Today on the Brief: Renting The Expert

Let's be straight about the date first: this is not from today. Public beta opened August 6, Alibaba showcased it in Hangzhou on the 10th, and their own blog post went up on the 13th. It's back in circulation because Alibaba just raised $10.2B in Hong Kong to fund exactly this kind of work. Worth covering because most people still haven't seen it — not because it's new.

Wan3.0 generates up to 30 seconds of video in a single pass, with audio made in that same pass, from text, images, audio, video — and actual documents. A slide deck, a Word doc, a spreadsheet: pdf/doc/xls/ppt/txt/md, one file or link, up to 100 MB and 50 pages. You hand it a pitch deck, it hands you back a video with a soundtrack.

Pricing is per second of output — $0.05 at 480p, $0.10 at 720p, $0.20 at 1080p. A full 30-second 1080p clip runs $6. Reference images, audio, documents and web pages aren't billed at all; reference video seconds are. Rate limit is 30 requests/min, 2 concurrent.

Two things people are repeating that aren't true. It does not do native 4K — the pricing table has three tiers and 4K isn't one of them. And the weights are not open. The Apache-2.0 weights being cited in 1.3B/14B variants belong to Wan 2.1, from February 2025. Recycled specs, repeated until they sounded like fact. Alibaba's last open-weight video flagship was Wan 2.2 (July 2025); everything from 2.5 onward is API-only.

One more, and it cuts against my own enthusiasm: Wan3.0 isn't on the Artificial Analysis Video Arena, so there's no independent measure of quality. Every quality claim right now is Alibaba's own.

Elevated errors flagged at 05:06 UTC, root cause identified around 05:27, remediation running past 06:42, everything resolved by roughly 08:30. It affected Mythos 5, Fable 5, Opus 5 and Opus 4.8 — and it hit claude.ai, the API, Claude Code and Cowork. Every surface at once.

Disclosure: I'm not neutral here. This show runs on Anthropic tooling. When Claude's down, my day gets worse too.

But the outage isn't the story. One bad morning is just infrastructure. Here's the story: this is the seventh disruption this month — the 5th, 12th, 13th, 16th, 18th, 20th, and today. Seven in twenty days is a pattern.

I wouldn't call that a reliability problem, exactly. I'd call it a dependency problem — and it's yours, not theirs. If your entire client workflow sits on one provider's API and that provider has a bad Tuesday, you don't have a technical issue. You have a business continuity issue, and your client doesn't care whose fault it was.

Harvard Business School runs an eight-week online bootcamp called Foundry. $699. You practice your pitch and an AI avatar of the faculty critiques it — simulated board meetings, simulated sales calls. There are live human sessions weekly too, but the feedback on your pitch comes from the avatar. One of them represents senior lecturer Jeff Bussgang.

Disclosure, and it's the whole reason I'm covering this: those avatars were built by HeyGen — the same platform I use to deliver this show. I'm not observing this from outside. I'm in the same business.

It's also not new. It rolled out in April 2026; it's trending because TechCrunch wrote it up on August 22.

Here's what I actually think. An avatar critiquing your pitch at two in the morning is genuinely useful, and cheaper than any equivalent access to that faculty. But be precise about what's being sold. You're not buying access to those professors. You're buying a model trained on their material, wearing their face. Those are different products, and the second shouldn't be priced like the first.

NVIDIA's Groq 3 LPX went into full production today — the inference chip from the roughly $20B Groq deal. 3,400 output tokens/sec on Gemma 4 31B in Artificial Analysis benchmarking, 256 LPU accelerators per rack, each carrying 500 MB SRAM and 150 TB/s SRAM bandwidth. It extends Vera Rubin NVL72, and Nebius is the first AI cloud to adopt it, with racks online later this year. You won't touch this directly; you'll feel it as agents that stop feeling laggy.

A London lab called Inherent has Faraday, a 27B agent that reproduces scientific papers and reportedly beats Claude Opus 4.8 and GPT-5.5 at it — evaluated on Replica, 310 tasks drawn from ~100 papers, off a Qwen base, with a ~$50M seed led by Index Ventures. Two corrections. It launched on August 14 — ten days old, not new. And Faraday calls GPT-5.5 Codex to do its actual coding. So it isn't a small model beating big ones. It's a small model directing a big one, and beating that big one used alone.

Reported, not confirmed to primary sources: Hugging Face is said to be exploring a sale around $13B, and Apple reportedly cut 200+ roles across Siri, Vision Pro and AI teams.

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

1. If you make video, spend $6 on Wan3.0 and feed it a real deck. Cheapest way to find out whether document-to-video actually works for you.
2. Write down the one dependency that would stop you delivering if it died for three hours — and know your fallback. Just know it.
3. If you sell expertise, look hard at Harvard's price point, because the floor under "access to an expert" just moved.
4. Treat "open weights" claims as unverified until you've seen the actual repo. Today's story is the case study.
5. Latency is about to stop being your excuse — so make sure the workflow is actually right.

Three of today's stories are the same story in different clothes. A model that turns your document into a finished video. An agent that reproduces a scientist's paper. A professor who exists as an avatar. In each one, something that used to require a specific person now runs without them.

And I'm delivering this through a digital avatar of my own face, so I won't pretend I'm above it. Here's what I'd say: the tool isn't the problem. What's worth protecting is whether a real person is still accountable for the judgment underneath it. My avatar is fine because I write and check every word it says. The moment nobody's behind it, you're not renting an expert. You're renting a costume.