@Proto Labs without experiments we don’t move it forward though so this is really really good work not just saying that I’m working in similar thinking that you’re doing execution and that’s where the real work gets done

@Joshua Hupp @Yoban Obregan I’ve been working around similar ideas https://wordstofilmby.com/rendertrust/platform

@Felix Mendoza well met Felix. I'm in cost segregation amongst other things. would love to connect as those are value add to each other

@Forest Howlett I bet you are a FunGuy! Welcome!

PS we are going to develop the pilot project we know it will be a little fuzzy in budgeting at first while we get a kind of content fly-wheel going and then dial in a formula where we know pretty close estimate for different types of AI work measured by 'Finished Minute Cost' as the KPI. Once that is in the bag, then a whole lot of other things can happen.

@Joshua Hupp in another group we were discussing this paper, and I related it to the harness and got this idea -> part of what i'd been working on is looking at scripts as code using https://fountain.io and this RLM paper just hit and makes me wonder on world building, which is a big part of what I work on. Chris Madia 2d (edited) • 🆕 News Recursive Language Models: Externalizing Long-Context Reasoning Source: “Recursive Language Models” (arXiv:2512.24601) https://arxiv.org/abs/2512.24601 This paper proposes “Recursive Language Models” (RLMs): instead of stuffing huge documents into a model’s context window, you keep the corpus outside the model (in a Python REPL-style environment) and let the model operate on it with code. The model iteratively reads, searches, slices, and extracts from the external data, and can optionally make a bounded recursive call to another model to solve subproblems on smaller spans. The key idea is that long-context failures aren’t just about token limits; they’re about task complexity and dense multi-step reasoning over many interdependent parts of the input. How this differs from a typical RAG approach: RAG’s core move is to retrieve a small set of “relevant” chunks and paste them into the prompt, betting that a compact working set is enough. RLM keeps the full corpus external and uses computation to navigate it step-by-step, which is better suited to problems where you must combine evidence across many distant sections rather than just fetch a few passages. They evaluate RLMs against direct prompting, a compaction/summarization approach, and a retrieval-agent baseline. RLMs can handle inputs far beyond native context limits and perform strongly on harder dense-reasoning tasks where retrieval or summarization alone tends to collapse. The tradeoff is operational: performance and cost are variable with heavy tails, so real use requires strict guardrails (step limits, tool-call budgets, recursion caps) and good observability.

@Crumbgrabber Trading Hello. Been working on the harness part as much as anything which is the real power of the tech they created. discreet skill set calling on the fly. I think Jan's implementation will mimic. In the meantime: https://www.skool.com/agent-zero/claude-code-harness-for-multi-agent-team-workflows?p=a02b5304

@Crumbgrabber Trading I've done several updates lately FYI much improved: https://github.com/bybren-llc/wtfb-safe-agentic-workflow/blob/main/docs/whitepapers/ITERATION-PATTERNS-COMPARATIVE-ANALYSIS.md

BREAKING: NVIDIA bringing local in hard https://x.com/askperplexity/status/2000589984818954719?s=46 BREAKING: NVIDIA just dropped an open 30B model that beats GPT-OSS and Qwen3-30B — and runs 2.2–3.3× faster Nemotron 3 Nano: • Up to 1M-token context • MoE: 31.6B total params, 3.6B active • Best-in-class performance for SWE-Bench • Open weights + training recipe + redistributable datasets You can run the model locally with 24GB RAM.

@Chris Theberge Framework seems to really be putting a best foot forward. Thanks for the post