In partnership with
> WHAT MATTERS
TODAY’s 3 MOST IMPORTANT
Meta unveiled a roadmap of four MTIA chips (300, 400, 450, 500), built with Broadcom at a roughly six-month release cadence. MTIA 300 is already in production, MTIA 400 has cleared lab testing and is heading to deployment. MTIA 450 and 500 are optimized specifically for GenAI inference, with mass deployment targeted for 2027. Meta will continue buying chips from Nvidia and AMD in parallel.
Why this matters: When one of the world's largest GPU buyers starts building its own inference silicon, the cost structure of AI shifts. This signals that full dependence on Nvidia is giving way to a portfolio approach where custom chips handle the most repetitive, cost-sensitive workloads.
According to The Information, OpenAI is preparing to embed Sora into ChatGPT, letting users generate video inside the chat interface. The standalone Sora app lost 45% of its installs month-over-month in January 2026 and dropped out of the US App Store top 100. Sora will remain available as a standalone app. No official launch date has been announced.
Why this matters: This is a distribution move, not a new product. OpenAI is betting that bundling text, image, and video creation into a single interface retains users more effectively than maintaining separate apps. The same playbook worked when DALL-E was folded into ChatGPT.
Stockholm-based AI app builder Lovable confirmed it crossed $400M ARR in February. $100M of that was added in February alone, with a team of just 146 people.
Why this matters: The revenue-per-employee ratio here is nearly unprecedented at this scale. Lovable is proving out an entirely new company template: small headcount, low fixed costs, non-linear growth. It is the benchmark every AI-native startup is now trying to replicate.
> SIGNAL HEADLINES
Capture the shift
The latest release adds free 1M context models via OpenRouter, fixes GPT 5.4 stopping mid-response, integrates Gemini Embedding 2 for memory, and adds Go support.
Replit tripled its valuation in six months, raising a Series D at $9B after previously being valued at $3B. Alongside the raise, the company launched Agent 4, which runs parallel agents for design, auth, and backend simultaneously. Replit claims it helps ship production-ready software 10x faster.
A 120B parameter model with 12B active parameters using a hybrid MoE Mamba-Transformer architecture. Inference runs up to 2.2x faster than GPT-OSS-120B, with support for 1M token context. All checkpoints, datasets, and model recipes are released openly.
Cursor expanded its Marketplace with plugins from Atlassian, Datadog, GitLab, Glean, Hugging Face, and PlanetScale. Each plugin bundles MCP capabilities with usage instructions, giving agents better context to act more independently across the full developer stack.
Meta acqui-hired Moltbook, a social network built for AI agents, folding the team into Meta Superintelligence Labs. The real goal was recruiting people actively experimenting with agent ecosystems, not acquiring the product itself.
> TRY THIS TODAY
One practical AI trick you can use immediately
Stop writing prompts from scratch. Use a library of expert prompts instead.
Most people type something like "help me write an email" and get a generic response back. The difference between a mediocre AI output and a genuinely useful one almost always comes down to the prompt itself.
Promptschat just shipped a major upgrade: from a simple list to a full platform, now with 151K GitHub stars. Try it today: promptschat
Search for the role you want the AI to play: senior developer, financial advisor, career coach, cybersecurity expert → Copy the prompt and paste it into ChatGPT, Claude, or Gemini.
New features worth noting: version control for prompts like GitHub PRs, private prompts for internal tooling, and MCP server integration for Claude Code and Cursor. All prompts are CC0 licensed with zero usage restrictions.
> PRESENTED BY MINTLIFY
Ship Docs Your Team Is Actually Proud Of
Mintlify helps you create fast, beautiful docs that developers actually enjoy using. Write in markdown, sync with your repo, and deploy in minutes. Built-in components handle search, navigation, API references, and interactive examples out of the box, so you can focus on clear content instead of custom infrastructure.
Automatic versioning, analytics, and AI powered search make it easy to scale as your product grows. Your docs stay accurate automatically with AI-powered workflows with every pull request.
Whether you're a dev, technical writer, part of devrel, and beyond, Mintlify fits into the way you already work and helps your documentation keep pace with your product.
> TRY THIS TODAY
Instead of one SKILL.md, build a Skill Graph
Most people write a single SKILL.md to guide their AI agent. That works for simple tasks, but as the domain gets more complex, a single file stops being enough.
A Skill Graph is the next step: instead of one large file, you build a network of small, linked markdown files connected through wikilinks. Each file captures one complete technique or concept. The agent reads an index, understands the landscape, then navigates to the right branch for each situation.
The simplest way to start:
Create an index file describing the main topics in your domain,
Each topic points to its own markdown file via wikilinks,
Each file is a self-contained thought with a short YAML frontmatter description,
The agent reads the index first, follows links when needed, skips what is not relevant.
The result: instead of loading all context upfront, the agent navigates to exactly what the current situation requires. That is the difference between an agent that follows instructions and one that actually understands a domain.
> WORTH READING
I read it and you should too
A long-form conversation with Steve Yegge, former engineer at Google and Amazon, on how AI is fundamentally reshaping the software engineering craft. Yegge argues that vibe coding is not a passing trend but a structural shift in how software gets built, and makes the case for why developers need to adapt now rather than later.
Schmid explains how an AI agent can run hundreds of training experiments overnight without any human involvement. Karpathy used the approach to squeeze 11% more speed out of GPT-2 training. Shopify CEO Tobi Lütke trained a 0.8B model overnight that outscored his previous 1.6B model. The piece makes a clear case that autonomous model optimization is no longer theoretical.
Found this useful?
👉 Forward it to someone trying to keep up with AI.
👉 Read online: techzip.beehiiv.com
Tech.zip Newsletter
| Zipping what truly matters in the AI era.
