ggcr

My personal bet is that this model should really shine in Autoresearch NanoGPT-style speedruns because its first-class integration with Tinker

It's a fork of OrientDB, isn't it?

I’m surprised they don’t just eject the injured worker from the colony

Wonder if this has something to do due with space constraints. If the study was done in a controlled nest, it must be space bounded one way or another. Dynamics might change when in real-world?

Fascinating. Hidden on the bottom of the article seems to be a video [1] showcasing how they track each ant out of the six colonies of 110 each.

I'd like to read the paper to skim over the methodology but it's not open-access :(

[1] https://www.uni-wuerzburg.de/fileadmin/uniwue/2026/0702Ameis...

With Reinforcement Learning, inference is very present in post-training stages now too

Fair point. Should've titled it "minimal dependencies" perhaps

Since C++ has no HTTP client in its std lib, I really had no other choice but to use curl. Same with OpenSSL. It'd be quite naïve of me to re-implement the whole HTTP stack and SHA256 from scratch =)

Yep! I'd also like to explore C++26 reflection for the XML FSM parser I currently have

Everything so far is synchronous. In our toy benchmarks I've found out Go to be quite performant, perhaps it's using go-routines behind the scenes (?) Haven't explored their internals yet but I was surprised

:(

I was hoping Microsoft would make it open weights, as they have done for years with the Phi models.

The era of big tech releasing models into the wild might be over, which IMO is counter-productive, as we are shifting from "the model is the product" to "the harness is the product"

Meta recently launched Muse Spark [1] and they themselves compare against Claude Opus 4.6 Max.

Here Microsoft is comparing against Claude Haiku, the smallest and least capable model from Anthropic.

[1] https://ai.meta.com/blog/introducing-muse-spark-msl/

Thanks, I've totally missed that

It's basically the same as with the Qwen2.5 and 3 series but this time with 1M context and 200k native, yay :)

From the HuggingFace model card [1] they state:

"In particular, Qwen3.5-Plus is the hosted version corresponding to Qwen3.5-397B-A17B with more production features, e.g., 1M context length by default, official built-in tools, and adaptive tool use."

Anyone knows more about this? The OSS version seems to have has 262144 context len, I guess for the 1M they'll ask u to use yarn?

[1] https://huggingface.co/Qwen/Qwen3.5-397B-A17B

alvaro's posts are must reads

https://ggcr.es/