Week 0: Friends Don't Let Friends Trust a Model Name

Nobody paid for this course

Rod teaches a module for Nebius Academy, AI Performance Engineering. That's a real, dated relationship. It's disclosed on this course's own page, not hidden in a footer: Friends of this site since 2026-08-20. Nobody paid for this course.

Three lessons already live elsewhere on this site were built straight out of that relationship, running real demos against Nebius Token Factory (Nebius's hosted inference API, the product this whole course is about) months before this course existed. This course holds itself to the exact same bar as every other course here: same lint floor, same obligation to name where the product breaks, same "friends don't get a softer review" rule stated to ourselves in writing before a word of this course was drafted.

What this course caught, in its own back yard

Here's the proof that rule is real, not a slogan. While researching this course, we checked whether the model names already used in this site's own existing Nebius lessons still match Nebius's current catalog. A "model ID" like meta-llama/Llama-3.3-70B-Instruct is the exact string a caller passes to Token Factory to pick which model answers the request. They don't match, and one of them is worse than "outdated": meta-llama/Llama-3.3-70B-Instruct, the exact model lesson 08 of this site's tools-memory-and-multi-agent-systems course uses in working code against api.tokenfactory.nebius.com, is discontinued August 31, 2026, per Nebius's own dated deprecation notice (docs.tokenfactory.nebius.com's August 2026 deprecation notice, fetched live and independently re-confirmed word-for-word during this course's own currency review), with a named replacement: Qwen/Qwen3-30B-A3B-Instruct-2507. That's not a hypothetical caution about catalogs moving fast in general. It's a dated fact about this site's own prior work, found by running the exact check this course teaches. The full receipt, including the replacement model and this site's own fix status, is in Lesson 3.

What actually renders on this course's own page

Tool — Checked against Nebius Token Factory 60+ open-model catalog, snapshot 2026-08-20.
Disclosure — Friends of this site since 2026-08-20. Nobody paid for this course.

Both lines, exactly as written, above the fold on this course's own page. Not a claim made here and left unverifiable. Go look.

What's not here

Two lessons were planned and cut. Fine-tuning was one of them. No first-party record exists of Rod having actually fine-tuned a model himself: a dedicated research pass searched this site's entire content/ tree for fine-tuning/distillation/LoRA/training terms, searched this program's own rulings and planning records for any mention of Rod fine-tuning, planned or executed, and checked his full public bio page for any fine-tuning-adjacent claim, before writing this course. It found none. Dedicated/enterprise endpoints was the other. Nobody here has stood one up, so it's cut too, matching the exact call already made on this site's sibling NVIDIA NIM course when it hit the same kind of gap.

Why cutting beats padding

This program's rule is simple: a lesson without a real receipt doesn't get written generically to fill a slot. A course that can't name its own gaps isn't earning the "friend" label. It's just advertising with a disclosure notice attached. Two cut lessons is the cost of that rule actually applying here, not an exception to it.

What you can do after Week 1

After lesson 1 you can explain what Token Factory actually trades against a self-hostable product like NVIDIA NIM. By lesson 3 you can work through whether an open model beats a frontier API for a specific workload, and query Token Factory's own live model list and deprecation notices yourself, the same habit that caught this course's own sibling lesson already sitting on a dated expiration.