Building @Runta, the execution layer that controls what AI agents can actually do. Ask Me Anything https://nitter.cf/t.co/AeLaIL0CWh
Seattle, WA
Joined June 2013
- Tweets7.1K
- Following489
- Followers5.2K
- Likes15.4K
This is what I thought about true RSI.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Scaling agents is all about keeping orchestration simple.
Current setup: 27 Grokbots, 19 Cursor cloud agents, Codex, Claude Code, and Devin. Just 1 Grokbot (Sharky) connects them all via Notion & Linear for high-concurrency jobs.
Edward retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
bbtree is a fast Rust B-tree: 5x faster than std's BTreeMap on one thread, with 30% less memory. The part I like most: speed is easy to measure, but correctness is not. It is guarded by fuzzing, Verus proofs, and deterministic simulation testing. blog.xiangpeng.systems/posts…
imi, Kimi, Kimi, Kimi
I'm begging of you please don't take my job
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
Verification is the hard part of AI security. Agents make it easier to find and patch bugs. It is much harder to verify that the agent did not mock data, inflate the impact, or claim a fix that never landed. Most of the time, the agent is an intelligent fuzzer.
One day I'm going to need a robot bodyguard just to protect me from everyone else's robot bodyguards. 🤖🥊
One possible reason decision models like Jev show such sharp gains is that their architectures leave massive headroom for sampling.
49% → 63% on aircraft recognition in 3 rounds. Zero weight updates.
Connected Google Research's Harness² to Jev:
• It rewrites its prompts
• Builds new tools on the fly
• Keeps lessons across runs
Repo: github.com/SASUKE40/harness2
49% → 63% on aircraft recognition in 3 rounds. Zero weight updates.
Connected Google Research's Harness² to Jev:
• It rewrites its prompts
• Builds new tools on the fly
• Keeps lessons across runs
Repo: github.com/SASUKE40/harness2
I have tried it. Amazing visual model. It thinks about images like a human does, with eyes.
My day with agents. Morning: Grok Bot fetches context from my central Notion database, then my coding loops run. Afternoon: Codex, Cursor and Devin review the work. Before standup: Grok summarizes today's progress from Linear.
Diagram below.
I set up coding loops that run until CI passes. Before lunch, I get a short ASD-STE100 report and a diagram.
Short text and one picture make it easier to understand what the agents did.
TIL from the Gemini 4 Argon briefing: the best way to make a model provider fix what you care about is to build an eval. Don't test on one-offs. Throw hard, long-horizon tasks at it and report the gaps.
People feel that the harness matters less because models are now better at tool use.
For Q&A-style agent workflows, where you hand off a task and let the agent finish it, today’s models are vastly overqualified. That’s why so many people feel the harness matters less and less.
But for agents that proactively drive a bigger mission on their own, for autonomous organizations, a harness that only does this is nowhere near enough.
As models get smarter, we should give agents bigger missions.
Right now, task size is capped by how much a human can think through and spell out in one go. Flip the workflow: hand the agent a long-horizon mission it clearly can’t finish alone, and keep humans in the loop to steer and align along the way.
Introducing FrontierPint
Introducing The Cheeky Pint Benchmark.
We made our intern watch every episode of Cheeky Pint to bring you this breakdown of pints drank by guests to the nearest sixth.
Congratulations to @elonmusk who is currently in the lead with three full pints of Guinness.
