rao-v

9 ms·

151.
by rao-v 7mo ago
This is nice, but it's not always the case that +3 indent is the right solution (e.g. if I'm copying already indented code it may be over indented). It's basically a non problem in most other languages, and a IDE formatter ho
152.
by rao-v 7mo ago
I messed with this at one point and gave up when I realized every device would have a permanent externally addressable IP within a block that is basically linked to me (good luck trying to change your IPv6 /48 every month or whatever y
153.
by rao-v 7mo ago
Is there a decent open video embedding model out there? I’d love to play with this without uploading video.
154.
by rao-v 7mo ago
I put together a little script to search for and list installed litellm versions on my systems here: https://github.com/kinchahoy/uvpowered-tools/blob/main/inven... It's very much not production gra
155.
by rao-v 7mo ago
The problem is that if you copy random code from the internet it cannot figure out the right indentation level - whitespace has meaning in python. What IDE can automagically handle this?
156.
by rao-v 7mo ago
Is there a non-tranformer based entity extraction solution that's not brittle? My understanding is that the cutting edge in entity extraction (e.g. spaCy) is just small BERT models, which rock for certain things, but don't have th
157.
by rao-v 7mo ago
Exactly. I genuinely do not understand how any significant user of python can handle white space delimitation. You cannot copy or paste anything without busywork, your IDE or formatter dare not help you till you resolve the ambiguity. One d
158.
by rao-v 7mo ago
Will do thanks - any standout features or clever things for me to look out for?
159.
by rao-v 7mo ago
I tried crush when it first came out - the vibes were fun but it didn’t seem to be particularly good even vs aider. Is it better now?
160.
by rao-v 7mo ago
Riding a motorcycle or even a bicycle around Waymos feels surprisingly safer. You can reliably predict so many things about how it will behave and to an extent even its traffic calming effect on other cars.
161.
by rao-v 7mo ago
I ended up on this journey using Dockge. Inoffensive and you can stick your compose files in a directory and manage with git vs. Portainer’s attempt to hide them.
162.
by rao-v 7mo ago
I generally agree but I value projects like this because there are smaller scale environments where many of these fallacies are perfectly fine working hypotheses. My home lab or a low volume, low 9s service etc.
163.
by rao-v 7mo ago
I’d love to believe this is real, but I’m pretty sure you will lose performance on a “fair” mix of tasks, even after fine tuning. I know multiple teams have explored recurrent layers (great for limited VRAM) but I don’t think it’s ever been
164.
by rao-v 8mo ago
Taking a step back, it seems fairly clear that wherever you set the bar, it should be possible to automate a system that reads code, generates some sort of intermediate representation at the acceptable level of abstraction and then regenera
165.
by rao-v 8mo ago
Agree re: progressive growing In terms of sub structure - in the old days of Core Wars randomly scattering bits of code that did things could pay off. I’m imagining similar things for LLMs - just set 10% of weights as specific known structu
166.
by rao-v 8mo ago
I’m genuinely surprised to see this not discussed more by the FOSS community. There are so many ways to blow past the GPL now: 1. File by file rewrite by AI (“change functions and vars a bit”) 2. One LLM writes a diff language (or pseudo co
167.
by rao-v 8mo ago
+1 I’ve always had the feeling that training from randomly initialized weights without seeding some substructure is unnecessarily slowing LLM training. Similarly I’m always surprised that we don’t start by training a small set of layers, st
168.
by rao-v 8mo ago
It’s interesting that you invest in mouse movements vs just targeting a click at X in Y milliseconds. CAD and video games are of course a great reason for this, but I wonder how much typical tool use can be modeled by just next click events
169.
by rao-v 8mo ago
+1 this does seem to be a genuine attempt to actually build an interpretable model, so nice work! Having said that, I worry that you run into Illusion of Conscious issues where the model changes attrition from “sandbagging” to “unctuous” wh
170.
by rao-v 8mo ago
So llama.cpp currently statically puts overflow MoE experts in RAM and inferences them on CPU, so you get a mix of GPU + CPU inferencing. You are rooflined by RAM->CPU bandwidth + CPU compute. With good predictability of MoE, you might s
171.
by rao-v 8mo ago
I'd love any keywords to search for to find active research on this topic!
172.
by rao-v 8mo ago
We’re probably a year from self hostable video LLM models that can identify sexual content etc. with high sensitivity (but probably poor specificity)
173.
by rao-v 8mo ago
I think your analysis is right this would make sense mostly for the 30B-3A style models that are mostly for edge / hobbyist use, where context length is precious so nobody is batching. Given that experts live per layer I dont think it
174.
by rao-v 8mo ago
Yeah I’ve often wondered why folks aren’t training two tier MoEs for VRAM + RAM. We already have designs for shared experts so it cannot be hard to implement a router that allocated 10x or 100x as often to “core” experts vs the “nice to hav
175.
by rao-v 8mo ago
I’d really like to see this optimized for the 50-120B parameter open source models that are local viable (gpt-oss-120b, qwen3-80b-3a etc.). For them I think it would be optimal to provide a tag per function and trust the llm to rewrite the
176.
by rao-v 8mo ago
Are we saying line of sights are not symmetric? Why not?
177.
by rao-v 9mo ago
Skills aren’t the right fulcrum to move LLMs on something as generic as well known language best practices. Trust the model owner to post-training on that. Use skills for more specific things (idiosyncratic patterns, specific library docs,
178.
by rao-v 9mo ago
I think it maybe time for us to think about what the sensible version of these capabilities are. Short term hacky tricks: 1. Throw away accounts - make a spare account with no credit card for airbnb, resy etc. 2. Use read only when it'
179.
by rao-v 9mo ago
Sub-agents are typically one of the major models but with a specific and limited context + prompt. I’m talking about a small fast model focused on purely curating the skills / MCPs / files to provide to the main model before it ki
180.
by rao-v 9mo ago
In a month or three we’ll have the sensible approach, which is smaller cheaper fast models optimized for looking at a query and identifying which skills / context to provide in full to the main model. It’s really silly to waste big mod

More ›