Coding agents/self-improvement @xai. Ex-@GoogleAI Resident/@augmentcode. Alum @HKUniversity. 💎 Terran @StarCraft II. Views are personal.
San Jose, CA, United States
Joined February 2016
- Tweets1.1K
- Following268
- Followers3.3K
- Likes1.7K
When pirates use machine guns, sailors can't rely using spears.
Coming to Android folds when 😡 @AnthropicAI?
Congrats @YuchenHe07 @ericzelikman!
No one changes the world alone
Introducing Ampersand - the first model built for us. Ampersand is a collaborator trained to help people achieve their individual and collective goals. In multi-month simulations, users collaborating with Ampersand accomplished more than with Fable and other frontier models
Available to everyone through our API
So true.
Always scale your rigor to your stake. No user, 100 users, 1M users, 1B users require very different levels of rigor.
Too much rigor is harmful not only to your productivity, but TO YOUR USERS TOO.
Pluto is such a milestone. I've wanted to make an AI RTS, possibly in VR, for a bit. The opportunity seems to be getting mature.
In 2016 I set out to solve Brood War with self-play RL from scratch. Today I believe that this quest has been completed youtube.com/watch?v=Lfo4C0xz… (by a 300M params model).
In the agentic coding era, maybe we should have a renaissance of personal offices?
I'm so much more productive in a solo meeting room than at my desk.
Appropriate timeline. Get AGI. Then play StarCraft: Dominion.
Here's my highly decorated Certificate of Enlistment BTW.
Recently I've seen a lot of social media videos of people catching almost the entire process of this flood.
It makes me wonder if we had some agentic analysis to aggregate these signals, we maybe can issue timely warnings to reduce the losses.
Like a TikTok Alert AI feature.
Fuller video compilation: xitter.cf/servasyy_ai/status/209…
这两天满屏都在追孙哥的新闻, 但我更关心的是另一件事
：吉隆口岸泥石流事件
给Codex一条指令：完成中国吉隆口岸事故还原视频
然后Codex + Computer Use + Chrome插件
找资料、搜现场视频、补堰塞湖和口岸被毁的画面、核对 5200 米到 1200 米那段海拔落差、做三维地形动画、生成旁白、合成导出。
网页、终端、本地工具, 它自己在里面来回切,没让我接手。
最让我意外的是它自己收尾检测: 旁白把"2026 年"念错了,它重做语音、重排字幕、重新导出,然后从最终成片里再识别一遍确认改对了。
以前 agent 的问题从来不是不会做,是做完不知道自己做错了。这次它把这一环补上了。
给它真实的工作环境和权限,一句目标能一路做到可交付。 交付物就是这支 8 分钟的片子👇
Fair analysis.
Speaking purely hypothetically here as someone who has no information:
It is a frontier cutting edge model with a lot of test time compute and agentic swarm. The chance that it was able to access some data no one thought it should access that would help it solve the problem is at least nonzero.
We’ve seen those in the past few weeks.
Codex w/ Astra is really good at using your browser to accomplish tasks.
But if you prefer an open-source harness, like @deepseek_ai Harness, @opencode, @pidotdev, here is an open-source MCP that will extend the same browser use capability to any agent:
github.com/cmsflash/agent-br…
If Dario took Sam's bar, would OpenAI need to stop research since the promise was to do so to avoid an AGI race if another entity can demonstrate they will achieve AGI in one year?
It's kind of crazy that Claude Code supports reading the answer out loud on laptop but not on mobile.
In a year, this scale will be regarded as short horizon, just like a few years ago, when people used long horizon to describe 32K-token contexts.
When rollouts become extremely expensive, rollout pruning techniques like this will become more and more important.
Excited to see your awesome work again @hyhieu226!
One Nexus is a neocloud in Vietnam, focusing on delivering @AMD technology to South East Asean users. We worked hard to quantize GLM 5.3 and its Flash sibling to MXFP4. Please try them and give us your feedback.
huggingface.co/OneNexus/GLM-…
huggingface.co/OneNexus/GLM-…
