CEO Xiao Hong and Chief Scientist Ji Yichao were summoned to Beijing's National Development and Reform Commission and told they cannot leave China. The investigation targets Manus' "Singapore bath" strategy: technology built in China, headquarters relocated to Singapore, state investors replaced with US venture capital before the Meta deal closed. Both founders are now separated from Meta's engineering teams. Integration is on hold.

So what? The offshore restructuring playbook for Chinese AI startups has become a compliance liability, not a legal shield. Beijing's message is substance over form: where the technology was built matters more than where the holding company is registered. Any investor or acquirer looking at a Chinese-origin AI startup needs to price this into deal terms before signing.

The Information reports that Apple's deal with Google goes further than previously known. Apple has full access to Gemini inside its own data centers, can edit the model directly, and can use Gemini's chain-of-thought reasoning to train smaller models that run offline on iPhones and iPads. Siri still relies on cloud-based Gemini for now. The distilled on-device models are not yet deployed. WWDC in June is expected to be the reveal.

So what? Google is not just selling an AI service. It is selling the ability to reproduce AI. Frontier models are becoming teachers that produce smaller, edge-deployable versions of themselves. For builders: on-device and offline-capable is shifting from a premium feature to a baseline expectation. Products built entirely on cloud inference should account for this.

The ARC Prize Foundation launched ARC-AGI-3 today. Unlike its predecessors, this benchmark drops AI agents into 135 interactive environments with no instructions and no stated goals. Gemini 3.1 Pro: 0.37%. GPT-5.4: 0.26%. Claude Opus 4.6: 0.25%. Grok-4.20: 0%. Every untrained human solved 100% on the first try. A $2M prize is live on Kaggle. All solutions must be open-sourced, no external API calls allowed.

So what? This is the first benchmark that measures learning efficiency directly, not just final scores. The AGI gap is no longer a talking point. It is a number. For anyone building AI agents: every sophisticated harness wrapped around today's models is the operator's intelligence doing the work, not the model's. That distinction matters when pitching autonomous systems to clients.

The compression algorithm optimizes the key-value cache that LLMs use to track conversation context. In testing on Nvidia's top server chips, TurboQuant achieved 6x memory reduction and 8x throughput improvement without degrading output quality. The paper will be presented at ICLR 2026 in April. AI memory stocks dropped 3-5% immediately after the announcement.

Faster responses, better task completion in noisy environments, and improved complex-instruction following. Available today for developers building real-time voice and vision agents.

The legal AI startup has now raised over $1B in total. Sequoia's decision to triple down in a single round is a strong signal about who they believe is winning the legal AI category. GIC co-led the round.

The startup is building a network of US open-source AI models positioned as a Western alternative to Chinese open-source AI. Investors describe it as the "DeepSeek of the West." It is currently in talks to raise $2.5B.

Most small businesses know AI exists but have no one to set it up for their actual processes. That gap is a product.

Here is how to build one:

Step 1. Pick a repetitive workflow your target client does manually: new client onboarding, weekly report generation, content repurposing, pre-call prospect research.

Step 2. Open Claude.ai, create a Project, and write Custom Instructions that define: Claude's role in this context, what inputs the user provides each time, and the exact output format expected.

Step 3. Test with three real examples from the client's actual work. Refine the instructions until the output is good enough to use with minimal editing.

Step 4. Package it as a service: a one-time setup fee plus a monthly retainer for maintenance and updates. A weekly report plugin, for example, is priced at $1,000-2,500 setup plus $300/month.

The math is straightforward: one plugin, ten clients, $2,000 setup plus $300/month each equals $56,000 in year one. No code required for most use cases.

Hate writing docs? Same.

Mintlify built something clever: swap "github.com" with "mintlify.com" in any public repo URL and get a fully structured, branded documentation site.

Under the hood, AI agents study your codebase before writing a single word. They scrape your README, pull brand colors, analyze your API surface, and build a structural plan first. The result? Docs that actually make sense, not the rambling, contradictory mess most AI generators spit out.

Parallel subagents then write each section simultaneously, slashing generation time nearly in half. A final validation sweep catches broken links and loose ends before you ever see it.

What used to take weeks of painful blank-page staring is now a few minutes of editing something that already exists.

Try it on any open-source project you love. You might be surprised how close to ready it already is.

Try Atlas today!

Friedman argues that the number that actually matters for frontier lab valuations is not the capability gap between closed and open models. It is the "monetizable spread": the subset of that gap that customers will actually pay a premium for. His central claim is that the monetizable spread is shrinking faster than the capability gap. If that is true, frontier labs are closer to utilities than platforms, and their current equity valuations reflect a moat that is quietly disappearing.

Why it made the cut: the clearest framework available for evaluating whether an AI company has a real competitive moat, beyond benchmark performance.

👉 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.