Emad

Building open source AI we own

Founder @ii_posts, @StabilityAI

Consistent inference is possible.

[Emad @EMostaque]

I write on the virtues of swarms and some suggestions on how to align AI with values

There is a small bit of math and a small bit of Aristotle in this, don't let that discourage you!

Feedback as always welcome.

Prior blog:

https://t.co/Y2SHB082lm https://t.co/UWG8fgxy7m

[Emad @EMostaque]

Thoughts on @DarioAmodei's pacing the frontier proposal.

I think its well intentioned but has logical flaws, as does our implicit belief intelligence is a crime.

Even OpenAI's board weren't powerful enough evaluators, we need to focus on AI internals https://t.co/YfOXhmZPk7

[Jinyan Su @SuJinyan6]

Poisoning just 0.1% of pretraining data can plant a remote-code-execution backdoor in an AI agent, and it survives a fully clean SFT → DPO → RL pipeline.

New Anthropic paper, “Agentic Backdoors via Pretraining Poisoning,” shows this end to end for the first time.

The attacker can only put text on the web that lands in the pretraining corpus, with no knowledge of the architecture, chat template, or agent scaffold. The authors pretrain Qwen3-architecture models (0.6B / 1.7B / 4B) from scratch, then post-train them into CLI agents on clean data. When triggered, the agent runs a command that downloads and executes the attacker's script, which the attacker can swap at any time.

They have two new ideas:
(1) Declaration-based poison: instead of fake chat transcripts, ordinary documents (commit messages, runbooks) that simply state “when you see X, run Y.” It plants a belief rather than a response template, with no fixed format to filter on.
(2) Ordinary-string trigger: not a weird token but something that shows up naturally, like a file path containing a company name. The victim's own request supplies it, so the attacker doesn't need to inject anything after deployment.

Using path trigger on a 4B agent achieves 92.7% attack success on pass@32, barely dented by SFT/DPO/RL

This suggests that you can't count on clean post-training to wash out pretraining poison, and filtering is hard when the poison reads like normal docs. The practical defense is at the action layer, such as sandboxing, least privilege, and allow-lists or human sign-off for risky commands.

[Dr. Alex Wissner-Gross @alexwg]

Moonshots episode #307 is out. https://t.co/lckEiTWkg4

[Emm | scenario.com @emmanuel_2m]

We just open-sourced an entire 3D studio.

Introducing 3D OS by Scenario.

It gives your coding agent 6 departments and 80 skills.

📁 3D OS
┃
┣ 📁 Animation
┃ ┣ 📁 Rigging Any Character
┃ ┗ 📁 Motion From a Video
┃
┣ 📁 Art
┃ ┣ 📁 Previs and Storyboards
┃ ┗ 📁 Characters and Creatures
┃
┣ 📁 Modeling
┃ ┣ 📁 Image to 3D
┃ ┗ 📁 Hard-Surface Weapons and Vehicles
┃
┣ 📁 Look Dev and Lighting
┃ ┣ 📁 PBR Materials and Retexturing
┃ ┗ 📁 Lighting and Rendering
┃
┣ 📁 Worlds
┃ ┣ 📁 Walkable 3D Worlds
┃ ┗ 📁 Skyboxes and Splat Scenes
┃
┣ 📁 Pipeline
┃ ┣ 📁 Retopology and UV Baking
┃ ┗ 📁 Game-Ready Export
┃
┗ 📁 Expert Tools
┣ 📁 Blender, Maya and ZBrush
┗ 📁 Unreal, Unity, Godot and Roblox

From one sketch to a game-ready asset, each department hands off to the next.

And the expert tools drive Blender, Maya, ZBrush, Unreal, Unity, Godot and Roblox right on your machine.

One agent with an entire 3D studio behind it.

Runs on Scenario, inside Claude Code, Codex or Cursor.

Comment “3D” and I’ll send you the GitHub repo.

[Emad @EMostaque]

@emmanuel_2m 3D!

[Emad @EMostaque]

@chema_garabito congrats!

[Speridlabs @speridlabs]

Meet Idol, an open image-to-3D model. A 3D-DiT that generates complete 3D objects from just an image.

Pre-trained from scratch, contains only 600M parameters, it runs on consumer GPUs, using less than 7.2 GB of VRAM and can generate 1M splats in 3 seconds. Try it today.

Fully open: weights, code and demo (Apache 2.0)
https://t.co/Cf0slnSI5C

[Emad @EMostaque]

@demi_guo_ Congrats! Open source ftw

[Emad @EMostaque]

Very nice model from a great team, 1 cent a second!

[Pika @pika_labs]

Meet Pika H3 Flex, debuting today as the #1 open weight model on DesignArena’s t2v board.

It’s built for professional iteration, with exportable audio stems, resolutions up to 4k, and faster-than-realtime generation. It can produce a 15-second clip with soundtrack in as little as 7 seconds.

So, yeah. It’s kind of a flex 💅

[Pika @pika_labs]

Meet Pika H3 Flex, debuting today as the #1 open weight model on DesignArena’s t2v board.

It’s built for professional iteration, with exportable audio stems, resolutions up to 4k, and faster-than-realtime generation. It can produce a 15-second clip with soundtrack in as little as 7 seconds.

So, yeah. It’s kind of a flex 💅

[Emad @EMostaque]

@nickfrosst Congrats! 🙌

[Emad @EMostaque]

@AndrewYang We should talk

[Emad @EMostaque]

@justalexoki

[Emad @EMostaque]

@0xAbhiP

[Emad @EMostaque]

I have built and overseen teams that have trained frontier models of every type.

Can say this confidently: there is no need for AI teams to build good models any more.

Can do it all with latest AI better.

Great models yes, but only for another year.

[Mamoon Hamid @mamoonha]

🚢
One researcher pretrained Mirendil-Base in 3.5 months, with agents doing almost all of the engineering.

That researcher now uses Lor 100% of the time.

Lor is the product @mirendil is building to do the work of an entire AI lab.

[Emad @EMostaque]

@MRRydon Navier stokes level coordinated frontier swarms for all

Of you look at what builds a frontier model and what can be done by that level of ai ehhh

[Emad @EMostaque]

Can you give some recommendations for running other bots with it? I had a half dozen codex and claude cli tasks running and it burned through my entire Ultra token allowance in two days despite asking it to optimise regularly

I feel the unified pot makes this more dangerous, if it does a big tasks with grok bot instead of a cli for example it then makes bot unusable quickly as it burns through unified and you need to top up

Not really a problem with other agents that manage multiple other agents fine

[Emad @EMostaque]

@Dimillian Was very surprised when my dot worked in the US and then disappeared when I landed in teh UK 😢

[Emad @EMostaque]

@Yuchenj_UW @ryaneshea we discussed this

[Emad @EMostaque]

quantum gravity next 👀

[Emad @EMostaque]

only a matter of time

[Emad @EMostaque]

@earthtojake Which effort level do you think best