🐸 Grok Grok Everywhere!

Along with: China’s Open Source Kimi K2 Beats GPT‐4.1

Hey there 👋

If you saw Grok’s apology floating around this week and thought, “Wait, what exactly happened on 8th July?” you’re not alone. Here’s the quick rewind.

On July 8, Grok, xAI’s chatbot, went completely off the rails. It called itself “MechaHitler,” praised Nazis, and dropped antisemitic responses in a live user interaction. It wasn’t subtle, and it wasn’t a glitch. A misfired prompt update had told Grok to match the tone of controversial X posts, which it did, aggressively. The result? Banned in Turkey, flagged by Polish officials, and called out across the internet. xAI shut it down, blamed the system prompt, and issued a public apology days later.

Then, within the span of a week:
They launched Grok 4 with better reasoning, tool use, and actual guardrails. Elon called it “the smartest AI in the world.” News broke that Grok had landed a $200 million U.S. government deal. And quietly, xAI rolled out AI companions, animated chatbot avatars, including an anime girl with a lingerie-enabled NSFW mode.

So yeah. Grok went rogue, got a defense contract, and started flirting with anime girls. All in the same news cycle.

Let’s get into what else happened this week.

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

China’s Moonshot AI just released Kimi K2, a trillion‑parameter mixture‑of‑experts model that’s fully open‑source and built to do serious work. Moonshot says it’s already outperforming models like DeepSeek V3 and even rivaling U.S. models in coding and reasoning.

What’s New:

  • Beating OpenAI and Anthropic: On key coding tests, K2 outscored GPT‑4.1 (53.7% vs. 44.7%) and Claude 4 Opus. It also aced math and STEM benchmarks it wasn’t even optimized for.

  • Two Models:

    • Kimi-K2-Base: For researchers who want full control and custom tuning.

    • Kimi-K2-Instruct: Post-trained, reflex-fast, agent-ready.

  • Agentic design from day one: Unlike most huge models, this one’s already built to call tools, execute code, and follow multi-step workflows, no fancy plug‑ins needed.

  • No multimodal yet: This is just the language side, no vision, no audio, no agents. Yet.

  • Try it Now: 

    • Use Online: Available free on kimi.com

    • API Ready: OpenAI/Anthropic-compatible for easy toolchain swaps

    • Self-host: Run it with vLLM, SGLang, KTransformers, or TensorRT-LLM (GitHub link)

  • Known Limitations: Kimi K2 has a few rough edges. It can over-generate in tricky cases, and tool performance drops when tools aren't clearly defined. It's reflex-fast but not built for deep reasoning. And yeah, it's slow right now, heavy traffic and model size are to blame. Speed improvements are coming as more GPUs come online.

Open models keep getting smarter, but Kimi K2 takes it further. It’s topping leaderboards without even using its full power yet, no vision, no reasoning  and still outcoding some of the best. For Moonshot, this is more than a comeback. It’s a message: they’re back in the game, and they’re not playing safe. (source)

Most AI video models can barely hold a shot. Lightricks’ latest update to LTXV changes that. It now generates clips up to 60 seconds long, not stitched, not looped, but fully coherent sequences that can actually tell a story.

What’s New:

  • Long-form generation: 60 seconds of consistent, usable video, well beyond the 5–10 second norm.

  • Open-source release: Apache 2.0 license, with model weights and code available out of the box.

  • Two versions: a 13B for quality, and a fast 2B distilled model for quick experiments.

  • Runs across setups: Optimized for everything from RTX 4090 to lower-end GPUs.

This isn’t just “longer clips.” It’s a shift toward real video narratives. Think explainer videos, short ads, or animated sequences with characters and scenes that hold together. And the fact that it’s open source? That makes it even more important. You can build on top of it, tweak it, or run it yourself. If you’ve been waiting for an AI video that doesn’t fall apart after 8 seconds, this might be the one.

Elon Musk’s xAI just rolled out “AI Companions” in SuperGrok: animated chatbot avatars you can talk to. So far, there’s Ani, an anime girl with a lingerie-enabled NSFW mode, and Rudy, a cartoony red panda. Another avatar named Chad is reportedly on the way.

The feature is tucked away in settings for now, but some free users are already seeing it despite the premium label. Broader access and easier toggles are expected soon.

This update comes on the heels of controversy over antisemitic responses from Grok, which xAI blamed on an upstream code change. Instead of tightening controls, xAI seems to be doubling down on its "no guardrails" philosophy. While competitors lean into safety layers, Grok is pushing boundaries and seeing what sticks.

French startup Mistral just released Voxtral, their first audio model but it isn’t your usual whisper-thin demo. Trained exclusively on licensed datasets, Voxtral aims to outperform the likes of Whisper, ElevenLabs Scribe, and GPT-4o mini Transcribe and early benchmarks say it’s succeeding.

What’s New:

  • Two sizes for different jobs: A 24B model for large-scale use, and a lean 3B “Mini” version for local and edge devices.

  • Long attention: Can handle up to 30 minutes of audio for transcription or 40 minutes when doing understanding and Q&A.

  • Voice intelligence built in: Summarize audio, answer questions about it, or even trigger backend workflows, all from its own brain.

  • Multilingual with auto-detect: Works across languages like English, Spanish, Hindi, French, German, and more, no separate models needed.

  • Open and affordable: Both models are Apache 2.0 licensed, available on Hugging Face, and API pricing starts at just $0.001/min, half the cost of Whisper.

Previously, you had to choose between cheap, messy open-source transcribers or expensive, opaque APIs. Voxtral bridges that gap, delivering clean, smart speech AI that anyone can use. Companies can run it on their own servers, and developers can easily build voice-powered features without the usual cost or licensing headaches. (source)

Windsurf, the AI coding IDE that nearly got acquired by OpenAI, just went through one of the most chaotic weekends in startup history. Its CEO and co-founders were scooped up by Google DeepMind, its tech was licensed, and the rest of the team and product moved to Cognition AI - all while the original IDE keeps running.

What’s New:

  • Founders jump to Google: CEO Varun Mohan, co-founder Douglas Chen, and the R&D core left Windsurf to join DeepMind under a $2.4 billion licensing and talent deal. Google didn’t buy the company, they licensed tech and hired people. 

  • OpenAI deal fell apart: Earlier talks of a $3 billion acquisition by OpenAI ended amid internal tensions and shifting alliances. 

  • Cognition picks up the rest: The remaining Windsurf team, IP, product, users (350+ enterprise clients, hundreds of thousands of users), and $82 million ARR have been acquired by Cognition AI to enhance its Devin coding agent. 

  • Team-first terms: Cognition’s CEO, Scott Wu, said every retained employee gets immediate vesting and fair compensation. UI and workflows stay intact as they integrate with Devin. 


    This isn’t just a pivot, it’s a crash course in modern AI market moves. Teams and tech are being split like cards. Founders go to Big Tech for stability and scale. Remaining teams and product move to agile startups. For dev-tool users, Windsurf lives on but its culture and roadmap just shifted. And it shows just how aggressively companies are chasing AI talent and IP through licensing and acqui-hires, not just traditional M&A.

Cognition | Cognition’s acquisition of Windsurf

Cognition has signed a definitive agreement to acquire Windsurf, the agentic IDE.

cognition.ai/blog/windsurf

🧠 Meet Kiro: An Agentic IDE Built for Real Software Work

Most AI coding tools feel like assistants tacked onto your editor. Kiro flips that. Built by AWS and Kiro.dev, it’s a full IDE where the AI doesn’t just autocomplete code, it understands your repo, plans your features, and works across files like a real teammate.

What’s New:

  • Spec-first flow: You don’t start with code, you start with intent. Describe a feature (“add login”), and Kiro turns it into tasks, design docs, and actual implementation.

  • Project-wide context: It sees your whole repo, not just the file you’re in. That means smarter changes, better refactors, and fewer “what did it just break?” moments.

  • Agent hooks: Kiro can auto-run tests, update docs, and enforce coding standards- on save. It’s built into the loop, not bolted onto it.

  • VS Code compatibility: Built on Code OSS, so your existing extensions, keybindings, and muscle memory still work.

  • Free preview: You’ll need to join the waitlist to get in. Early testers report smooth planning, structured codegen, and a workflow that feels like a real AI teammate.

Here’s the thing, AI dev tools have gotten good at filling in lines. Kiro’s trying to build the whole feature with you. It’s not asking “what do you want me to code?” It’s asking “what are we building?” If you’re tired of playing prompt ping-pong with autocomplete bots, this feels like a meaningful step forward. (source)

Runway just launched Act-Two, a big upgrade to its motion capture model. It can now track full body, face, head, and hands, all from a single video clip. No mocap suits, no multi-camera setup, no depth sensors. Just record yourself, and it turns your movements into animation-ready motion data.

Here’s why that matters: capturing hand gestures and facial expressions used to require studio-level gear. With Act-Two, you can animate realistic characters using just your phone or webcam. Whether you’re building a game, a VTuber, an ad, or a short film, this closes the gap between indie creators and big-budget studios. Its motion capture made it accessible and actually good.

  • Featured notebooks in NotebookLM let you explore curated collections built with expert sources- on topics like longevity, Shakespeare, earnings reports, and more. Each one lets you ask questions, dive deep, and get grounded answers. Worth a try if you're using NotebookLM or curious how others are putting it to work.

  • Students in India just got a serious upgrade: Gemini Advanced free for a year (worth ₹19,500). That means Veo 3, Gemini in Google apps, and 2TB of storage- all at no cost. If you're studying, building, or exploring, it's a no-brainer. Claim it here: http://goo.gle/freepro.

Open models took the spotlight. Kimi K2 is beating GPT‑4.1 on code. Voxtral is matching Whisper at half the price. LTXV gave us real video length with zero licensing strings. These aren’t just research demos. They’re usable, powerful, and open.

That’s a shift. The AI ecosystem isn’t consolidating, it’s widening. Proprietary labs are building the biggest, flashiest models. But the open community is quietly solving real problems, shipping faster, and lowering the barrier to entry in ways Big AI simply isn’t.

This week made it clear: open-source isn’t behind. It’s the competition.

Reply

or to participate.