Today on the Brief: Everything Moved Closer To You

Slack Code launched today. You tag a coding agent into a conversation and it spins up a project-specific code channel. Everyone in that channel sees the same context the agent is working from — you audit code diffs as they're proposed, get live previews of HTML output, leave feedback the agent incorporates, and sign off before anything ships. It's available on every Slack plan, starting today.

One correction to what's circulating: OpenAI is not live in this yet — it's listed as "available soon." The four partners you can actually use right now are Claude, Devin, GitHub Copilot and Vercel.

Here's the part I find more interesting than the launch. This is not the first coding agent in Slack. Linear shipped agent coding sessions as live cards in Slack on August 11. Ara launched a Slack-native coding agent on August 16, pitched explicitly on the idea that the IDE is no longer where the work happens. There's also a Channels SDK from August 4 for bringing any agent into Slack or Teams.

Hao AI Lab at UCSD released FastMetal, bringing their FastWan video model family to Apple Silicon through an MLX/Metal runtime. Their own claim: a five-second, 480p video generated entirely on a Mac in about thirty seconds. No CUDA, no cloud, no discrete graphics card.

Let me put an honest number next to that. The same model family does that job in 1.8 seconds end-to-end on an RTX 5090. So a Mac is roughly seventeen times slower. That sounds bad until you remember what you're comparing — one of those requires an expensive GPU or a cloud bill that scales with every clip. The other is a laptop you already own, running offline, at zero marginal cost.

OpenAI confirmed it will continue offering Zero Data Retention for frontier models, and is previewing something called Private Safety Processing — designed to identify misuse patterns across related interactions without giving OpenAI personnel access to the underlying content. Broader rollout is expected in September. This pulled the biggest engagement of anything in today's brief.

If you do client work, this is the thing your clients keep asking about. Not "is it good," but "where does our stuff go." Being able to say the content is never retained and no human at the vendor can read it is a genuinely different conversation than the one most freelancers are having right now.

Liquid AI released quantized checkpoints of its LFM2.5 series — 230M, 350M, 1.2B and 2.6B parameters, small enough to run on a phone. They used the same quantization-aware distillation technique that made the Mac video release possible. That's two different labs, on the same day, shipping the same idea: squeeze the model so it runs on hardware you already own.

A correction to Sunday's episode. Stripe has now officially announced the OpenRouter acquisition at approximately $7.5B, with roughly $1.5B allocated to OpenRouter's founders. On August 17 we said "more than seven billion," based on the reporting that existed then. Now it's official, and it's a bit higher.

And a number I'm deliberately not calling closed. UK chip startup Fractile is reportedly in talks to raise around $600M at a ~$6.5B pre-money valuation, following an initial ~$250M deal with Anthropic. That is in talks — not closed. After what I got wrong on Tuesday about Etched, I'm not calling a round done until somebody announces it.

Here's what I'd actually do with this:

1. Try Slack Code this week on something small if you collaborate with anyone at all. It's on every plan and costs you nothing to test.
2. Install FastMetal and generate something tonight if you have an Apple Silicon Mac. Not for client work yet — just to feel what zero-cost iteration does to how you work.
3. Read OpenAI's Zero Data Retention terms now if you handle client data, so you can answer the question before a client asks it.
4. Watch the on-device model space seriously. Two labs shipped the same compression idea on the same day, and that's usually what a real trend looks like early.
5. Keep treating funding numbers as provisional until the company itself says them out loud.