Latest Innovations in AI: An Overview

Hello everyone, Keshav here! As Ben is out this week attending an AI Engineering event, I’ve taken the reins to bring you the latest updates.

Introducing Claude Mythos

Last week, a slip-up leaked news about Anthropic’s upcoming model, Claude Mythos. It’s officially in the works and boasts significant enhancements over Opus 4.6.

However, we won’t be able to access it immediately. The reason? Claude Mythos excels at identifying and exploiting software vulnerabilities. When attempting to generate exploits for Firefox, Opus achieved only two functional exploits after countless tries, while Mythos achieved an impressive 181.

This model has already uncovered several decades-old bugs in vital software projects such as OpenBSD (27 years old) and FFmpeg (16 years old).

Instead of making it widely available, Anthropic is permitting 12 companies to access a preview version through Project Glasswing. This initiative aims to discover vulnerabilities in critical software, with Anthropic dedicating $100 million in model usage credits and $4 million in donations to open-source security organizations as part of this effort.

Theo shared an insightful video on the topic, making an interesting comparison: “Mythos is to Opus what Opus is to Sonnet.”

Meta’s Latest Model: Muse Spark

In the wake of my recent tweet detailing the companies Meta has acquired in the past year with minimal output, Meta has unveiled their latest model – Muse Spark. From a quick glance, it appears to fit between Sonnet 4.6 and Opus 4.6. Although it’s not yet usable, API access is forthcoming, and there are promises of open-source developments (goodbye LLaMA).

Despite the online criticism regarding Meta’s non-innovative model release after substantial investment and a lengthy silence, I consider this a step forward. Have you noticed the significant improvements in Instagram’s search functionality recently? That’s all thanks to AI!

State of Frontier Models

Ethan Mollick provided an excellent recap on the current landscape of frontier models. The main players — Google, OpenAI, and Anthropic — continue to lead, while Meta is catching up. In contrast, xAI has lost momentum, and the best models from China remain lagging by 7-9 months.

New Tools and Applications

On another note, Factory’s desktop app is now officially out of beta. It offers a cloud computer feature, seamless use of other applications on your device, and effortless management of multiple Droid sessions.

Featured Innovations

• Chronicle: Cursor for slides. No more starting presentations from scratch; quickly transform ideas into stunning slides!

• OpenRouter Spawn: Effortlessly deploy OpenClaw and other agents to your preferred cloud environment. Compatible with all models on OpenRouter.

• Zapier’s SDK is now open to everyone. Enjoy programmatic access to all of Zapier’s functionalities during the beta phase. (See the documentation for details.)

• Kiro.dev: This spec-driven IDE from Amazon is reinstating its startup credits program for startups with teams of up to 30 members.

• Cogito: A Markdown editor for Mac. I’ve been using Clearly (recently updated) for easy edits and viewing of markdown files.

• Graphify: Create a queryable knowledge graph from any codebase or directory.

• Pi and Mario, the creator of Pi, are joining Earendil, a company founded by Flask’s creator. The core harness remains open-source, with upcoming features showcasing a mix of enterprise and fair source models (initially proprietary, soon to be open-source).

• Impeccable: Offers free design skills for coding agents, equipped with 21 commands to audit and fix common design errors.

• Superset and Builder 2.0: Two new user interfaces for operating parallel agents. Superset resembles Codex (terminal-first, with worktrees), while Builder has a kanban-style layout integrated with Slack and Jira.

• CSS Studio by Motion: Allows you to make on-the-fly design adjustments in your browser and pass those changes to your agent for implementation.

• S3 Files from AWS: This feature enables data to be stored as a file system, simplifying data usage for agents.

• Every is now running two parallel organizational charts — one for humans and another for each employee’s OpenClaw agents.

Recent Announcements

Today, Perplexity announced an initiative called the Billion Dollar Build. This 8-week competition will see teams utilizing Perplexity Computer to create a company aimed at reaching a billion-dollar valuation. Finalists will have the chance to secure up to $1 million from the Perplexity Fund, along with an additional $1 million in Computer credits.

For additional insights, Chris Tate officially introduced a new skill called Email Emulation, allowing for the testing of magic links and verification codes without sending actual emails. This can be achieved via the Resend SDK, facilitating retrieval from a local inbox and code extraction to complete authentication flows.