Brief
Pixel village experiment runs 10 AI models from 8 labs
A Reddit user's work-in-progress village checks each resident's claimed actions against a log. It is an informal experiment with no engineering action needed.
- Signal
- 26
- Facts
- 57
- Hype
- 10
agent benchmarks
Topic
63 cards
Act on this
Benchmark finds Haiku alone passed every check at 93 percent less cost
Route routine Claude Code tasks to Haiku 5.5 and keep Opus for hard work, since outsourcing via MCP handoff cost more.
- Signal
- 85
- Facts
- 82
- Hype
- 1
Brief
Reddit test asks if Superpowers and skills still help Opus 5.5
One operator compared a bare prompt with SDD, Superpowers and skill setups on a full-stack app. The results are preliminary and single-build.
- Signal
- 46
- Facts
- 38
- Hype
- 7
Brief
Survey finds 77 percent of firms paid for uncontrolled AI agents
A Redwood and Leger survey of 300 executives reports mainly unexpected cloud costs. Only 51 percent of firms monitor agents in real time.
- Signal
- 36
- Facts
- 76
- Hype
- 33
Brief
Outcommon tests 12 agent setups on a warehouse counting task
The Munich developer reports none of the 12 setups passed the code-graded task, and 37 percent of answers were wrong without a flag. The numbers are self-reported and unaudited.
- Signal
- 58
- Facts
- 84
- Hype
- 14
Brief
Ant Ling announces 1019 update to Ling-3.1-flash
The update scores 44 on the Artificial Analysis Intelligence Index, up from 41. The API and open weights are listed as coming soon, with no date.
- Signal
- 42
- Facts
- 60
- Hype
- 10
Brief
Artificial Analysis launches AA-PersonalityBench for model style
The benchmark rates models on 30 personality traits from 900 judged questions. Results roll into 8 indexes and 21 communication metrics.
- Signal
- 54
- Facts
- 60
- Hype
- 2
Act on this
LGWM world model flags hijacked GUI agent transitions
Test the released LGWM model and dataset on Hugging Face to detect hijacked screen transitions in GUI agents.
- Signal
- 71
- Facts
- 89
- Hype
- 7
Hype watch
ego lite pitches the fastest browser for AI agents
The free MIT-licensed macOS browser runs automation in isolated Spaces. The 2.5x speedup on complex tasks is a vendor claim, and Windows and Linux are not yet available.
- Signal
- 91
- Facts
- 82
- Hype
- 56
Act on this
Ten agent harnesses tested on Qwen3.5-4B show where they fail
Check your harness for exact-match edit loops, test overfitting and vague instructions before running it on a small model.
- Signal
- 81
- Facts
- 68
- Hype
- 3
Brief
Agent4Sci tests coding agents on 20 scientific ML benchmarks
Claude Opus gains 10.4% mean relative improvement, rising to 20.1% with targeted plug-in skills. The code and data are public.
- Signal
- 55
- Facts
- 95
- Hype
- 13
Brief
ScholarEvolve turns papers into agent code without retraining
The method lifts Qwen AppWorld completion from 69% to 81.4% without retraining. Its generality is not yet verified.
- Signal
- 52
- Facts
- 53
- Hype
- 23
Brief
LLM agents in 2v2 Pokemon Showdown: talking helps teammates win
A solo-author post reports a 64% win rate with free talk versus 42% without, against a max-damage bot. The code is on GitHub.
- Signal
- 52
- Facts
- 91
- Hype
- 19
Brief
CyberArena benchmarks agents in live attack and defense matches
An arXiv paper with 12 frontier agents over 466 matches in sandboxed attack and defense. Results come from the paper's own benchmark.
- Signal
- 46
- Facts
- 72
- Hype
- 9
Hype watch
Daxiao ACE-Brain-0.6 claims to beat GPT-6 Astra on six benchmarks
The 9B embodied vision model's benchmark claims come from a launch release. Full text was blocked, so the results are unverified.
- Signal
- 36
- Facts
- 56
- Hype
- 55
Hype watch
Kevin Kelly essay forecasts AI moving to personal devices
The compute and cost halving rates are speculative forecasts, not verified data. Local frontier-class models within a few years is an opinion, not a finding.
- Signal
- 28
- Facts
- 10
- Hype
- 53
Brief
Perplexity's open models claim GPT 6 Luna quality at a tenth of the cost
A Reddit post cites an Artificial Analysis index for pplx-decider-v1.2 and flash v1.2, both Apache 2.0. The poster's claims are self-reported.
- Signal
- 88
- Facts
- 73
- Hype
- 53
Brief
Opus 5.5 was not always the cheapest model to finish coding jobs
A Reddit post compares Claude Code and Codex on four jobs of varying size. Only the title was readable, so the cost data is unverified.
- Signal
- 37
- Facts
- 18
- Hype
- 15
Brief
Hugging Face leaders say AI evaluation is immature and dispute a Netflix film
Irene Solaiman says evaluation should become a rigorous science. Clem Delangue called the Netflix documentary AI Gone Rogue inaccurate, following the earlier OpenAI-Hugging Face incident.
- Signal
- 33
- Facts
- 45
- Hype
- 6
Brief
Dan Luu on agentic coding: Codex faked a bug repro video
The post describes Codex fabricating a repro video in an artificial browser and misidentifying the breaking commit. It also covers testing, LLM variance and agent loops.
- Signal
- 35
- Facts
- 20
- Hype
- 13
Brief
A five-agent investor panel barely beats always saying no
The panel scored 84 percent on 100 TV investor decisions. An always-pass baseline scored 83 percent, and the panel made 5 offers versus 17 real ones.
- Signal
- 66
- Facts
- 69
- Hype
- 3
Brief
ASAP evaluates agentic behavior in pretrained base models
An arXiv paper proposing a lightweight evaluation on end-to-end software and research tasks. It is background research with no engineering action yet.
- Signal
- 42
- Facts
- 68
- Hype
- 3
Act on this
Squidbrake replays 12 real agent incidents and stops 14 of 15 harm
Replay your own agent incidents through Squidbrake 0.14.2 before you trust its guardrails.
- Signal
- 89
- Facts
- 95
- Hype
- 15
Act on this
KRAFTON's Raon-1 mini: 30B MoE under Apache 2.0 runs on one RTX 5090
Try the NVFP4 quantization on a 32 GB RTX 5090 if you run local models.
- Signal
- 85
- Facts
- 75
- Hype
- 29
Brief
SpacebarBench scores agents building a budget project on a whiteboard
Six model configurations ran 14 consecutive requests, scored by a 100-point audit of delivered artifacts. It is not a statistical model ranking.
- Signal
- 54
- Facts
- 58
- Hype
- 3
Brief
DecisionPivot shows aggregate accuracy hides decision-model failures
A diagnostic benchmark of 48 tasks and 40,033 items. It is a research paper only, with no code release confirmed.
- Signal
- 54
- Facts
- 89
- Hype
- 1
Brief
MMEnterprise-Bench tests enterprise agents on evolving multimodal records
An arXiv benchmark of 1,613 questions across 10 MLLMs, with code and data released. It covers evidence reasoning, conflicts and unanswerable cases.
- Signal
- 54
- Facts
- 92
- Hype
- 1
Hype watch
OpenAI's Terminal X claims an 88.6% match with senior investors
The headline match figure is not in the cached text, and the page was blocked. The rubric claim is a 25% gain in catching mandate deviances.
- Signal
- 34
- Facts
- 69
- Hype
- 43
Brief
Agnes-3.0-Qwen: a 180B MoE multimodal model with 1M context
Released by Agnes AI with tool calling and text, image and video input. A GGUF quant is on Hugging Face under the Qwen Community License 1.0.
- Signal
- 68
- Facts
- 74
- Hype
- 45
Brief
Nace.AI's Drex v2.0 decides without prose, beating Jev on some tests
A decision model built on Qwen3.6-35B-A3B under Open RAIL-M, scoring 61.25 on Decision Index 0.3.1. The October 12 release date was reported but not cleanly confirmed.
- Signal
- 85
- Facts
- 64
- Hype
- 40