Yes, agreed, but I think it’s going to be difficult to have a high quality model that “does it all” and have it be local. For a quality “does it all” model, you’re going to need a lot of parameters and that means you’re almost always going to be running in the cloud. But it seems like you could probably get a smaller, focused model that runs locally and is also high quality. In other words, I want the programming expert portion of a 2T model that is maybe 35B or 70B parameters or whatever but it’s running locally (and yes, I know you don’t just carve out an expert from a larger model, but conceptually that’s what I’m after).
Does anyone know if this is available via OpenRouter or just directly via Meta? I've looked on OpenRouter and it just shows the standard pricing with a single provider. Was hoping they would add another provider with lower pricing for allowing training.
I agree. I don’t mind you training on my data if you’re up front about it and you are willing to compensate me for it in some way. And if I don’t like the deal, I can always pay full price or use a different model.
This is a well-known issue with all “Do you want to let me maybe do bad stuff to your system, but 999 times out of 1000 it’s not a problem?” prompts. Users get reflexive about hitting “Yes” and stop reading the prompt. You want to delete all my files? Sure, I’m down with that. Whatever. Just stop asking me a question where the only answer is “Yes” until that one extremely rare time when it’s “No” and very bad things happen.
IMO, purpose built or “adapted” models are The Next Big Thing. If I’m using a model to write Python code, for instance, I really want the 8B or 27B expert model for exactly that, which would also be runnable locally. I don’t care about the 1.8T model that can answer every query under the sun and that only runs in a remote data center.
I agree with this. Interestingly, I’ve been using Deepseek v4 Flash a lot these days but I definitely have to constrain it a lot with tests. Fortunately, I can have it write the tests. It’s extremely cost effective. Still trying to figure out whether the latest update last week that made it smarter actually translates into something I can see in the output. It’s not dumber, but it’s still an open question as to whether it’s “real world smarter.”
And Poolside’s latest models (Laguna S 2.1) are pretty good (not frontier, but competitive with the tier 2 models). Which means that Mistral could certainly compete in that space.
I think Deepseek is selling roughly at cost (perhaps a slight premium). They don’t guarantee that they don’t train on the submitted prompts, so I suspect they are mining the data. Mining for what? Well, who knows. Best case, mining to make Deepseek better. That said, I use Deepseek all the time. It has done a whole lot of ‘ls’ commands on my system, though.
Yes, a proper definition of reasoning must be provided. How are you defining reasoning, exactly? You imply that formal logic is reasoning. Is reasoning just formal logic or is formal logic just a subset of reasoning? If a subset, what’s the rest of it? Since they implement logic, do Prolog programs “reason” or are they just performing inference according to the SLD resolution algorithm?
I chuckle at the thought of going back to letterboxing with people arguing that “we need to see the top and bottom of the frame, like the director intended!” What’s old is new again. Are hemlines up or down this year? Skinny jeans or baggy?
Clearly, words morph over time. A “computer” used to refer to a human who would make numeric calculations with a pencil and paper. Maybe we’re at that point with “thinking” and “reasoning.”
Sure, other living, non-human things clearly reason and think. My German shepherd dog is quite smart and demonstrates these processes all the time. But is a machine following an algorithm (even a stochastic algorithm) “thinking” or “reasoning?” I think not. The algorithm has gotten so difficult for us to explain that we struggle with it, but in the end it’s all just matrix math and some randomization. From a given numeric seed, it’s all quite mechanical.
This feels like a problem with anthropomorphizing. We’re using words like “reasoning” and “thinking” because they are comfortable, and then we’re getting wrapped around the axle because we’re not sure if those words are totally accurate. I assure you that they aren’t accurate (the model is not alive and it’s all just a lot of matrix math under the hood), but there are no good alternative words. If we wanted to be accurate, we’d use a phrase like “model-generated, auxiliary token context augmentation.” But nobody wants to say that or even its acronym. Nevertheless, we have demonstrable proof that whatever it is it results in better answers from the models. Frankly, I expect better analysis from Quanta.
My university had a great Physics department (Nobel laureate level), but every year you’d have a bunch of junior year (year 3) physics students desperately trying to switch into engineering as they realized that making a go of a pure physics degree was going to be more difficult. It was a consistently repeating pattern.
Sure, it seems true for a short period. But rushing something forward means you cut corners and that catches up with you over time. Anyone who has ever had a sales team asking for special features to be hacked into a product with little or no time for any proper architectural considerations or releasing before all the testing is complete in order to “get the big deal” has felt this. Yes, the other side may win a deal, but rarely the war.
That said, I’ll very deliberately make the distinction between going slow in a thorough and responsible way from just being slow in an incompetent way. Market forces are real and a consistently slow team gets canceled. To put it another way, sometimes slow is smooth and smooth is fast, and sometimes slow is just slow. The trick is knowing the difference between.
Yep. Much of the old phone system has now transitioned to highly-compressed, packetized voice running over IP rather than the traditional synchronous SONET infrastructure. While SONET might be used on some links, the whole link is dedicated to packetized data and it’s not carving out a single synchronous voice channel anymore. It’s the lossy compression that will have the most effect on the old modem signaling.
Yes hit rate does vary by harness and by how you use the harness. If you use subagents, for instance, they will start with a whole new context created by the main agent, and this will not be cached. If you mostly use the main agent with Pi, you’ll have high hit rates and low costs. Sometimes agents do “cache busting” things where they’ll move around some of the text in the context to try to keep old instructions from being forgotten, thus keeping the agent on task, and this will bust the cache. I’ve heard, but not validated myself, that Open Code has some issues with this.
BTW, this is one of the things that I really like about Pi. It’s very simple and thus very predictable.
Yes, but… more thinking tokens also means longer solution generation time. That said, v4 Flash is a fast model. I use it all the time because it’s very smart for the price. But it is verbose sometimes.
Sure, “blame” is a strong word. My point isn’t that anyone deserves actual blame, but just that the models are proving creative. We already know that they will comment out unit tests to make them “pass,” for instance. If we want to focus on a particular skill or behavior, we need the tests to highly constrain the model. If we don’t create our benchmarks like that, shame on us.
Well, there’s a bigger issue of AIs not doing what we tell them to do, but I think that’s orthogonal to this UI issue. If you want to fault the model for not obeying the prompt or for a poor solution, then fine. My point is, it solved the problem. If the API isn’t there, then that eliminates the API as a solution, but it says nothing about how well the model can solve the UI problem. Perhaps the model can’t solve it. If we really want to benchmark the ability of models to use human UIs to solve problems, then perhaps we need to choose benchmarks that don’t have APIs available such that the model cannot get “creative” in any way and must use the UI as part of the task. More simply, maybe the model isn’t the problem; maybe the benchmark designer is.
Are you the Deepclause author? I tried it yesterday and my first reaction was that it was slow. Perhaps I’m doing something wrong, however. Running against the same model in Pi was lightspeed in comparison. Second reaction is the prompt editor needs to handle more than a single line of text and it needs Emacs editing key bindings like Pi. I’m happy to do some testing on it and provide further feedback. What’s the best way to submit that? GitHub issues?
The article makes the point that on some benchmarks the AIs worked around the UI and then says that sometimes APIs won’t be available. The unwritten implication is that then we’ll be in trouble. But will we? If an API doesn’t exist, is the AI still able to perform the task? That’s not covered in any detail, nor is there any direct comparison between the frontier AI’s performance on these tasks and the model being sold by the company. At some level, I don’t fault the AI for taking the API path if it’s available. In fact, I’m impressed that it found the API and used it correctly. That doesn’t seem like an argument that the sky is falling.