide economy by 2029.

• The scenarios highlight issues like geopolitical risks, AI’s deployment into military systems, and the necessity for understanding internal reasoning.

• Kokotajlo left OpenAI in 2024 and led the ‘Right to Warn’ open letter, speaking out against the AI labs’ lack of safety concerns and whistleblower protections.

Why it matters: While many dismiss AGI and ASI predictions, this forecast comes from researchers with direct, insider experience at leading AI labs. These scenarios suggest we could have only a temporary window to make sure AI stays controllable before it surpasses our abilities — making current safety and policy decisions critically necessary.

QUICK HITS

🛠️ Trending AI Tools

• 🕺 DreamActor-M1 – Turn images into full-body animations for motion capture

• 📦 ‘Buy for Me’ – Amazon AI agent that makes purchases from other web sites

• 🎥 Adobe Premiere Pro – Features like Generative Extend, Media Intelligence

• ✨ ActionKit – Add 1000+ integration actions to your AI agent

💼 AI Job Opportunities

• 📑 The Rundown – Account Manager

• 🛠️ Cresta – Senior IT Engineer

• 🗂️ Deepmind – Administrative Business Partner

• 🤝 Horizon3 – Mid-Market Sales Manager

📰 Every little thing else in AI today

Sam Altman revealed that OpenAI is changing its roadmap, with plans to release o3 and o4-mini in weeks and a “significantly better than originally thought” GPT-5 in months.

Midjourney rolled out V7, the corporate’s first major model update in a yr, featuring upgrades to image quality, prompt adherence, and a voice-capable Draft mode.

OpenAI has reportedly explored acquiring Jony Ive and Sam Altman’s AI hardware startup for over $500M, aiming to develop screenless AI-powered personal devices.

Microsoft showcased its game-generating Muse AI model’s capabilities with a playable (but highly limited) browser-based Quake II demo.

Anthropic Chief Science Officer Jared Kaplan said in a brand new interview that Claude 4 will launch within the “next six months or so.”

A federal judge rejected OpenAI’s motion to dismiss The NYT lawsuit, ruling the latter couldn’t have known about ChatGPT infringement before the product’s release.

COMMUNITY

🎥 Join our next live workshop

Join our next special workshop this Tuesday at 4 PM EST to learn find out how to create custom AI-powered workflows with Nick Huber, Ecosystems Lead at Poe.

RSVP here. Not a member? Join The Rundown University on a 14-day free trial.

🤝 Share The Rundown, get rewards

We’ll at all times keep this text 100% free. To support our work, consider sharing The Rundown with your folks, and we’ll send you more free goodies.

See you soon,

Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team