╔═════════════════════════════════════════════════════════════╗
║ ║
║ ▀█▀ ▄▀█ █▄▀ █▀▀ █ ▀█▀ █▀▀ █▀█ █▀█ █▀▄▀█ ║
║ █ █▀█ █ █ ██▄ █ █ █▀ █▀▄ █▄█ █ ▀ █ ║
║ ║
║ ▀█▀ █ █ █▀▀ ▀█▀ █▀█ █▀█ ║
║ █ █▀█ ██▄ █ █▄█ █▀▀ ║
║ ║
╚═════════════════════════════════════════════════════════════╝
// Tech news at terminal velocity
cat 2026-05-07.md
$ cat TLDR.md
▸ • Snap and Perplexity quietly kill their $400M AI search integration deal.
▸ • Google shuts down Project Mariner, folding its autonomous web agent dreams into Gemini.
▸ • Zyphra drops ZAYA1-8B, a tiny open-source model trained on AMD that punches way above its weight class.
Snap and Perplexity have 'amicably ended' their $400M deal that would have brought the AI search engine directly into Snapchat. Turns out, integrating a massive AI search engine into a disappearing photo app might have been more complicated than a handshake deal. Snap's Q1 guidance now assumes zero contribution from the startup.
Google has quietly pulled the plug on Project Mariner, its experimental AI browser agent designed to autonomously surf the web and book trips for you. The dream isn't entirely dead, though—Google is cannibalizing the tech and folding its agentic capabilities into Gemini and Chrome's 'auto-browse' features.
Zyphra just released ZAYA1-8B, a Mixture-of-Experts model with only 760M active parameters that somehow matches DeepSeek-R1 on math and coding benchmarks. The real kicker? It was trained entirely on AMD Instinct MI300X GPUs, proving that the NVIDIA monopoly might actually have some cracks in its armor.
Nintendo just surprise-announced the first new Star Fox game in a decade, slated for the upcoming Switch 2. Fox McCloud is officially back, and our thumbs are already preparing for the inevitable blister-inducing boss fights.
The mad scientists at Unsloth teamed up with NVIDIA to squeeze even more juice out of LLM training. By caching packed sequence metadata and using double-buffered async gradient checkpointing, they've managed to speed up training by roughly 25% with zero loss in accuracy. It's a masterclass in low-level optimization.
On May 5th, the registry operator for the .de TLD accidentally published broken DNSSEC signatures, effectively wiping millions of German domains off the internet. Cloudflare's postmortem dives into how their 1.1.1.1 resolver handled the chaos and how 'serve stale' features cushioned the blow for users.
How do you test an AI coding agent when there isn't a single 'correct' answer? GitHub engineers break down their approach to building a 'Trust Layer' for Copilot using dominatory analysis. It's a fascinating look at moving beyond brittle test scripts and black-box judgments in the era of generative AI.
A new open-source scaffolding tool called Agent-Harness-Kit just dropped, offering a provider-agnostic way to build multi-agent workflows. It supports the Model Context Protocol (MCP) out of the box, making it easier to string together complex AI behaviors without getting locked into a single ecosystem.
Unsloth launched a new API inference endpoint that gives local models like Qwen and Gemma advanced agentic superpowers. It integrates seamlessly with tools like Claude Code and OpenAI Codex, bringing self-healing tool calling and bash execution to your local rig.
Anthropic just raised the usage limits for Claude Code, and they're crediting a massive new deal with SpaceX for the infrastructure boost. When you're helping build rockets, apparently you get the good servers.
Privacy advocacy group Noyb is arguing that LinkedIn's feature showing who viewed your profile shouldn't be locked behind a Premium paywall. They claim that under GDPR, users have a fundamental right to know who is accessing their data. Time to see who's been lurking.
The AI boom is incredibly thirsty for power, and Taiwan is feeling the crunch. TSMC is now heavily backing wind power and other renewables to keep their massive chip foundries running without browning out the entire island.