A little video that - explains the relationship between conventional LLMs and reasoning models (and agents), - philosophizes a about "from scratch" approaches, - and explains how to install Python & PyTorch requirements with uv.

And a link to the youtube version:

This playlist will keep me company this winter

You said almost all current models are reasoning models, yet they still have low/medium/high effort settings Can they still run as normal non-reasoning LLMs?

Right, reasoning models with different reasoning effort settings. Many/most also have a "None" setting

worth flagging one thing on the uv part since it bites people. uv add torch does not resolve to the same thing everywhere. linux pulls the cuda build straight off pypi, cu130 for torch 2.11 right now, while macos and windows get cpu only. so a pyproject that works on your machine can hand someone else a different accelerator with no error anywhere. pinning it means an explicit index entry against or whichever variant you want, plus sys_platform markers under tool.uv.sources. and torch-backend auto is still uv pip only, it does nothing under uv add or uv sync

Yeah, I added a uv.lock to help a little. But pinning too much can also cause issues. I.e., there are people who can't run the a specific version (due to CUDA compatibility) with specific drivers and so on. The default works for most people but not everyone, yes.

Great insight

Hey, what's upd Just learned something new from Sebastian Raschka about LLMs and conventional approaches!.

🐐

Do you expect the playlist to be around 8-10 videos like your previous book?

Good question. Let's see how it goes...

Fantastic explainer!

Legend!

nice that the GPU section treats Apple MPS as a first-class path instead of a footnote

Numerical stability-wise it's still tricky in some cases but has got a bit better over the years

fair - first-class as a path, still picky on numerics

This good stuff

uv for deps is the real vibe coder move

What that relationship did to comparison gets left out. A reasoning model and a conventional one can post identical accuracy on one set while spending an order of magnitude apart per answer, so the number is not comparable unless the budget is fixed. Does the video cover that?

Well yeah, it depends on the use case, there is no one size fits all.

The "from scratch" framing is where this gets philosophically tricky, since reasoning traces blur the line between training artifact and runtime behavior.

The from-scratch framing helps most where people conflate layers: model capability, harness behavior, and granted permissions get argued as one thing.

Torch wheels split by OS, reasoning models keep effort knobs; from-scratch still hits those

Nice mix, connecting the LLM to reasoning to agents story with an actual uv setup walkthrough is genuinely useful

keine ahnung von garnichts aber lamentieren yo

I take that for picking a model; it is the comparison I get stuck on. The MT-Bench grader sides with whichever answer runs longer in 68% of pairs, and reasoning models run longer by design, so length earns credit the reasoning did not. Does the video cover graders at all?

uv for PyTorch installs is a game changer. The reasoning model breakdown sounds fascinating too.

does the from scratch section argue for reimplementing things to learn them, or against reinventing infra that already works

actually both

The reasoning-vs-agent framing is the one most teams blur. Reasoning buys a better single answer; agents buy retries against a real environment. Different failure modes, different eval.

How does uv handle complex dependency conflicts compared to pip?

much faster, it uses a Rust-based solver

Putting uv beside reasoning models makes the path from watching an explanation to running code feel much shorter.

jack of all trades master of none but make it llms. how long until you cover conda too or is uv winning?

Using uv to set up PyTorch environments has completely transformed my daily python workflow

reasoning is a property of a single call, agency is a property of the loop around it. most of the disappointment i see sits in the second one and gets blamed on the first.

Sounds handy. Thank you for sharing.

Reasoning models are still doing next-token prediction with more structured scaffolding. The compression of world models into language has not changed.

exactly

From scratch is the interesting part. If language is how we compress world models, a different approach needs a different compression.

From-scratch can name LLM, reasoner, and agent in one sitting. That is a syllabus, not a niche. Three labels on one chalkboard are still one occupancy. #LLMs #Agents

There's a 43 minute video because agent ate the other two words.

Explaining conventional language models, reasoning models, and agents together helps separate architecture from workflow. The distinctions become much clearer when each layer is tied to an example of what it changes in practice.