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: https://t.co/sMkqVscWZD
Code: https://t.co/J3GxdlKeav
Google DeepMind co-founder Demis Hassabis:
“AGI is coming, yet schools are still teaching children skills that AI will be better at. The entire education system needs to be flipped.”
Training with fewer bits is faster, but quantization error can interfere with learning.
Today on Connectionism, we detail how we trained Inkling 2 Small with a native 4-bit forward pass from the first step of pretraining through the final RL rollout.
https://t.co/CIMECSEpD0
slightly more specific example: increasing diversity of thought by training multiple models, and increasing inference spend by running them in huge swarms over a long time with real-world feedback that gets put back into the weights is possible, and would almost certainly increase creativity a lot. But it's very expensive.
Gemini 4 Carbon is exceptionally good at blogs, handover docs, and assignment.md files. Feels like it actually models intent, structure, and context rather than just generating prose.
One successful run proves very little.
That's why I love pass@4 in evals. It reveals whether an agent can find a path to success, even when its first attempt fails.
Everyone commenting on the nature of "log scale" here, all while the smallest footnote I've ever* seen helpfully says “the placement of the data points on the graph [...] is solely for illustrative purposes”
It's crazy how people look at public benchmarks for non-coding agents and think we're already seeing the frontier of what's possible.
Insurance, finance, legal, and other knowledge work are a completely different story.
The gap between public benchmarks and the internal evals we've been building with enterprises and AI labs is genuinely massive.
We've been finding some really interesting failure modes, capability gaps, and breakthrough points that public benchmarks barely capture.
Excited to open-source some of this work over the coming weeks and months.
I think people are seriously underestimating what's coming for non-coding agents.
Jensen Huang is one of the most successful entrepreneurs and executives in history.
He’s a role model to many of us and a champion of causes many consider good and threatened.
It will be a new era for sure, but it’s a good moment to celebrate achievements of the man who pioneered the GPU and transformed it into dominant AI accelerator