rao-v
rao-v on Hacker News
RL environments. RL environments as far as the eye can see
It’s important to understand that it had pytorch to actually run subsets of the weights. So its more like a dedicated person running the model to find the token it shouldn’t generate, then following the weight chain back up while calculating the NN outputs to find a place where if it changes some weights (again with a lot of testing via pytorch) the result is better.
Its not natively reading the weights in the Matrix “symbols scrolling on a screen” sense. Still bloody impressive and not an outcome I would have bet on!
Just inferencing a model this size is a bit of a feat let alone poisoning it!
Looks like this might actually be model weights of some kind, hope someone messes with it (safely) and reports back.
You have to understand that Strata and HyperQwen are giving you good looking tokens per second but are not giving you anything close to the bf16 model’s capability (they quantize, worse they quantize KV caches, they approximate expert routing etc.)
The models are so impressive that getting 70-80% of their capability can be helpful, but beware, you are using a significantly dumber model.
+1 I tried quite hard to make things work with tmux, zellij, and even some light custom stuff, but if you have multiple agents running on multiple boxes herdr just works unreasonably well. I really didn't want to like what seemed like a random mix of features, but they've stabilized it to a sensible capability set.
I’m assuming the bitcoin crash is some sort of “if Binance is vulnerable to AI powered (doesn’t have to be true) hacks, then all crypto is less a safe bet and thus worth less?”
Very nice! I'm actually running a autoresearch / modded-gpt comparision to try the complex math operation idea right now! I suspect it's too hard for small quick model training runs to figure out how to use them.
I'll put your knowledge graph nn idea into the mix of things to try.
Ask the least smart LLM you have access to (I tried gemini lite flash) to make a list of ways for processes in microVMs to use a sidechannel attack to talk to each other and you'll get quite a list (including obviously cache hits/misses). Heck a smart CS junior would come up with it. It's well studied!
I’d do this as an RL post training step maybe…
The model already ingests text, so you just to attach a tiny VLM head that reviews the image and read / writes scene continuity text tokens.
Once you have that you can probably just RL with scoring on continuity issues and eventually the VLM head will learn to update dense continuity tokens?
So many fun wierd experiments to run if you have some compute to mess with. There are competent open video diffusion models that you could test this concept easily enough.
I’ve always wondered why we don’t add a few frozen “layers” to models that perform complex computations 2/3rds deep in the model. Some set of steerable fft + some common vector math operations could be reusable units of capability that the model learns to leverage during training.
Given that models learn to use residual connections optimally, it can’t be hard for them to steer around the fixed computation layer when it’s not needed.
I’m sure giant models invest a few layers in reproducing these functions, but surely smaller models can benefit?
I think it’s more that you need really good latency guarantees for 6DOF to not feel terrible - it’ll be pretty hard to get that from a multitasking phone
The Oculus devices have and continue to have terrific support from Facebook
I mostly agree with all of this, but it doesn't really invalidate the horror / tragedy that it really was a lot of war / murder (?) alongside.
FWIW:
It looks like the Neanderthals had javelins as ranged weapons: https://www.mdpi.com/2571-550x/6/1/17
The paper I linked shows that we were in and out of Neanderthal territories (or vice versa) a bunch.
Punctuated equilibrium (is that still in fashion?) suggests that the inbreeding / limited population eras of humanity are when we got a lot of our special capabilities!
I don’t understand why these world models aren’t designed to be backed by an updatable world state. It could probably just lay out the world onetime and reference it as needed vs trying to invent from wholecloth via diffusion. Object persistence!
Oh I’m aware - my point is it doesn’t even have “sorta” interesting attempts and pads with Claude for paperwork. OpenAI’s equivalent will have a bunch of cool sounding projects (which I agree will struggle to have the promised wins - with frontier general purpose models anyway)
I’m honestly a little surprised by the sparsity of examples? One example of improving an existing therapeutic (Generate Biomedicines), one example of speeding up an existing clever play in food allergies (Divergence Labs), one fairly generic lead optimization play, and then we get … Claude helped write better paperwork and documentation for Enveda?!
I know Anthropic are brilliant at knowing when to focus where, but I genuinely expected them to be talking about 3-5x more serious scaled up drug discovery plays at this stage.
Is this not a focus, or did the big plays not want to be part of a roundup post?
There’s a tiny segment of folks (like me) who have been waiting a very long time for an 8ish inch OLED tablet. For some reason, until mid 2025, they were extremely rare, even in the Android multiverse. I’m excited because I read a ton with white text on black, and an OLED display (especially with this nanotexture stuff) will make a huge difference.
There are so many just barely sub-exponential function shapes between 2^(n / sqrt ln n) {this result} and 2^(log n) {polynomial time}!
I’m not sure what meaningfully changed except for the QCD / yang mills result (where leaders in the field are going to do a review and write up first).
The way to think of this is that (current) inferencing hardware has natural cut points of efficient scale (assuming some optimal architecture which at this point is almost always an MOE variant), so you always have an incentive to provide 3-4 model sizes (which nicely enough often means diff speed to answer also).
You are also rolling the training of your models by size and the lat est small model can sometimes beat an older model.
For a while we thought you just train the big model first then distill smaller models from it but since we've learnt the importance of on policy RL (basically the way a small model figures out a problem is very diff from a big model so it's best to RL optimize each model distinctly)
GPUs that were planned for cloud streaming are much more valuable now. It (unfortunately) makes sense for the only real cloud streaming player to significantly revisit pricing.
This is really interesting and fun!
Reactive optimizers could actually be helpful if you used them to dynamically tweak training data mix and use them to make choices about RL environments (if you recall from the Mimo training dashboards, it’s quite tricky to figure out how to value and chose RL traces during post training)
I mean Claude and codex both have their uses for me and I’m one of the deranged running my own inferencing box with Qwen Next so I need pi for it.
That’s like half a day to get to sensible useful defaults that everybody needs. Any chance you’d be up to push what you got to git?
What’s the best way to pull (ie archive) and search these across a few harnesses? I remember seeing a project or two trying to standardize them, but can’t find anything that seems to be a good sensible solution
Great resources thank you!
What I want to know is how good is it for moonlight streaming? The prev versions had odd latency issues.
EDIT: Brutal! Apparently it doesn’t support VRR or 120 hz at 4K. Surprising miss to not support HDMI 2.1 properly.
Further EDIT: Basically we are entering 2027 with no way to get a tv add on box that can stream 4K apps (Disney+ etc. need Widevine L1), 4K at 120hz (i.e. support HDMI 2.1 properly) and decode AV1. Just a bizarre gap in the market.