Karpathy's AI agent reality check

PLUS: Reverse-engineer winning ads to create high-quality AI videos

Good morning, AI enthusiasts. The “Year of AI Agents” prophecy came true, in certain respects — agents are everywhere, in every product, pitched by every company. But according to Andrej Karpathy, they don’t actually work.

The respected researcher just delivered a reality check, calling current agent output “slop” and saying the tech needs another decade to deliver on its hype-filled promises.

In today’s AI rundown:

  • Karpathy gives reality check on AI agents

  • Gemini gains live map grounding capabilities

  • Reverse-engineer winning ads to create high-quality AI videos

  • Anthropic co-founder: AI is a ‘real and mysterious creature’

  • 4 new AI tools, community workflows, and more

LATEST DEVELOPMENTS

OPENAI

Image source: The Dwarkesh Podcast

The Rundown: Former OpenAI and Tesla researcher Andrej Karpathy threw cold water on the AI agent hype during an interview with Dwarkesh Patel, projecting a decade-long timeline before autonomous AI systems can deliver on current promises.

The details:

  • Karpathy believes industry messaging is overselling current agentic coding capabilities that output “slop,” saying the models “aren’t there yet.”

  • He said that agents “just don't work” due to fundamental gaps like insufficient intelligence, multimodal limitations, and lack of continual learning.

  • Karpathy also called reinforcement learning “terrible” and “noise,” but it looks good because “everything we had before it is much worse.”

  • Elon Musk challenged Karpathy on X to compete against Grok 5, though Karpathy said he’d rather collaborate with the model than compete against it.

Why it matters: As one of the most respected researchers in AI, Karpathy’s words hold significant weight — and provide a major technical reality check to the “Year of the AI Agent” hype. But despite the harsh critiques, it’s also possible that systems that fail to impress a top mind are still massively productive for the other 99% of users.

TOGETHER WITH SAMSARA

The Rundown: When 24,000 bottles of Guy Fieri’s tequila vanished on a highway, it proved one thing: visibility saves value. Samsara’s Complete AI Safety Solution uses AI dash cams, in-cab alerts, and coaching to detect risky or unauthorized activity before losses happen.

AI helps fleets:

  • Detect unsafe or off-route behavior

  • Prevent theft and crashes in real time

  • Protect drivers, assets, and cargo

Access the report and see how AI prevents risk before it happens.

GOOGLE

Image source: Google Maps

The Rundown: Google just plugged Gemini into Maps, giving its AI direct access to real-world location data and letting developers tap the company's massive geographic intelligence trove.

The details:

  • The capability pulls from Google’s 250M venues worldwide, feeding Gemini current business hours, customer ratings, and venue specifics via API calls.

  • Developers can display interactive map widgets within their applications, preserving the Google Maps interface alongside AI-generated responses.

  • The system automatically IDs when geographic context enhances a query, retrieving relevant metadata without requiring triggers from users.

  • Pricing starts at $25 per thousand location-enhanced prompts, positioning the feature as a premium offering for enterprise apps.

Why it matters: This integration hands Google a competitive moat not easily replicable by rivals — the infusion of its already widely used mapping infrastructure into its advanced AI models. While the steep pricing may lend itself to more enterprise-focused needs, the combo opens a new level of location-aware AI-powered apps.

AI TRAINING

The Rundown: Create professional marketing videos with Sora 2 by analyzing successful UGC ads, converting them to JSON formats, and generating polished AI videos that match proven patterns.

Step-by-step:

  1. Search TikTok/Instagram for winning UGC ads in your niche, download videos with strong hooks and clear product demos you want to emulate

  2. Upload to Google AI Studio with Gemini 2.5 Pro and prompt: “Analyze this video shot by shot. Return strict JSON with: scene, description, camera_angles, transitions, voice transcript, on-screen text. Constrain to 15 seconds”

  3. Paste JSON in ChatGPT to adapt: “Take this JSON and adapt to [your niche/product]. Keep camera angles and pacing. Replace script with [your messaging]. Output Sora-compatible JSON”

  4. Paste final JSON into Sora, generate, and review for script completion, logo fidelity, readable text, and clean transitions

  5. Clean up with free tools: remove watermark, enhance speech with Adobe Podcast, upscale via Replicate, and strip AI metadata using video remixer

Pro Tip: Regardless if you’re on the free or paid plan for Sora, I’d recommend cleaning up your video in order to stand out on the “For You page,” as our feeds are dominated by AI video slop that can easily be identified as AI and not marketing-grade content.

PRESENTED BY IBM

The Rundown: IBM Network Intelligence introduces a new human-AI model that blends analytical AI with “reasoning” AI — a dual approach developed to help enterprises filter noise, identify issues, and scale operations seamlessly.

With IBM Network Intelligence, you’ll experience:

  • Fast, accurate insights

  • AI that scales while humans guide strategy

  • Less tool bloat, more efficiency

AI RESEARCH

Image source: Reve / The Rundown

The Rundown: Anthropic co-founder Jack Clark published a new essay titled “Technological Optimism and Appropriate Fear,” describing modern AI systems as mysterious entities exhibiting unexpected self-awareness rather than predictable tools.

The details:

  • Clark cautioned against considering AI just a tool, saying “what we are dealing with is a real and mysterious creature, not a simple and predictable machine.”

  • He said the recently launched Sonnet 4.5’s situational awareness has grown, and now “acts as though it is aware it is a tool.”

  • Despite calling himself a “technology optimist,” Clark said he's “deeply afraid” — especially of AI models helping design their own successors.

  • Clark believes AI firms need to “do a better job of listening” to concerns from the public and expand the conversation beyond tech elites.

Why it matters: Anthropic has been one of the few frontier labs truly considering the idea of AI as a “being” instead of simply a machine, and its co-founder’s latest words only reaffirm that — though hearing words like “deeply afraid” and “mysterious creatures” from a frontier leader likely won’t help reassure AI safety advocates.

QUICK HITS

  • 🗂️ Skills - Claude’s new folder-based system for loading new capabilities

  • ⚙️ SWE-grep - Cognition’s fast, agentic coding model

  • 🎥 Sora 2 - OpenAI’s social AI video platform, with new Storyboards and extended video lengths

  • 🎬 Veo 3.1 - Google's new upgraded AI video model

Uber is launching “digital tasks” in its driver app, letting U.S. drivers earn extra cash by completing simple AI training work like uploading menus or recording audio samples.

Elon Musk revealed that he estimates the probability of xAI’s upcoming Grok 5 model achieving AGI is “10% and rising.”

OpenAI announced the pause of video generations featuring Martin Luther King Jr. on Sora following a request from the King estate.

Anthrogen unveiled Odyssey, a 102B parameter protein language model that uses a new “Consensus” architecture to design and optimize proteins more efficiently than traditional approaches.

Meta announced new parental controls coming in 2026 that will let parents block teens' chats with AI characters on Instagram and monitor conversation topics.

COMMUNITY

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 Barry in Australia:

“I’m a performance marketing manager for a global brand, and started using a new workflow for reporting. I use Gemini/ChatGPT/Claude with a fixed prompt with my specific needs and also provided my past framework on insights. I provide 3 different Google analytics screenshots along with some Google ads data and run the prompt. I then use that insights and paste it in a Notion database with the week. On a quarterly basis I ask Notion AI to find trends and insights based on my weekly insights. I've been able to find so many helpful and informative insights with this workflow, both on a weekly and quarterly basis.”

How do you use AI? Tell us here.

That's it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve The Rundown experience for you.

Login or Subscribe to participate in polls.

See you soon,

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

Reply

or to participate.