>The same dynamic runs through the open evals, smaller models that feel dumber in practice still outscore better ones on sites like Artificial Analysis. This isn’t hypothetical, Thinking Machines’ Inkling Small scored within a point of its flagship sibling on the Artificial Analysis Intelligence Index with less than a third of the parameters and beat it on Humanity’s Last Exam, GPQA Diamond and SciCode. Public, static, famous test sets leak into training data and fine-tuning choices.

This sounds unconvincing, because a) pelican test is subjective, there's simply nothing to leak as it has no available direct answers and maybe an extremely faint preference signal, and b) the same small models actually do perform well when you change the subject. Some models are genuinely trained to be better at some domain, such as 2D layouts or vector graphics in this case. It all depends on particular recipes and datasets. This is the actual reason these tests are poor as vibe checks: they don't do anything to disentangle generalization, memorization, and specialized training.

The benchmark I want to see people adopt is:

Build a flowsheet based steady state chemical process simulator, then use it to simulate and optimize a full scale oil refinery.

1) Building a solver engine that works at this scale is not a trivial problem, and the successful ones rely more on heuristics than some categorically different solution approach.

2) Defining the engineering equations relevant to this task is relies on understanding what level of fidelity is required to answer the questions people ask of steady state process models.

3) Knowing the thermophysical properties of chemicals and crude oils is possible from the open literature, but the information is diffuse and different correlations are applicable in different situations.

4) Creating a GUI which converts a flowsheet into matrix math is non-trivial, although a sequential modular approach is a bit easier.

5) Defining large scale models in such a way that they solve robustly is as much art as science. For example, completely closed recycle loops like refrigeration systems are a nightmare for solvers, so it is often better to define them in an open-loop way.

6) Optimization involves knowing the relevant commodity prices, but more importantly how to define the constraints on the model so it doesn't just say to produce infinite gasoline.

7) Troubleshooting the inevitable convergence failures is also as much art as science. There are a large number of diagnostic techniques, but fundamentally you need to be able to relate what is happening during the solver iterations with the intent of your model because more often than not the problem is that you've asserted something impossible, redundant, or irrelevant.

I keep seeing Astra make beautiful 3d stuff online, yet when I feed it some old school RuneScape assets (even tried with some very detailed guidelines) and asked it to generate some new plausible assets it failed horribly.

I think there’s still something really off with current (frontier) models when it comes to creating “novel” stuff? Even 2004 style graphics..

Or am promoting it wrong?

Though I somewhat think some benchmarks are silly like the article says. I saw someone on YouTube recently take a picture of a building across the street from them (seemed like it was in NYC), and asked GPT 6 Astra to make it in blender. It did a surprisingly good job in 30 minutes. So though these benchmarks don't seem to mean much you could always add a touch of randomness to them like the person in the YouTube video did, but the problem with that is how would you compare the benchmarks in any clear way if they aren't even consistent? Regardless it appears LLMs are getting this good at the general task and not just at the particular instances of said task.

> That’s the problem, these tests can’t tell you how good a model is anymore because it’s trivial for labs to optimise for exactly these tests by the next release.

But they don't prove the claim. Are the models amazing at recreating Minecraft, but the second you swap the word Minecraft out with another game or a custom game, it shits the bed?

That's not what I see. My feed is full of people using Astra to recreate all sorts of games from Diablo to some random idea they came up with, in ridiculously polished detail like animations that would have taken me weeks of iteration in gpt-5.6-sol but it was a single shot by Astra.

@notch Hey Markus, can I get a refund on my alpha distro of minecraft? I think the currency is worth more than it used to be considering how many versions there are now :)

I cant select text nor click links on this page with firefox.

I’ve used Astra for the past day and a half. My layperson’s review is that it is impressive at computer use and 3D reasoning, and fails in similar ways to 5.6 Sol at similar rates when it comes to coding. I have no idea how it scored so high on SWE benchmarks because so far it has been very “mid” as the kids say.

Most of my and my peers PortCos run their own eval and benchmark sets, simply because they know what they need best.

The reality is, capabilities have largely converged across foundation models over the last 18 months, and much of the value add is coming from the harness layer itself now.

This has been the operating assumption for me and my peers, and has largely played out that way.

That said, this has always been an issue with benchmarking since the very beginning. DB Benchmarks, compute benchmarks, and others that were external facing were always inherently a content and product marketing tool. The actual internal benchmarking used to model, understand, and enhance your product was always a closely held secret.

Most of these conversations are happening, but largely in person and not on HN.