ilaksh

The $2000 version is 10 X better.

Right and they already said they have a massive backlog and are trying to increase by 500%.

I think there is still a little bit of room for green screen. At least maybe until Seedance 3.0 in a few months.

But it is interesting how video generation turned the corner from several months ago to now. I think the next AI project they do probably has to be with a generative tool even though they may be very hesitant to do that because they don't want to emphasize AI models that threaten their own jobs.

But I remember when they did the Project Greenlight thing which resulted in a massive list of desired projects from subscribers that they could not possibly make a dent in.

I think in the next several months they may be able to revisit that or at least some of the projects that had to be shelved due to lack of time. To make that possible they would leverage the latest generative AI models and workflows, possibly a few that haven't been released yet. But we are rapidly approaching the point where the models plus tooling is good enough to replace 95% of the technical work, and it's down to creativity. I think it's only a couple of years before that progress bar really maxed out.

You can start with Artificial Analysis benchmarks but, then you need to test in the real tasks anyway.

From my testing, the problem Cerebras has is they need to give people a taste to get new customers in, but since it requires so much hardware to get that speed and they don't have OpenAI-type money (< $3 billion compared to over $100 billion) and it's just incredibly hard regardless to manufacture that many of these types of crazy giant wafer nodes, there is just no way for them to keep up with the demand.

So what they do is they have these cheap plans that let anyone get a little taste of the incredible speed, but they aren't reliable at all, because they have to share limited resources dedicated to sales. I believe that is really the main purpose of like the $50/month or whatever plans -- part of the sales funnel. The cheap plans compete directly with resources they need for their "real" paying enterprise customers.

The "real" plans start at like $1500 per month, and I assume those have reasonable service level reliability.

But actually I believe those $1500/month plans are totally worth it for some applications.

The robotics structure, dexterity and sensors are already 90-95% there in leading edge humanoids. What makes it able to be a plumber or not are the machine learning models, which are usually built with transformers (VLAs -- vision language actions models), often include frontier VLMs or models with the same architecture as a top level control, and use the same or very similar types of reinforcement learning for training.

GPT-2 was nowhere near being able to create web pages. It was largely incoherent. But three years later gpt-3.5 and similar were able to create simple web pages and programs and became very popular. And a few years after that, only a small portion of software engineers generally write code by hand still.

Humanoid robotics may progress faster than that because they are building on existing AI. But it will be capable of some plumbing work within a few years.

Look all of the progress in robotics. Companies like Physical Intelligence etc. There has been massive progress in all sorts of areas from picking strawberries to assembling kits, folding clothes, etc.

Huge progress in spatial understanding with frontier models now building architecture and complex mechanical systems in Blender.

Plumbing is coming within a few years.

Well for AI it's more like training than product design. I have switched from software engineering to AI training for software engineering (partly because honestly it is just much easier now to get that kind of gig work versus the programming contracts I was getting before).

They have a ton of tasks that pop up on my dashboard for all sorts of industries that I know nothing about. I think they are ramping up in non-software areas.

Since it's brand new there are going to be a lot of subpar solutions on the market. When SSDs first came out, there were only a few models that were actually reliable.

It is relatively easy to create and deploy a transcription model. It's pretty hard to find one that is really accurate and to configure it with a good microphone setup and pair that with a fast but also intelligent LLM to clean things up. If you execute any of that poorly or pick a cheap not very smart model for privacy and or cost saving you could easily end up with nonsense.

Great. I wonder.. is this based on Mochi or PersonaPlex? Also, how big is the model, and do we use it?

Cerebras can do like 4000 tokens per second on gpt oss 120b with it's new chip and around 1000 tokens per second on frontier models using its new chip.

Were you talking about Talaas 8b model? They will probably do something like Bonsai or Qwen 27 B next I hope. But it makes might be an old model.

Quite possibly. Because the models are "eating" the harness --they don't need it to tell them to do multiple steps with planning, implement, review, or to spawn sub tasks, etc. anymore.

I also assume that in some way, in order to produce correct code without having to do a lot of but fixes, the larger models have gained implicit to execution engines or at keast predictors in their weights.

There are also several projects that have trained VMs into layers. Or embedded them in various ways.

There are two major potential advantages if you can really make it work well. One of them is potentially huge speed improvements by iterating inside the accelerator without deciding and running external programs. The other is being self-contained, eliminating the need for most of the harness complexity.

Another reason I think this is likely to become a real thing is that it seems with hardware improvements with flash and within a few years fully in-memory compute, we are going to get much larger, faster and smarter models. So they will be able to more easily absorb the complexity of emulating systems with more fidelity, as well as operate more quickly.

You can also imagine that as models get larger and faster, it is a natural next step in training. The top models now automatically run the software the generate and test it thoroughly. They do not need to be told to do this. By feeding enough of those parts of the sessions back into a large enough model to train on, the system gains the ability to predict the outcome without having to run the code. So it's the core part of a training pipeline already built into recent models.

There also might be some convergence with continuous learning. Because altering weights based on some token context seems like it could be similar to keeping a program resident. And maybe the continuous learning part provides the "compilation" step.

How did they integrate Voxtral Large exactly? Does it output a special token to indicate it is asking for help, or is it running a command?

Anyway, if it is trained to decide when to think/ask for help that's good.

Too bad it's not open source though. But for large companies in Europe or elsewhere that want to control their data, this might be one of the best options now.

Although the frontier is true full duplex. Anyone know if there is anything that is open source for that other than MoshiRag or PersonaPlex?

Also did the CSM update some months back fix it? I never got a chance to test it.

Someone please tell the author thank you for working on that. I hope it gets into normal Arch distribution ASAP.

I recently moved to CachyOS and have been hoping to eventually save up for some kind of laptop (or more likely the cheaper N1 when it comes out, my income is very unstable) like this. I just assumed I would be able to put Linux on it.

I see this new chip/setup as being an obvious technological upgrade and more cost efficient, even though the memory bandwidth is slightly disappointing. But I see it it as a superior and more advanced architecture with the memory sharing and much more efficient integrated chip and more efficient ARM.

To me this is just clearly the logical way to build efficient laptops and PCs now. I do hope that we don't have to wait that long to get something with better memory bandwidth though.

I will admit that the description of prefill times is a little disappointing. Someone please tell the author thank you for working on that. I hope it gets into normal Arch distribution ASAP.