• AI Emergence
  • Posts
  • 🚀Grok 4 Just Landed. Elon Talked the Talk - Now It’s Test Time

🚀Grok 4 Just Landed. Elon Talked the Talk - Now It’s Test Time

Along with: DeepMind’s AI Drug Company Is Almost Ready to Test on Humans

Hey there đź‘‹

In the past few weeks, Meta has hired several top engineers from OpenAI and Apple, including Ruoming Pang, who led Apple’s in-house AI model efforts. Reports suggest some of these offers include massive compensation, with total packages going as high as $100 to $200 million over a few years. It’s all part of Meta’s new Superintelligence Lab, now staffed with big names like Alexandr Wang (Scale AI) and Nat Friedman (ex-GitHub).

OpenAI has responded with $4.4 billion in stock grants to keep its core team in place. Apple is reportedly bracing for more exits, too. None of the companies have confirmed the exact details, but one thing’s clear: top AI researchers are being courted hard, and the hiring game just got serious.

While the talent shifts behind the scenes, this week brought major updates from xAI, Perplexity, and DeepMind. Let’s get into it.

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Algorithm - Resources for learning

  • The Output - Our reflection

Table of Contents

After weeks of benchmark leaks and Elon teasing “the smartest AI in the world,” xAI has officially launched Grok 4 in two versions: the base Grok 4 and a souped-up Grok 4 Heavy. Both come with a serious compute upgrade and actual performance to back the talk.

What’s New:

  • Two Versions:

    • Grok 4: Strong general-purpose model, trained with tool usage in the loop

    • Grok 4 Heavy: Adds multi-agent reasoning , think multiple Groks debating, solving, and combining answers in parallel

  • Tool-Augmented Training: The model didn’t just learn facts, it was trained to use tools like calculators, search, and code execution while learning. That shows up in its reasoning.

  • Built on 200K GPUs: Powered by xAI’s internal “Colossus” supercluster. This is 10Ă— the compute of Grok 3, and 100Ă— more than Grok 2.

  • Voice Mode Gets a Glow-Up: 5 new voices, faster latency, more expression. Less robotic, more responsive.

  • Benchmark Flex:

    • Grok 4 Heavy scored 44.4% on “Humanity’s Last Exam” -  ahead of Gemini 2.5 Pro

    • Base Grok 4 scored 25.4% with tools turned off -  still solid

  • How to Use : $30/month for Grok 4 and $300/month for Grok 4 Heavy (with multi-agent reasoning), available via grok.x.ai or inside X (Twitter) for Premium+ users

  • Coming soon: AI coding model in August, a multimodal agent in September, and full video generation in October

For months, Grok felt like a side project with a meme engine. That’s over. With real benchmarks, tool use, and a heavyweight version built for deep reasoning, Grok 4 finally feels like it’s trying to solve real problems, not just reply to tweets. And with coding, agents, and video gen on the roadmap, xAI’s just getting started.

🧠 Meet Comet: The Browser That Thinks With You

After months of quiet, Perplexity is making noise again, this time with Comet, an AI-first browser built from the ground up. It’s not just search with AI on top, it’s your browser, co-piloted by an assistant that stays in context.

What’s New:

  • Context-aware Sidebar Assistant: Comet’s assistant can summarize articles, explain content, compare products, book meetings, send emails, all without needing to re-prompt or copy-paste. It works across tabs and understands what you’re doing as you go.

  • Chromium Core, AI-First UX: Built on Chromium, but designed around assistant workflows, not bookmarks and search bars. The AI isn’t an add-on, it’s the interface.

  • Invite-Only Rollout: Available for Perplexity Max users ($200/month), currently on Mac and Windows. Public launch is still TBD, but early invites are live. Join the waitlist here. 

We’ve seen a dozen “AI browser” demos, but most are just Chrome with a chatbot stuck on the side. Comet feels different,  more like a serious attempt at an AI-native browsing experience that doesn’t break your flow. (source)

After quietly working behind the scenes for years, Isomorphic Labs (the DeepMind spinout built on AlphaFold) says it's “very close” to testing its first AI-designed drugs on humans, likely starting with cancer trials later this year.

What’s New:

  • Trials are about to begin: Their lead candidate is headed into clinical trials in the coming weeks, marking one of the first times an AI-designed molecule will be tested on real patients.

  • AlphaFold 3 under the hood: The drugs are designed using AlphaFold 3’s deep modeling of protein interactions, not just folding predictions, but how proteins actually bind and behave in the body.

  • $600M funding + pharma deals: Isomorphic raised $600M in April, partnered with Novartis and Lilly, and is also running internal oncology programs.

This isn’t just research anymore, this is AI designing new drugs, getting them into labs, and pushing them into real-world testing. If it works, it could reshape how the entire industry approaches drug discovery. (source)

Moonvalley, a new startup founded by ex-DeepMind researchers, just launched Marey, an AI video model trained entirely on licensed content. No scraped data. No legal grey areas. Just full-scene control for people who actually want to direct.

What’s New:

  • Full-scene control: Adjust motion, camera angles, lighting, objects, and simulate handheld shots, shot-by-shot, not just prompts.

  • Physics-aware rendering: Motion stays grounded in real-world logic.

  • Background + character editing: Drop in custom actors, change tone or setting mid-shot.

  • Pricing: $14.99 for 100 credits, up to $149.99 for 1,000. Generates 5-second clips, more features coming soon.

Marey isn’t built for memes. It’s designed for directors. It’s already being used in real productions -  including a Carl Sagan documentary and a sci-fi feature backed by Natasha Lyonne’s studio. Moonvalley says it’s cutting indie production costs by up to 40%, with deeper lighting and character tools on the way. (source)

Hugging Face and Pollen Robotics just launched Reachy Mini, a small, expressive desktop robot made for building, learning, and coding. It moves, listens, talks, and you can program it however you want.

What You Get:

  • Starts at $299, ships as a DIY kit

  • Built-in camera, mic, speaker, motorized head + antennas

  • 15+ plug-and-play behaviors via Hugging Face

  • Test in real life or in simulation before hardware arrives

  • Full open-source stack- hardware, software, and SDKs

    Great for learning, hacking, or building real AI apps from your desk. (source)

Shortcut is a new AI agent built for Excel- and it’s fast. Really fast. It one-shots most spreadsheet tasks and solves Excel World Championship problems 10× faster than humans, scoring >80% in under 10 minutes.

What’s New:

  • Automates business tasks: Charts, models, scenario planning, insights- done in seconds.

  • Excel-native: Import, export, and edit files directly. No tool-switching.

  • Built-in smarts: Handles everything from financial projections to complex formulas.

  • Plain language input: You ask. It builds.

Still in early dev. Expect bugs with large files, formatting, or long chats. But even now? It’s a serious time-saver. Early preview is live- Join the waitlist at tryshortcut.ai.

HeyGen just dropped the first Creative OS for video- a full-stack AI agent that turns your ideas into polished content.

What’s New:

  • Upload a doc, clip, or prompt.

  • It finds the story, writes the script, and casts the actor.

  • Then it edits: pacing, motion, captions- end to end.

From ads and TikToks to product demos and short films, it handles the whole workflow like a human team- but faster. Join the waitlist to try it.

The Velvet Sundown racked up over a million Spotify listeners in weeks, complete with two albums, a moody backstory, and a verified artist profile. But there were no bandmates- just AI-generated music, faces, and lyrics. After denying accusations and insisting they were “real musicians,” the creators quietly updated the bio, revealing it was all a synthetic project designed to challenge ideas of authorship and identity. Deezer flagged their album as 100% AI, citing growing concerns as nearly 20% of uploads are now machine-generated. The stunt has reignited pressure on Spotify around disclosure, royalties, and the future of AI in music. (source)

  • In Episode 2 of the OpenAI Podcast, Mark Chen and Nick Turley sit down with Andrew Mayne to talk through how ChatGPT was built, how product decisions take shape inside OpenAI, and what’s coming next with agentic coding and multimodal assistants. Solid behind-the-scenes insight into the direction this tech is heading.

  • In this episode of the Joe Rogan Experience, Dr. Roman Yampolskiy dives into the risks of advanced AI, the idea of an “AI endgame,” and why unpredictability in intelligent systems is a serious concern. Yampolskiy is a leading voice in AI safety, and this conversation goes deep into the stuff most headlines skip.

  • A practical proposal by Anthropic on how we can hold the biggest AI labs accountable- without slowing innovation. This piece outlines a flexible transparency framework for frontier model developers, covering safety disclosures, secure development practices, and public documentation. Worth a read if you're tracking AI policy, governance, or safety.

  • This week’s spotlight: a free Generative AI Program packed with everything from the basics to advanced tools. You’ll go deep into LLMs, prompt engineering, RAG systems, AWS GenAI services, and how it all ties into real-world use cases. Whether you're new to GenAI or already building with it, this one’s worth your time.

Grok 4 was the breakthrough this week. Not just because it’s smarter but because it finally feels engineered with direction. Multi-agent reasoning, tool use, voice, and serious infrastructure behind it. It’s no longer just Elon’s chatbot. It’s a real contender in the stack.

But what made this week stand out wasn’t just Grok 4. It was how many different forms AI is starting to take, a browser that thinks with you, a drug pipeline shaped by protein models, an Excel agent that outpaces humans, a creative OS for video, a desktop robot for experimentation. None of these are gimmicks. They’re early signals of AI settling into specific roles, workflows, and industries.

The question is no longer “what can AI do?” It’s “where is it quietly becoming essential?”

This week gave us one big answer and a dozen smaller ones.

 đź‘‹See you next week!

Reply

or to participate.