AI Revolution
AI RevolutionMar 29
Tech

Anthropic’s New Claude MYTHOS Is The Most Powerful AI Ever!

13 min video5 key momentsWatch original
⚡
TL;DR

Anthropic's leaked documents reveal Claude Mythos, a new AI model tier above Opus with major cybersecurity capabilities that the company is rolling out cautiously to early enterprise customers due to offensive hacking risks.

Key Insights

1

cybersecurity risk management — Anthropic is treating its new Mythos model like a national security issue, limiting it to early enterprise access because internal testing showed it's better at finding and exploiting security vulnerabilities than any competing AI system.

2

brain activity prediction accuracy — Meta's Tribe V2 can predict how a human brain responds to video, audio, and text by training on real fMRI scans, and it works so well on new people that it sometimes predicts group brain responses better than actual individual recordings.

3

persistent task memory — Most AI agents today reset when you change tasks midstream, but Gwen Claw uses a three-layer memory system that keeps your broader context, work history, and live task state all running at once, letting it handle complex multi-step jobs without losing track.

4

self-improving agent design — Gwen Claw has a self-evolution loop where it logs failures and negative feedback, analyzes root causes, and improves over time instead of staying frozen after launch, which is a fundamental shift from how most agents work today.

5

open-source chip strategy — Alibaba's Schwantc950 chip is built on the open RISC-V architecture instead of proprietary ARM, giving the company freedom from licensing fees and better positioning in a world where US export restrictions limit access to Nvidia GPUs.

Deep Dive

Anthropic's Leaked Mythos Model Brings New Security Headaches

Anthropic accidentally left nearly 3,000 assets in a publicly accessible data cache, including draft documentation for a new model called Claude Mythos, also referred to internally as Capiara. This represents a new tier above the company's current top-tier Opus model. The leak revealed the model is already trained and undergoing testing with early enterprise customers. Anthropic confirmed it's working on a major step change in reasoning, coding, and cybersecurity capabilities. The company is being extremely cautious about rollout because internal testing shows Mythos is far ahead of competing models at finding and exploiting vulnerabilities. Anthropic has prior experience with this problem. A Chinese state-linked group once used Claude code in real attacks against 30 organizations including tech companies and financial institutions. Because of this risk, Anthropic is limiting initial access to enterprise customers and even planning an invite-only CEO retreat in the UK to give business leaders early exposure to unreleased capabilities.

Meta's Brain-Reading AI Predicts How Humans Respond to Content

Meta's Fair team introduced Tribe V2, an AI system that predicts how the human brain responds to video, audio, and language by combining Llama 3.2, VJEPA 2, and WAV2VEC 2.0 into a shared transformer model. The system was trained on 451.6 hours of actual fMRI brain scan data from 25 people and then evaluated on 1,117.7 hours of data from 720 people. It predicts activity across more than 20,000 cortical points with remarkable precision. What's striking is that Tribe V2 can make zero-shot predictions on new people without additional training, and in some cases those predictions better capture average group brain responses than real individual recordings. When given just one hour of new data, the model improves significantly and beats older linear methods by two to four times. Meta says the model can run virtual brain experiments in silico and naturally organized itself around five major brain networks including auditory, language, motion, default mode, and visual processing.

Gwen Claw Builds AI Agents That Actually Finish Tasks

Gwen Claw attacks one of AI's core problems. Most agents sound competent in conversation but lose track halfway through real work when priorities shift or tasks change. Gwen Claw uses a three-layer memory system with a stable identity layer, long-term background layer, and dynamic trajectory layer to keep your broader context, working history, and current task state all running simultaneously. It also uses context slimming to cut junk data while preserving important details, preventing the system from drowning in its own context or running up massive token costs. Unlike most agents that work in isolated demo environments, Gwen Claw operates directly in your local browser where it can use real login states, cookies, and cached information like a normal user would. The killer feature is self-evolution. Instead of staying frozen after launch, Gwen Claw logs failures and negative feedback, analyzes root causes, and turns those insights into targeted improvements. That means the agent gets smarter through repeated real use. It integrates with Huawei, ChatGPT, Telegram, WhatsApp, and web access, and supports private deployment for companies worried about data control.

Alibaba's New RISC-V Chip Sidesteps US Export Controls

Alibaba revealed the Schwantc950 CPU designed specifically for AI agent workloads. While most people focus on GPUs for training, Alibaba is betting that CPUs matter more than people think, especially for inference where agents run multi-step actions. The chip claims more than 30% performance improvement over mainstream competitors because it can be customized for specific inference patterns. The strategic move is that Schwantc950 uses the open RISC-V architecture instead of proprietary ARM, which means Alibaba avoids licensing fees and maintains independence from Western chip designs. This matters because Chinese companies face heavy US export restrictions on advanced Nvidia GPUs. Alibaba doesn't sell these chips to other companies. Instead it uses them internally to strengthen its own cloud AI services. Analysts say the real value isn't in immediate revenue but in building supply chain resilience, reducing costs, and maintaining control over AI computing power at a time when access is becoming harder to secure.

Takeaways

  • ✓Anthropic's Mythos model is so good at finding security vulnerabilities that the company is limiting it to approved enterprise customers and treating the release like a national security issue.
  • ✓Meta's brain-reading AI works so well that it can predict group brain responses better than individual scans, opening the door to virtual neuroscience experiments without live human subjects.
  • ✓AI agents are moving from static chatbot mode to self-improving systems that learn from failures and get better at complex multi-step work over time.
  • ✓China's domestic AI chip efforts are accelerating due to US export controls, with Alibaba choosing open-source RISC-V architecture to gain independence and reduce long-term costs.

Key moments

2:00Anthropic's accidental leak of Mythos model

“Some draft content, including what looks like a full blog post about a new model, got accidentally left in a publicly accessible data cache. And we're not talking about one or two files. There were nearly 3,000 assets sitting there.”

4:00Cybersecurity risks of new model

“The documents make it very clear that cyber security is a major concern. They literally state that this model is currently far ahead of any other AI model when it comes to cyber capabilities and they're worried about what that means in the real world.”

7:00Meta's brain prediction accuracy breakthrough

“One of the most surprising parts is how well it handles new people it has never seen before. Usually, if you want a model like this to work on a new subject, you'd expect to need a lot of fresh data. But Tribe V2 can make zero-shot predictions.”

9:00Gwen Claw's self-improvement capability

“Most agents today are basically fixed. If they fail, you get an error and move on. Gwen Claw adds a self-evolution loop where failures and negative user feedback get logged, analyzed for root causes, and turned into targeted improvements.”

12:00Alibaba's chip strategy against US restrictions

“Risk 5 is an open architecture unlike ARM's design model where companies pay royalties to use the blueprint. So choosing Risk 5 gives Alibaba more freedom and potentially lower costs.”

Get AI-powered video digests

Follow your favorite creators and get concise summaries delivered to your dashboard. Save hours every week.

Start for free