There's a quite intense backlog of new stuff that hasn't made it to prod. (I would have told you in 2023 that we would have ex. switched to Mamba-like architectures in at least one leading model)
Broadly, it's probably unhelpful that:
- absolutely no one wants the PR of releasing a model that isn't
competitive with the latest peers
- absolutely everyone wants to release an incremental improvement, yesterday
- Entities with no PR constraint, and no revenue repurcussions when reallocating funds from surely-productive to experimental, don't show a significant improvement in results for the new things they try (I'm thinking of ex. Allen Institute)
Another odd property I can't quite wrap my head around is the battlefield is littered with corpses that eval okay-ish, and should have OOM increases in some areas (I'm thinking of RWKV, and how it should be faster at inference), and they're not really in the conversation either.
Makes me think either A) I'm getting old and don't really understand ML from a technical perspective anyway or B) hey, I 've been maintaining a llama.cpp wrapper that works on every platform for a year now, I should trust my instincts: the real story is UX is king and none of these things actually improve the experience of a user even if benchmarks are ~=.
Oh yeah, that's exactly what I was thinking of! Seems like it would be very useful for expert models with domains with more definite "edges" (if I'm understanding it right)
As for the fragmentation of progress, I guess that's just par the course for any tech with a such a heavy private/open source split. It would take a huge amount of work to trawl through this constant stream of 'breakthroughs' and put them all together.
For sure read Stephenson’s essay on path dependence; it lays out a lot of these economic and social dynamics. TLDR - we will need a major improvement to see something novel pick up steam most likely.
This reminds me of having to go into spotify's package files to track down their version of this animation (an animated svg of a bar chart) and kill it, because it would destroy performance on my PC so badly that it was affecting other programs, causing hitches and freezes.
The animation's still there — and my PC is better now, so it doesn't stutter — but I'm willing to bet it's still burning waaay too many watts, for something so trivial.
Picking up a bundle of loose pipes actually seems like a great benchmark for humanoid robots. Especially if they're not in a perfect pile. A full test could be something like grabbing all the pipes, from the floor, and putting them into a truck bed, in some (hopefully) sane fashion
I have my personal multimodal benchmark for physical robots.
You put a keyring with bunch of different keys in front of a robot and then instruct it pick it up and open a lock while you are describing which key is the correct one. Something like "Use the key with black plastic head and you need to put it in teeths facing down"
I have low hopes of this being possibe in the next 20 years. I hope I am still alive to witness if it ever happens.
You're telling me I've had 'Cry of the Chasmal Critter Chain' in my regular rotation for 20 years!?
I'd implore anyone with even a passing interest in VGM to find their favorite game on OC Remix and listen to some of the different remixes people have made. It's all very good stuff!
Blender's got a constraint solver for IK, right? How much spaghetti code do we need to add to give it a full CAD kernel? It already does everything else!
I've honestly wished I could use it to make vector graphics sometimes, but that also needs some of the basic elements of CAD (parallel edges, radius constraints etc). It's so close to parametric modeling too, with the mesh modifiers, drivers, and now geo-nodes.
Of course, I believe there are a few CAD plugins, but I've never used them, so I can't speak to their efficacy.
There's a bit more to it than that. There's an underlying library which can support solid modelling, but Blender has (or had) such an outdated version that it just wasn't possible.
Back in 2020 someone submitted code to get it working, in order to make solid modelling possible:
Unfortunately it looks like no official Blender developer ever took the time to review it, let alone merge it.
Super unfortunate, as it was only about 15 lines changed. Probably would have needed at least one revision though, as one of the changes was just commenting out some lines. That'd likely have needed to be a better conditional instead.
There is the excellent CAD sketcher plugin for Blender; this adds a basic 2D parametric/constraint based editor into your workflow, which can convert it's output into a mesh to integrate into your blender model. For more complicated models I typically make 2 or 3 2D constraint models, and use the blender boolean tools to combine this into the final 3D model.
I'm not an expert either, but it could be compared to bitmap graphics vs. vector graphics.
3D modelers like blender (or even OpenSCAD) work with a bunch of triangles - there is often not some higher level representation of the geometry. You could put a drill hole in a part, but it ends up as just a ton of triangles that approximate that drill hole, vs. a file format which semantically encodes "there is a cylindrical drill hole at this location, with this vector direction, and this radius".
That's what things like BRep (Boundary Representation) and STEP files give you is that semantic data which describes the part "here are the edges, faces, dimensions, etc.", vs. "here's a bunch of triangles, good luck machining this"
> but it could be compared to bitmap graphics vs. vector graphics.
This is very much how I internally understand it and explain it to people, yes!
It is a good analogy for e.g. why it's often a challenge to get something milled with a CNC when you only have an STL file.
STL is like a PNG line drawing: it can be high quality, but it's not describing the drawing. STEP is like SVG: it's more effort to render it, but it contains the instructions to draw it.
Isn't the problem with that analogy that there are things like NURBS which are pretty directly analogous to vectors (and isn't a surface a boundary?)
Edit: Along with the fact that blender has a lot of non destructive workflow steps (that usually get baked out into the "bitmap", to further your analogy)
It's not a perfect analogy, and Blender does have some parts which resemble the sort of data structures you'd use in a CAD modeler.
I think with enough plugins and customizations, you _could_ twist blender into something that resembles a CAD modeler, but it's really an uphill fight compared to selecting tools which were designed with that goal in mind from the start.
I've played with it a bit (I made a local nonsense social media thing, reading llm content), and its surprisingly easy to get text to display in sync with the audio. It works with voices that are installed on the users system.
