Not sure if I would tradeoff speed for accuracy. Yes, it's incredible boring to wait for the AI Agents in IDEs to finish their job. I get distracted and open YouTube. Once I gave a prompt so big and complex to Cline it spent 2 straight hours writing code. But after these 2 hours I spent 16 more tweaking and fixing all the stuff that wasn't working. I now realize I should have done things incrementally even when I hav…
Wouldn't it be possible to trade speed back for accuracy, e.g. by asking the model to look at a problem from different angles, let it criticize its own output, etc.?
Just have it sample for longer or create a simple workflow that uses a monte carlo tree search approach. Don't see why this wont improve accuracy. I would love to see someone run tests to see how accurate the model is compared to similar parameter models in a per time block benchmark. Like if it can get same accuracy as a similar parameter autoregressive model but with half the speed, you already have a winner, besides other advantages of a diffusion based model.
> it can adjust a whole block of tokens when it encounters some kind of disjunction. This is true in principle for general diffusion models, but I don't think it's true for the noise model they use in Mercury (at least, going by a couple of academic papers authored by the Inception co-founders.) Their model generates noise by masking a token, and once it's masked, it stays masked. So the reverse-diffusion gets to dec…
The exact types of path dependencies in inference on text-diffusion models look like an interesting research project.
Yes, the problem is coming up with a noise model where reverse diffusion is tractable.
Ok. My go to puzzle is this: You have 2 minutes to cool down a cup of coffee to the lowest temp you can You have two options: 1. Add cold milk immediately, then let it sit for 2 mins. 2. Let it sit for 2 mins, then add the cold milk. Which one cools the coffee to the lowest temperature and why? And Mercury gets this right - while as of right now ChatGPT 4o get it wrong. So that’s pretty impressive.
Depends on the shape of the cup! You can contrive a cup shaped like an exponentially flaring horn, where adding the milk increases the volume a little, which massively increases the surface area, and so leads to faster cooling. Or you can have a cup with a converging top, like a brandy glass, where adding the milk reduces the surface area, and makes cooling even slower.
And add the twist: what if the cup is made from a highly conductive material?
Not sure if I would tradeoff speed for accuracy. Yes, it's incredible boring to wait for the AI Agents in IDEs to finish their job. I get distracted and open YouTube. Once I gave a prompt so big and complex to Cline it spent 2 straight hours writing code. But after these 2 hours I spent 16 more tweaking and fixing all the stuff that wasn't working. I now realize I should have done things incrementally even when I hav…
AI field desperately needs smarter models - not faster models.
Definitely needs faster and cheaper models. Fast and cheap models could replace software in tons of situations. Imagine a vending machine or a mobile game or a word processor where basically all logic is implemented as a prompt to an llm. It would serve as the ultimate high level programming language.
AI field desperately needs smarter models - not faster models.
Definitely needs faster and cheaper models. Fast and cheap models could replace software in tons of situations. Imagine a vending machine or a mobile game or a word processor where basically all logic is implemented as a prompt to an llm. It would serve as the ultimate high level programming language.
I think natural language to code is the right abstraction. Easy enough barrier to entry but still debuggable. Debugging why an LLM randomly gives you Mountain Dew instead of Sprite if you have a southern accent sounds like a nightmare.
There are so many models. Every single day half a dozen new models land. And even more papers. It feels like models are becoming fungible apart from the hyperscaler frontier models from OpenAI, Google, Anthropic, et al. I suppose VCs won't be funding many more "labs"-type companies or "we have a model" as the core value prop companies? Unless it has a tight application loop or is truly unique? Disregarding the team c…
Word on the street is a lot of money is going into vertical application AI companies this season. Makes sense - the bitter lesson means capturing a market and proprietary data is a good play, while frontier models keep getting better at using what you (and only you) own.
The linked page only compares to very old and very small models. But the pricing is higher even than the latest Gemini Flash 2.5 model, which performs far better than anything they compare to.
Sort of. The benchmarks showing Flash 2.5 doing really well are benchmarking its thinking mode, which is 4x more expensive than Mercury here
Flash 2.5 without thinking mode is also exceptionally good fwiw.
Definitely needs faster and cheaper models. Fast and cheap models could replace software in tons of situations. Imagine a vending machine or a mobile game or a word processor where basically all logic is implemented as a prompt to an llm. It would serve as the ultimate high level programming language.
I think natural language to code is the right abstraction. Easy enough barrier to entry but still debuggable. Debugging why an LLM randomly gives you Mountain Dew instead of Sprite if you have a southern accent sounds like a nightmare.
I'm not sure it would be that hard to debug. Make sure you can reproduce the llm state (by storing the random seed for the session, or something like that) and then ask it "why did you just now give that customer mountain dew when they ordered sprite?"
Would have been nice if along to this demo video[1] comparing speed of 3 models, they would have share the artifacts as well, so we can compare quality.
Not sure if I would tradeoff speed for accuracy. Yes, it's incredible boring to wait for the AI Agents in IDEs to finish their job. I get distracted and open YouTube. Once I gave a prompt so big and complex to Cline it spent 2 straight hours writing code. But after these 2 hours I spent 16 more tweaking and fixing all the stuff that wasn't working. I now realize I should have done things incrementally even when I hav…
So my personal belief is that diffusion models will enable higher degrees of accuracy. This is because unlike an auto-regressive model it can adjust a whole block of tokens when it encounters some kind of disjunction. Think of the old example where an auto regressive model would output: "There are 2 possibilities.." before it really enumerated them. Often the model has trouble overcoming the bias and will hallucinate…
Unfortunately the intuition and the math proofs so far suggest that autoregressive training is learning the joint distribution of probabilistic streams of tokens much better than diffision models do or will ever do. My intuitive take is that the conditional probability distribtion of decoder-only autoregressive models is at just the right level of complexity for probabilistic models to learn accurately enough. Intuitively (and simplifying things at the risk of breaking rigor), the diffusion (or masked models) have to occasionally issue tokens with less information and thus higher variance than a pure autoregressive model would have to do, so the joint distribution, ie the probability of the whole sentence/answer will be lower and thus diffusion models will never get precise enough. Of course, during generation the sampling techniques influence the above simplified idea dramatically and the typical randomized sampling for next token prediction is suboptimal and could be beaten by a carefully designed block diffusion sampler in principle in some contexts though I havent seen real examples of it yet. But the key ideas of the above scribbles are still valid: autoregresive models will always be better (or at least equal) probabilistic models of sequential data than diffusion models will be. So the diffusion models mostly offer a tradeoff for performance vs quality. Sometimes there is a lot of room for that tradeoff in practice.