My implementation of the Recursive Language Model (RLM) paper by @a1zhang , Kraska, and @lateinteraction .
Key insight: "Treat long context as an external environment, not something to stuff into a context window."
Applied to video understanding — instead of encoding 38K frames into a prompt, the agent:
→ Treats video as an environment
→ Writes code to explore segments
→ Uses recursive LLM sub-calls for analysis
Tested: 20+ min video, 7 steps, $0.002
Paper: arxiv.org/abs/2512.24601
Code: github.com/mohammed840/RLM-i…
Wanted to share something I've been working on: Text2WetLab 🧪
Been building this with a couple of Cambridge PhD researchers, turning scientific papers into executable protocols for lab robots.
We're now collaborating with labs to take it further. It's fully open source, and we'd love more people to contribute!
github.com/PhysicalAIBenchma…
SFT is more like getting a haircut with a chainsaw than it is “fine-tuning”! Should just be built into primary training.
And having a “system prompt” is silly! Default behavior should be sensible.
The above are already true for Tesla AI (not yet deserving of “super”). It is the most efficient AI on Earth.
FSD V15 goes to wide release next month. It will be something special.
Moe retweeted
Harnesses used to be a crutch for weak LLMs. In 2026 and onward, with stronger models, harnesses now offer a mechanism for composing an LLM’s existing and new capabilities to generalize well beyond their training environments.
In our new paper, LM Harnesses are Compositional Generalizers, we explore how various harness abstractions enable LM systems to generalize over task length, task strategies, recursive task depth, and more through composing existing LLM capabilities. These abstractions provide a means for more data-efficient training, especially over long tasks, by keeping trajectories locally in-distribution (LID) — i.e., shaping all individual LLM calls to be in-distribution.
For tasks that take a very long time to roll out, this implies that we can even perform surrogate harness training, meaning we train an individual LLM call instead of doing the very long harness rollout, using purely synthetic tasks, with fake tool calls + sub-agents!
A common idea floating around is that LLMs will absorb harnesses and make them obsolete. We disagree, and argue that we haven’t sufficiently explored what harnesses can provide.
The full paper, logs, codebase, and data are available: alphaxiv.org/abs/2611.harnes…
w/ @SOURADIPCHAKR18 @lateinteraction
1/7
Our work with insurers shaped how we tested Inkling on @tinkerapi. We built 15 underwriting cases from public sources to test how models handle new evidence.
Pass@4
Inkling: 60%
Inkling 2 Small: 40%
Inkling Small: 20%
1
4
328
7/7
The benchmark is public. On GitHub you can read each case turn by turn and inspect its verifier. On Hugging Face you can download the data and case-level results.
GitHub: github.com/mohammed840/inkli…
Hugging Face: huggingface.co/datasets/moha…
19
7/7
The benchmark is public. Download the 15 cases, inspect the rubrics or run the evaluation yourself.
GitHub · code, tests and run instructions
github.com/mohammed840/inkli…
Hugging Face · data and results
huggingface.co/datasets/moha…
19







