AI-delirious, Linux-lover, builder and maker, (former for now) Minecraft content creator | https://xitter.gq/t.co/oWIujc2u3l | https://xitter.gq/t.co/HIOQKi9RC8
Denver, CO
Joined January 2007
- Tweets5.8K
- Following261
- Followers5K
- Likes1.3K
I don't know why this is the thing that freaks me out about AI, but it is.
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…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Ironic: @poteto is vibing a game in her beef with @thsottiaux like it's 1999.
Me over here watching my grok @bot quota evaporate as I approach my reset in two days.
No t me over here watching my grok work on Hermes while they work it out. The backup could become the daily.
Guys I found why the cost of serving Grok Bot was unreasonably high last week.
Day 28/ tibo's game has landed
i spent $5.99 on this domain, but i had a blast making it with grok bot and cursor! i hope you enjoy it and find all the easter eggs 🤍
if i did it right you should be able to play it inside of X too
tibo.dog
Anthropic and Grok engineers love to talk about how we're just using 10% of what their models and harnesses can do and how they run dozens of agents shipping thousands of PRs.
I'd use 100% of my quota in 2 days trying.
SAAS providers that don't make their services headless just don't get it.
Stop blocking agents. Open up your APIs or provide MCP access.
I don't want to use your precious web interface. Get over yourself. If that's your only value, you don't have any.
Between their presence here on X and stuff like this @SpaceXSI is absolutely crushing community.
Maybe I'm in a Grok Bot bubble right now, but from here, no other lab touches them.
Ambassador applications for @spacexsi / @cursor_ai are open
Details + apply here: anysphere.typeform.com/space…
Grok @bot v0.77.0 is out, and if you run more than one bot, the new Permissions page is the part to look at.
Every Auto-review rule sits in one place, split into what your bots always ask about and what they can just do.
Approval cards also stay open 30 minutes now, so a card doesn't expire while you're away from the app. I feel like I want even longer with how much I get distracted, but I'll take the 30 minutes.
x.ai/changelog/bot
New update: If you run more than one Grok Bot, you know the approval cards pile up fast. @bot v0.75.8 goes after that.
Auto-review now takes a second look before it interrupts you, so you see fewer cards. A bot saving or editing its own routines no longer gets flagged as risky.
And when you keep approving the same kind of action, your bot can propose a change to your Auto-review rules on a card. You confirm it once, and you stop tapping yes all day.
x.ai/changelog/bot
I can't decide if this is an unneeded layer or something useful. Anyone tried this?
Also feels a bit like buzz.xyz in that you can self-host the a relay. I assume their paid service is the hosted version of that.
Can these bots talk to each other in groups in this app? If Dots and Grok Bot agents could talk to each other, that would be interesting.
@grok , meet Beeper 👋
Manage your Grok Bot conversations from your Beeper inbox, one chat for each bot.
Jangro retweeted
今天把 Railway 上跑的 Hermes Agent 直接删了。花了一下午，全部工作流和项目都切到了 Grok Bot，也重新给这些 Bot 理了一下岗位和汇报协作关系。bot 不是越多越好、非必要也不要拉群，要平衡 好 Token 效率。
顺手给当前的 9 个 Personal Agent 排了个「从夯到拉」，纯属个人看法:
🔴 夯：
Grok Bot @bot ，对 X 的用户非常友好，每天有1000次免费 X search 额度-根本用不完，这意味着基本是拥有实时 X 数据，而 X 上英文区有非常多优质信息源，30刀/月的 Premium 订阅才能用上
🟠 顶级：
Muse @Muse ，Meta 是少有的擅长ToC 的美国公司，上手门槛非常低，免费额度很多
Duet @claudeai ，Opus 5.5 驱动，不 NB 也不行，20刀/月的订阅就能上手
🟡 人上人：
Dot @OpenAI ，藏在 ChatGPT 里，容易忘事定时任务也不按时跑，100刀/月的订阅才能用上，开放爱感觉是有点嗨了
Instinct @instinct ，不收费随便蹬，就是有点慢，适合砍账单和买东西
Cue @CueAgents ，我最喜欢的是能打电话收发短信
⚪ NPC：
Hermes Agent @NousResearch ，开源适合开发者折腾，需要自己部署和配置各种 Key，上手有点难度
Gemini Spark @GeminiApp ，被 Gemini 并进了 Task，但其实是一个 类 Personal Agent 的产品，普通用户绝对够用。
⚫ 拉完了：
OpenClaw @openclaw ，一言难尽，我已经退坑很久了。
这个排名，估计会有很多人不同意。
为了跑 Hermes 才订阅的 Railway，现在上面还剩 8 GB 内存、8 vCPU、100 GB 硬盘的额度，得想想怎么利用一下。
TIL Hyprland (I'm on Omarchy) has a per-workspace setting that toggles the layout: dwindle vs scrolling.
Try it!
Super+L toggles it.
Super + ← / Super + → scroll through them
Resizing panes doesn't resize the others.
This is useful when I end up with several related tiny panes in one workspace and I don't want to promote them to their own desktops.
#omarchy
