An important aspect of AI is that it needs to be seen as moving forward all the time. Plateaus are the death of the hype cycle, and would tether people's expectations closer to reality.

In college, I took an interim elective course on geology of the national parks. On the first day of class, the professor asked an icebreaker for students to say which national park they lived closest to. I said Ohio - Cuyahoga Valley.

Well some snot nosed boy scout confidently piped up that there were mostly certainly no national parks in Ohio, and the professor agreed. This is a deep personal grudge that I still hold to this day.

https://huggingface.co/datasets/CnakeCharmer/CnakeCharmer

This project started from a belief that llms should be better at doing python to cython code translations than they are. So we started setting a large set of parallel implementations.

Then I realized that Claude code was much better at working on the data using tools (mcp) to check and iterate. The scope transformed into an platform for creating the SFT agentic trace dataset using sandboxed tools for compilation, testing, linting, address sanitizing and benchmarking.

We still need to bulk up the GRPO dataset with a large number of good unmatched python examples. But early results using SFT only on gpt-oss 20b are quite good.

* We need to completely deregulate these US companies so China doesn't win and take us over

* We need to heavily regulate anybody who is not following the rules that make us the de-facto winner

* This is so powerful it will take all the jobs (and therefore if you lead a company that isn't using AI, you will soon be obsolete)

* If you don't use AI, you will not be able to function in a future job

* We need to lineup an excuse to call our friends in government and turn off the open source spigot when the time is right

They have chosen fear as a motivator, and it is clearly working very well. It's easier to use fear now, while it's new and then flip the narrative once people are more familiar with it than to go the other direction. Companies are not just telling a story to hype their product, but why they alone are the ones that should be entrusted to build it.

Mid-sized models like gpt-oss minimax and qwen3.5 122b are around 6%, and gemma4 31b around 7% (but much slower).

I haven’t tried Opus or ChatGPT due to high costs on openrouter for this application.

In my opinion, the reason people don't use DSPy is because DSPy aims to be a machine learning platform. And like the article says -- this feels different or hard to people who are not used to engineering with probabilistic outputs. But these days, many more people are programming with probability machines than ever before.

The absolute biggest time sink and 'here be dragons' of using LLMs is poke and hope prompt "engineering" without proper evaluation metrics.

> You don’t have to use DSPy. But you should build like someone who understands why it exists.

And this is the salient point, and I think it's very well stated. It's not about the framework per se, but about the methodology.

https://hindenburgresearch.com/smci/

There was certainly a time when it seemed they were shopping for engineers opinions of what to do with it, but I think they quickly determined it would be a much smaller market anyway from ssds and didn’t end up pushing on it too hard. I could be wrong though, it’s a big company and my corner was manufacturing and not product development.

That Cyrix machine was already miles ahead of the 386 that was handed down to me to play text based games on and learn dos through hard knocks. I remember leafing through old hard drives that had 10mb of capacity and realizing they had no value despite not being that old.

Later in college, I had the confidence to build my own first desktop with parts cobbled together from sketchy resellers. Athlon A1 single core 1ghz. Man that thing could fly.

Turing test is really in the rearview, huh?

Humans need machines to detect if a machine wrote the text, because humans aren’t sure.

The "new" on huggingface banner has weights that were uploaded 11 months ago, and it's 2B params. Work on this in the repo is 2 years old.

The amount of publicity compared to the anemic delivery for BitNet is impressive.

Comments URL: https://news.ycombinator.com/item?id=47331058

Points: 1

# Comments: 0

If you're looking for a zero-shot classifier, tasksource is in a similar vein.

https://huggingface.co/tasksource/ModernBERT-large-nli

It doesn't seem terribly common yet though. I think it is challenging to keep it stable.

[1] https://www.opencompute.org/blog/amd-arm-intel-meta-microsof...

[2] https://www.opencompute.org/documents/ocp-microscaling-forma...

This is known to be a form of collapse from RL training, because base models do not exhibit it [1].

1. https://arxiv.org/abs/2505.00047