9 of xAI's original 11 co-founders have now left the company. This week, two more departed after Musk acknowledged at an all-hands that xAI's coding tools are not competing with Claude Code or Codex. Macrohard, the project meant to build an AI agent capable of replacing white-collar work, is on pause. SpaceX and Tesla executives have been brought in to evaluate and cut staff. Of the founding team, two co-founders remain.

→ Coding tools are where AI labs generate real revenue, not chatbots. xAI is losing ground on the most commercially important front at the exact moment the market is beginning to consolidate around a few winners. With a SpaceX IPO anticipated and xAI now operating as part of that entity, a cash-burning unit that is trailing on product is not a story Musk can afford to let run.

After 12 years building Meta's AI research operation, Turing Award winner Yann LeCun has left to found AMI Labs in Paris. The startup just closed a $1.03B seed round, the largest in European startup history, at a $3.5B pre-money valuation. Backers include Nvidia, Bezos Expeditions, Cathay Innovation, and Eric Schmidt. AMI is not building an LLM. Instead, it is developing world models using the JEPA architecture, systems that learn how the physical world operates rather than predicting the next token. Year one is entirely R&D. No product. No revenue.

→ This is the first time one of the architects of the LLM era has put $1B behind the argument that the current paradigm has fundamental limits. With Nvidia and Bezos Expeditions in the cap table, this is no longer an academic position. If JEPA delivers, it opens AI that can reason and plan in physical environments: robotics, healthcare, industrial systems - domains where LLMs still hallucinate in ways that matter.

Gavriel Cohen built NanoClaw in 48 hours after discovering that OpenClaw had downloaded all of his personal WhatsApp messages to his machine in plain text, including messages with no connection to his work. NanoClaw is 500 lines of code, running on Apple container technology to sandbox all data access. After Karpathy tweeted about the project, the repo hit 22,000 GitHub stars. This week, Docker signed an agreement to integrate Docker Sandboxes into NanoClaw as the default runtime.

→ When a community builds a secure alternative in a weekend, it signals that the underlying problem is real and urgent. Bitsight found over 30,000 publicly exposed OpenClaw instances. Snyk reported 7.1% of ClawHub skills leak credentials. Security and sandboxing are becoming required infrastructure for any agent deployment at enterprise scale, not a feature to be added later.

Multiple Claude agents run in parallel to analyze GitHub pull requests, cross-check codebase logic, filter false positives, and rank bugs by contextual severity. Low false-positive rate and strong internal adoption at Anthropic from day one.

Processes text, images, video, audio, and documents into a single unified vector space within one API call. For RAG pipelines, this removes the need for separate embedding infrastructure per modality and opens new possibilities for how enterprise data gets indexed and retrieved.

A $400M Series D led by Georgian, with a16z, Coatue, Y Combinator, Databricks Ventures, and Accenture. Replit now has 40 million users and claims 85% Fortune 500 penetration. CEO Amjad Masad is targeting $1B ARR by end of 2026, up from $150M annualized today.

Processes text, images, video, audio, and documents into a single unified vector space within one API call. For RAG pipelines, this removes the need for separate embedding infrastructure per modality and opens new possibilities for how enterprise data gets indexed and retrieved.

You don’t need to be a coder to make AI work for you. Subscribe to Mindstream and get 200+ proven ideas showing how real people are using ChatGPT, Midjourney, and other tools to earn on the side.

From small wins to full-on ventures, this guide helps you turn AI skills into real results, without the overwhelm.

Get Your Free Guide

If you run Claude Code or OpenClaw, you have likely hit this: Claude Code stores OAuth refresh tokens in macOS Keychain but does not use them to refresh automatically. Every connected service expires independently. On a bad day, you are re-authenticating two or three times before lunch.

Composio routes authentication through managed OAuth flows with encrypted credential storage and automatic token refresh, SOC 2 and ISO 27001 compliant. Instead of pasting API keys into config files, you authenticate once and your agent orchestrates across 1,000+ services including Salesforce, Slack, GitHub, Google Workspace, Notion, and Datadog.

Setup takes three commands:

From there, a single Claude Code prompt like "Search YouTube for the top trending videos on [topic], identify content gaps, generate 5 ideas, and post to my Slack channel" runs end-to-end without stopping to ask for a credential.

Techzip note: The real leverage is not just the auth fix. Once authentication stops being friction, workflows start compounding. An agent can pull context from Datadog, cross-reference Notion, draft a response, and post to Slack without a human acting as a bridge between each step.

XSkill introduces a framework where AI agents extract two types of reusable knowledge from their own past runs: experiences (action-level tool selection) and skills (task-level workflow planning). Agents compare successful and failed rollouts to distill high-quality patterns, then retrieve and adapt those patterns when facing similar situations. Results across five benchmarks: average success rate up from 33.6% to 40.3% on Gemini-3-Flash; tool errors down from 29.9% to 16.3%, all without any parameter updates.

Why it made the cut: This is one of the first papers to quantify systematic improvement in agents that accumulate knowledge from their own trajectories. A related thread from @omarsar0 surfaces the sharpest open question in this space: the bottleneck for self-improving skills is not generation but evaluation. Agents can propose skill updates all day. What they cannot yet do reliably is judge whether an update actually made things better.

👉 Forward it to someone trying to keep up with AI.

👉 Read online: techzip.beehiiv.com

Techzip Newsletter
| Zipping what truly matters in the AI era.