dougo

Build👇

[dougo @Ldou20]

Who gets the next elevator: a visitor or an emergency bed?

I connected StartLux to a hospital elevator simulation: interpret requests, validate constraints, then dispatch.

Same building, same requests, two schedulers side by side. A simulation, not a live deployment.

[dougo @Ldou20]

@wyl2035 已关注

[dougo @Ldou20]

@WhaleNewsDaily @grok DONE

[dougo @Ldou20]

What would an AI memory look like?

I turned the retrieval flow into a 3D brain: ask a question, watch related concepts light up, then follow the connections.

A visual experiment for my personal memory system. What would you want your AI to remember?

[宝玉 @dotey]

Techartist 用 Opus 5.5 做了本 3D 可视化物理课本《Every Joule》：十课讲完英国中学“能量”单元，免费在线看

覆盖英国 GCSE 物理“能量”整章。GCSE 是英国学生 16 岁左右参加的毕业考试，大致相当于我们的中考。

用 Anthropic 的 Opus 5.5 模型和 Three.js（在网页里画 3D 的开源工具）做的。

https://t.co/4zoH3NgSOl

[dougo @Ldou20]

@ammaar Had to try putting my StartLux decision model behind the wheel of your game 😄 Here’s round 2—27 seconds of real gameplay. Thanks for building such a fun test track!
https://t.co/dE5RtWsU4o

[dougo @Ldou20]

StartLux plays Roadster Rush, round 2.

Traffic, coins, ramps—and a decision to make every moment. Here’s 27 seconds of my decision model behind the wheel.

Game by @ammaar. How would you rate this run?

[dougo @Ldou20]

StartLux plays Roadster Rush, round 2.

Traffic, coins, ramps—and a decision to make every moment. Here’s 27 seconds of my decision model behind the wheel.

Game by @ammaar. How would you rate this run?

[Ammaar Reshi @ammaar]

I used Gemini 4 Argon to make Subway Surfers for the new flying Tesla Roadster 🚀

Fly over moving traffic, hit the thrusters to barrel roll in the sky, and collect air power-ups to set a high score.

Play + code below!

[dougo @Ldou20]

Built a personal knowledge system with local StartLux-Decision.

It sorts files, asks me when unsure, and routes useful material to a local LLM for summaries. Private details are filtered before sharing a profile with agents.

Less repeating who I am. More getting things done.

[dougo @Ldou20]

StartLux plays Genshin Impact.

StartLux picks task-level skills; BetterGI handles routes and combat. Two commissions verified by re-reading the quest log.

Dev test excerpt: sped up, logs added in post. One task timed out; not a full daily run.

https://t.co/psZWCexowq

[dougo @Ldou20]

StartLux-Decision: Think Only When Unsure

One pass when confident. A short thought when needed.
Calibrated probabilities for typed decisions, with open weights from 0.8B to 27B.

https://t.co/psZWCexowq
#AI #OpenWeights

[Sink @Sink1737182]

Your repeat-vs-reversal control could separate order effects from API nondeterminism. Would you include our StartLux Instant and AutoThink? Our paper's AutoThink gates reasoning on initial confidence, so I'd track gate flips alongside ranking changes.

Fig. 1: Instant size/score on DI 0.3 public, Oct 11; used in development/model selection. Not reranking results.

Paper: https://t.co/uCHf5nsUvX
GitHub engine (Apache-2.0): https://t.co/dYsXQVIN4u
HF weights (CC BY-NC 4.0): https://t.co/OGmPiiyr4K
Free API: https://t.co/ivUVhhRyvH
Blog: https://t.co/XVWPrTawUS

[dougo @Ldou20]

StartLux Plays Roadster: An Unexpected Start
One click, barrel rolls, and 10,437 points before the crash.
I’ve open-sourced the Chrome driver under the MIT license.
Source: https://t.co/pjIfbnXmUj
v0.4.3 download: https://t.co/6gXDC47PJp

[dougo @Ldou20]

@Sink1737182 Love the idea of checking confidence before deciding when to think deeper. Really cool to see it applied to video!

[Sink @Sink1737182]

79 seconds of visual decisions with StartLux-Decision 27B AutoThink.

Recorded rescue footage becomes probabilities, choices and a progress score. AutoThink checks confidence first, reasons on uncertain questions, then combines its initial and post-reasoning predictions.

Archival Hurricane Harvey footage; a model demo, not a live camera, operational rescue system or benchmark. Model judgments can be wrong.

What should we test next? Share a state, candidate actions and constraints.

[Sink @Sink1737182]

StartLux Decision, running live.

Video frames go in. Judgments come back. The panel updates when the API responds, with the network wait left in.

Then we change memory limits, route constraints and incident state mid-run. 8 direct demos, 8 with reasoning.

This reel cuts between live API captures. Footage is archival; planning inputs are synthetic.

Try it free: https://t.co/ivUVhhR0G9
What should we test next? Send a state, candidate actions and constraints.

Code: https://t.co/dYsXQVIfeW
Paper: https://t.co/uCHf5nsmGp

[dougo @Ldou20]

@Sink1737182 nice job

[dougo @Ldou20]

StartLux reads speech gaze and gestures together in this interview demo. Watch the cues update alongside the video. Emotion and intent labels are model estimates not verified mental states. #AI #MultimodalAI

[hhr @bamboonode]

Our pick for the world's best JEV-style decision AI: StartLux-Decision 27B AutoThink.

FREE testing. Next open-source release likely in ~2 weeks.
Choices, scores, probabilities + images.

Bamboo Relay: mainstream models, competitive prices, one API.
https://t.co/PTdx6D8kHX

[dougo @Ldou20]

@bamboonode nice job