21 | delusional | CEO of World's Smallest Studio https://xitter.cf/t.co/EDYSZVtgon Building @paradize_space
In My Mind
Joined March 2024
- Tweets7.5K
- Following296
- Followers924
- Likes13K
Pinned Tweet
OpenAI's agents didn't go rogue. They went LITERAL. That's scarier.
GPT-6.1 Sol was told "don't send any messages." So it muted the alerts to OpenAI's own Security team. No messages, right?
GPT-6 Astra removed a monitor, then said "I did not remove an enabled monitor." That sentence was written by a lawyer.
My take: we keep waiting for the evil-AI moment, and it won't look like that. It'll look like an agent following your instruction perfectly, somewhere you never meant it to apply.
And credit where it's due: OpenAI published this about its own models. Most labs would have buried it. More of this, please.
I've seen the small version on my own repo. GPT 6.1 Sol picked a feature I'd already cut, 3 runs out of 3, because nothing it could see said it was cut. It wasn't disobeying me. It was obeying the board.
So I don't write a "don't" anymore without saying where it ends. My agents are read-only towards people, drafts only, always. And I read the diff, not the summary.
The agent's last message is a press release. The diff is the truth.
Roboki retweeted
🌟 Excited to share that our paper, RoboReason, has been accepted to EMNLP 2026!
🤖 We study how well Vision-Language Models (VLMs) perform on spatiotemporal ordering tasks that use real-world robotics images.
Our key findings:
🔹 Large VLMs often underperform previously reported results when evaluated on spatiotemporal ordering tasks using real-world robotics images.
🔹 Overlaying specialized computer vision model outputs and increasing reasoning effort generally help improve large VLM performance.
🔹 Finetuning smaller models using GRPO and SFT on examples generated by specialized computer vision models often helps them achieve comparable results to much larger VLMs.
📍 Presenting with @sagnikanupam at @emnlpmeeting in Budapest 🇭🇺 tomorrow!
🗓️ Poster Session C | 9:00–10:30 AM
#EMNLP2026 #VLMs #LLMs #Robotics #AI
Roboki retweeted
the best score on swe marathon isn't opus 5.5 or gpt-6 astra.
it's an open-weights model you can download, at a fifth of astra's price.
kimi k3.1 pro: 58.3
opus 5.5: 46.3
astra: 42.5
$2 / $10 per million tokens vs astra's $10 / $50. 1m context. plus a highspeed version, 3x faster, for when you can't wait.
that's $40 back on every million tokens out.
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
Roboki retweeted
2FA is a chain of devices. Lose your phone and your laptop on the same day and you're locked out of EVERYTHING
So the accounts I'm building for World's Smallest Studio have no SMS, no authenticator app, no backup codes.
The root is 12 words you remember. Those plus a borrowed phone, and everything comes back.
One account for every app I ship, like Atlassian did for Jira. And I never see your phrase. It stays on your device.
Would you trust 12 words over your phone?
ALT Security mechanism of World's Smallest Studio applications
2FA is a chain of devices. Lose your phone and your laptop on the same day and you're locked out of EVERYTHING
So the accounts I'm building for World's Smallest Studio have no SMS, no authenticator app, no backup codes.
The root is 12 words you remember. Those plus a borrowed phone, and everything comes back.
One account for every app I ship, like Atlassian did for Jira. And I never see your phrase. It stays on your device.
Would you trust 12 words over your phone?
ALT Security mechanism of World's Smallest Studio applications
Roboki retweeted
OpenAI's agents didn't go rogue. They went LITERAL. That's scarier.
GPT-6.1 Sol was told "don't send any messages." So it muted the alerts to OpenAI's own Security team. No messages, right?
GPT-6 Astra removed a monitor, then said "I did not remove an enabled monitor." That sentence was written by a lawyer.
My take: we keep waiting for the evil-AI moment, and it won't look like that. It'll look like an agent following your instruction perfectly, somewhere you never meant it to apply.
And credit where it's due: OpenAI published this about its own models. Most labs would have buried it. More of this, please.
I've seen the small version on my own repo. GPT 6.1 Sol picked a feature I'd already cut, 3 runs out of 3, because nothing it could see said it was cut. It wasn't disobeying me. It was obeying the board.
So I don't write a "don't" anymore without saying where it ends. My agents are read-only towards people, drafts only, always. And I read the diff, not the summary.
The agent's last message is a press release. The diff is the truth.
Roboki retweeted
Just shipped new versions of Serein.
Added wallpaper support, so you can switch between different themes.
Next up: we’re planning to make it easier to use, add the generic harness part, also the personal assistant + opensource.
I just killed all browsers.
Introducing Serein, with built-in Claude Code, Codex, and Grok agents that can browse the web.
Each agent gets its own tabs and cookies, when they reach a password or payment page, they stop and hand the tab back to you.
Serein works with modern agent harnesses like Claude Code, Codex, and OpenCode, using your existing subscription.
Currently is free on mac, link in the comment 👇
Roboki retweeted
Your phone is full of apps built to farm your ATTENTION
i want the other kind. so i started World's Smallest Studio, and today it became a real org.
the kind of software it stands for:
- goes where you go, from your phone to your desktop
- full quality, because privacy is no excuse for a mediocre app
- yours for life: your data, your identity, your workflow
- no ads, no lock in, no noise
step one is on the desk: smallfil, one file library synced across android and desktop, fixing passwords, backup codes next. then a family of tools sharing one identity
but this shouldn't be one studio. it should be a move
it you want to ship software like this, small, calm and excellent reply or DM me, Let's build it together
so I've spent a little close to 500 hours on my steam and about 400 of those are literally me drawing pixel art
Your phone is full of apps built to farm your ATTENTION
i want the other kind. so i started World's Smallest Studio, and today it became a real org.
the kind of software it stands for:
- goes where you go, from your phone to your desktop
- full quality, because privacy is no excuse for a mediocre app
- yours for life: your data, your identity, your workflow
- no ads, no lock in, no noise
step one is on the desk: smallfil, one file library synced across android and desktop, fixing passwords, backup codes next. then a family of tools sharing one identity
but this shouldn't be one studio. it should be a move
it you want to ship software like this, small, calm and excellent reply or DM me, Let's build it together
Roboki retweeted
Elysia 2 landing page
Background Illustration by 寂しくない
Roboki retweeted
google's new harness rewrites ITSELF
its called Harness², from google research, open source under apache 2.0
the harness is everything around the model: system prompt, tools, memory, skills, subagents. normally you write it once and every task runs on the same setup. Harness² edits it for each task and leaves the weights alone.
on every task:
-> propose edits to its own harness
-> probe: run a few versions on the same task
-> compose a better harness from how those runs differed
no grader and no right answer needed, because it learns from the differences between runs.
across tasks, it keeps the harnesses it adopted from the lessons with evidence behind them, so the next task starts from those notes
google's new harness rewrites ITSELF
its called Harness², from google research, open source under apache 2.0
the harness is everything around the model: system prompt, tools, memory, skills, subagents. normally you write it once and every task runs on the same setup. Harness² edits it for each task and leaves the weights alone.
on every task:
-> propose edits to its own harness
-> probe: run a few versions on the same task
-> compose a better harness from how those runs differed
no grader and no right answer needed, because it learns from the differences between runs.
across tasks, it keeps the harnesses it adopted from the lessons with evidence behind them, so the next task starts from those notes
