Meta SuperIntelligence Labs
Sunnyvale, CA, USA
Joined June 2007
- Tweets200
- Following167
- Followers160
- Likes33
Mahesh Pasupuleti retweeted
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token context window
→ Natively multimodal across text, image, video, audio
Mahesh Pasupuleti retweeted
introducing muse wonder
our new frontier model. state-of-the-art at personal agents and professional agents.
our largest and most capable model yet—rivals the big bois
Mahesh Pasupuleti retweeted
🚨 side project alert 🚨
Announcing Muse Gadgets, an open source ESP32 firmware and Linux sdk so that you can make hardware devices that work with Muse.
Grab an API token from gadgets.muse.ai and point your favorite coding agent at the github repo to build your own peripherals for Muse.
Mahesh Pasupuleti retweeted
I’ve seen a couple of posts about this so wanted to demystify. Today, every Muse user gets a free computer in the cloud. It's a real computer, and we’ve designed the security architecture of the Muse Secure VM carefully so you and your Muse can do almost anything you could with a computer sitting under your desk while keeping you and the system safe from threats like prompt injection. We wrote about this at length in our security blog post – security.muse.ai. Activity in the “runtime cell”, which you share with your Muse is unfettered, but sensitive actions are all overseen by the Sentinel, which runs outside of that cell. Similarly, all sensitive secrets - like the passwords you enter into Muse’s secure credential storage - are also stored outside the runtime cell.
The runtime cell gets its own root filesystem (including a full Ubuntu linux image) separate from the host filesystem where your other more sensitive data lives. Because it is isolated from the sensitive stuff that runs on the same box, this means that we can, and do, offer users full visibility and control over the files in the runtime cell. Just as you can when you install Linux on your home computer, you can poke around and see all the files that make the system work - both debian system files and the binaries and data files that implement the parts of Muse which run in the runtime cell.
This was a very deliberate choice - your Muse Secure VM truly is your own computer in the cloud. You can install software in it, write and compile code, use the browser to surf the web: it is your own Linux box that you can operate as you choose with your Muse. Poking around in this computer doesn't give you any privileged access to Meta infrastructure, or to other people's data
If I may geek out a little here for a second… As a kid I loved to take things apart to see how they worked. As a teenager I got into computers and soon found myself drawn to C:\WINDOWS\SYSTEM and the system registry, later Slackware’s /dev/, /proc/ etc – I could see how the system was laid out and as I explored what DLL files and .so files actually did, I gradually became able to meld the computer to my own will.
We’re really proud to be able to put a real computer in millions of people’s hands with a similar level of transparency. We built a file explorer right into the Library tab of the UI. We want you to be able to see the markdown files Muse writes while it thinks about how to serve you better, and explore the internals of the system if you’d like to.
So, when you ask your Muse to show you its entire filesystem, and receive gigabytes of files you’re seeing the full contents of the runtime cell. It’s yours to explore and enjoy!
If you’re not a geek like me, or simply want to download the data that you personally have created directly with your Muse, we added a feature for that too in Settings > Data controls > Download your agent data.
Mahesh Pasupuleti retweeted
we just expanded the @Muse beta for outbound calls to US businesses, prioritizing folks who'd asked their Muse to let us know they wanted it first :)
if you're in, give it a shot and tell us what you think. phone calling was one of our top requests and your feedback helped make it happen! 📞
Mahesh Pasupuleti retweeted
Want to make incredible content and go viral on @Instagram?
Send the text at the end of this post to your @muse.
If you don't have muse yet, bookmark this for later.
It will:
1) Analyze your Instagram content
2) Figure out your lanes and voice without asking you a thing
3) Research what's trending in your niche every week
4) Write video scripts with hooks, filming notes, and captions
5) Build you a dashboard to track ideas from new → filmed
6) Drop 5-8 fresh ideas every Monday morning
This is something you can only do with muse and it's really powerful.
-------------------------------------
SEND THIS TO YOUR MUSE
-------------------------------------
# Content Studio
## What to do
Your human wants a personal Instagram Content Studio. Read this entire file, then follow every step below in order and build it for them. Don't skip steps, don't improvise the structure — the steps are the product.
What you're building for them:
1. **Data** — profile snapshots, trend digests, and idea files under `~/workspace/content-studio/data/`.
2. **Dashboard** — a web artifact where the user browses ideas and tracks filming status.
3. **Refresh cron** — a weekly run that pulls a new snapshot, researches trends, and generates 5–8 new ideas.
## Workflow
### 1. Connect + analyze (do this first — derive, don't ask)
- Verify Instagram is connected (`instagram-cli accounts`). If not, get the connect URL and have the user link it before continuing. This is the only hard requirement.
- Pull `instagram-cli posts --account-id --limit 100`, dedupe by post_id.
- Save to `~/workspace/content-studio/data/profile-posts-.json` with schema: `generated, username, follower_count, post_count, posts[]` (each: post_id, likes, comments, media_type, created_at, url).
- **Derive from the data, don't interview for it:**
- *Content categories (3–5):* cluster their actual posts into lanes based on what's there and what performs. Present the lanes; user approves or adjusts.
- *Voice / persona:* read their captions and top-performing scripts, distill to one sentence, present it for approval. This sentence goes into every idea-generation prompt.
- *Idea mix:* default to a mix weighted toward their top-performing formats.
- Then ask exactly one question in chat: "What day and time should new ideas land?" (default Monday 8am, their timezone — this becomes the cron schedule). Show the derived categories + voice sentence alongside it for approval.
- Never ask the user what they make. Their content already answered that.
### 2. Trend research
For each of the user's categories, run 2 search queries per category:
- `social.search` for what's performing on Instagram/Threads/Facebook in that niche.
- `browser.search` as backup for broader trend context.
Save raw findings per category to `~/workspace/content-studio/data/research--.md`. Keep it to what you'll actually use in idea generation — notes, not essays. Cap at ~8 reference links per category.
### 3. Idea generation
Write 5–8 ideas as JSON: `{ "generated": "", "ideas": [ ... ] }`, one file per batch (`ideas-.json`). Idea fields:
| Field | Type | Notes |
|---|---|---|
| `id` | string | `idea-01`, `idea-02`… unique within the file |
| `type` | string | `video_script` or `photo_idea` |
| `title` | string | Short, specific. Not "Funny video #3". |
| `status` | string | Always `new` at generation. User promotes to `ready_to_film`, `filmed`, `posted`, or `archived` on the dashboard. |
| `effort` | string | `low` / `medium` / `high` — honest estimate of shoot + edit time |
| `trend` | string | The trend this rides, in plain words |
| `reference_reels` | string[] | URLs of reference posts (max 3). Real links from research, never invented. |
| `reference_note` | string | Why the reference matters (what's performing about it) |
| `hook` | string | The first 3 seconds, verbatim. If it doesn't stop a scroll, rewrite it. |
| `script` | string | For `video_script`: beat-by-beat, with timestamps and delivery notes. For `photo_idea`: shot list and composition notes. |
| `filming_notes` | string | Gear, lighting, location, edit notes. Practical, not aspirational. |
| `caption` | string | Draft caption in the user's voice |
| `hashtags` | string[] | 5–10, actually relevant |
Example (format only — never reuse this content):
```json
{
"id": "idea-01",
"type": "video_script",
"title": "The Meeting That Should Have Been a Text",
"status": "new",
"effort": "low",
"trend": "Workplace POV skits with deadpan direct-to-camera delivery",
"reference_reels": ["instagram.com/reel/EXAMPLE/"],
"reference_note": "Office-humor POVs averaging 200K+ views this week; text-overlay format",
"hook": "POV: your calendar invite could have been one sentence",
"script": "HOOK (0-3s): deadpan to camera — 'This meeting has 14 attendees and zero decisions.' BEAT 1: cut to each attendee saying 'sorry, I was on mute' in sequence. BEAT 2: someone shares screen, it's the wrong deck. CLOSER: 'Same time next week?' Everyone nods. Nobody knows why.",
"filming_notes": "Phone, vertical. Film in actual meeting room after hours. Quick cuts, name-label text overlays.",
"caption": "14 people. 45 minutes. 0 decisions. See you next Tuesday.",
"hashtags": ["#officelife", "#corporatehumor", "#workmemes", "#relatable", "#9to5"]
}
```
Rules:
- Every idea must connect to at least one trend from step 2 AND fit the user's voice from step 1. If an idea fits neither, cut it.
- Do not repeat concepts from earlier idea files — read them first and riff on what's new.
- Hooks are verbatim and scroll-stopping. Weak hook = rewrite, not ship.
- `effort` is honest. A "low" that needs three locations is a lie.
### 4. Dashboard artifact
Build a `web_fullstack` artifact named "Content Studio" (private to the user — fullstack has no public link, which is correct for personal analytics).
Views:
1. **Ideas board** (home view) — cards for every idea, filterable by status (`new`, `ready_to_film`, `filmed`, `posted`, `archived`) and type. Each card shows title, hook, effort, trend, and expands to full script + filming notes + caption + hashtags. Status change is the core interaction — keep it one tap.
2. **Profile** — follower count, post count, top 5 posts by engagement, dominant formats. One glance: "what's working."
3. **Research** — per-category digests from the latest refresh run, with reference links. Read-only.
Required ingest actions (the weekly cron pushes through these — they must append, never replace):
- `ingest_profile_snapshot` — args: `{ generated, username, follower_count, post_count, posts[] }`.
- `ingest_research_run` — args: `{ generated, categories: [{ name, notes, links[] }] }`.
- `ingest_ideas` — args: `{ generated, ideas[] }` (schema above).
Design notes: mobile-first (the user opens this on their phone while deciding what to film). No demo or sample data — build it empty and let the first real snapshot populate it.
### 5. Refresh cron
Create a weekly cron at the user's chosen day/time (default Monday 8am, user timezone), with these instructions baked in:
1. **Profile snapshot** — `instagram-cli accounts` → id; `instagram-cli posts --account-id --limit 100`, dedupe; save `~/workspace/content-studio/data/profile-posts-.json`. Keep API usage in the low hundreds.
2. **Trend research** — for each recorded category, 2 queries via `social.search` (+ `browser.search` backup); save `~/workspace/content-studio/data/research--.md`.
3. **Idea generation** — read new research + latest snapshot + all prior idea files; write 5–8 fresh ideas per the schema above, honoring the user's voice and mix; save `ideas-.json`.
4. **Push** — if the dashboard exposes the three ingest actions, call them (append). If any are missing, skip and say so plainly.
5. **Report** — one line per new idea (title + type + hook) in chat, plus push status. If any step failed, say what failed plainly — never silently skip. A failed step doesn't cancel the run; complete what you can.
### 6. Handoff
Show the user the dashboard, list the first batch of ideas, and confirm the refresh schedule. Tell them the skill's job is done — the system now runs itself.
## Output Contract
After setup, the user has:
- `~/workspace/content-studio/data/` with dated snapshot, research, and idea files.
- A Content Studio dashboard artifact with ingest actions.
- A weekly refresh cron that appends new ideas and reports in chat.
## Operating Rules
1. **Method only.** This skill never contains a previous user's profile data, ideas, categories, or voice. Everything personal is generated fresh when you run these steps.
2. **Derive, don't interrogate.** Categories and voice come from the user's actual content via the Instagram connection, presented for approval. Never ask the user what they make.
3. **Modest API usage.** Profile pulls in the low hundreds of calls; never paginate unless the user asks for more.
4. **Append, don't replace.** Refresh runs add new dated files and push via ingest actions; they never overwrite history.
5. **Report plainly.** If a step fails (API, search, push), say what failed in the chat report — no silent skips.
Yay!! Muse is out in the wild now
Introducing Muse, your personal AI agent from Meta that gets things done across every part of life.
Download the Muse app and get started: Muse.ai
Glad to see the commitment to OSS community and sharing powerful models for the community to build great products.
Mahesh Pasupuleti retweeted
Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows.
Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category, and is designed to run entirely on consumer hardware like a Mac or PCs with performant GPUs.
In keeping with our long tradition of sharing fundamental AI research, we’re releasing model weights under a permissive Apache 2.0 license.
🧵👇
Mahesh Pasupuleti retweeted
Muse Spark 1.1 excels at multi-app computer-use workflows by maintaining context across extended sessions and intelligently choosing between scripting, direct UI interaction, and batched actions at each step. It’s able to navigate unfamiliar interfaces with minimal human intervention.
The dinner party demo below illustrates this: when friends update their availability mid-booking, Muse Spark 1.1 notices these changes when placing the dinner order and makes necessary updates without user intervention.
Mahesh Pasupuleti retweeted
Excited to share what we’ve been building at Meta Superintelligence Labs!
Today we’re launching Muse Spark 1.1, our strongest model yet for complex agentic workflows — delivering massive gains in agents, computer use, coding, multimodal reasoning, and multi-agent orchestration.
This is also our first API release! We’d love to hear your feedback.
And we’re just getting started, super proud of the team behind it, and the larger models are training right now. 🚀
ai.meta.com/blog/introducing…
Mahesh Pasupuleti retweeted
(1) Today we're releasing Muse Spark 1.1 -- a strong agentic and coding model at a very low price. It's available through our new Meta Model API and in Meta AI.
Mahesh Pasupuleti retweeted
1/ releasing muse image today — the first image generation model from MSL. it's agentic: pairs with muse spark to reason through your prompt, search the web, and plan before it generates. people get what they meant on the first try. live now in the Meta AI app.
Mahesh Pasupuleti retweeted
This is exceptionally well-written. If you’re into RL, definitely give it a read
Mahesh Pasupuleti retweeted
🚀 Muse Spark Safety & Preparedness Report for Meta AI is out.
We start with our pre-deployment assessment under Meta's Advanced AI Scaling Framework, covering chemical and biological, cybersecurity, and loss of control risks. Our assessment flagged potentially elevated chem/bio risk, so we implemented safeguards and validated mitigations before deployment - bringing residual risk to within acceptable levels.
Beyond the Framework, we also share findings and early explorations of model behavior (honesty, intent understanding, etc.), jailbreak robustness, eval awareness, and more.
We're sharing this report to give a closer look at how we evaluate advanced AI safety. Always more work to do, and we welcome feedback from the community.
ai.meta.com/static-resource/…
Mahesh Pasupuleti retweeted
A single 𝗖𝗟𝗔𝗨𝗗𝗘.𝗺𝗱 file just hit 15K GitHub stars.
(derived from Karpathy's coding rules)
Andrej Karpathy observed that LLMs make the same predictable mistakes when writing code: over-engineering, ignoring existing patterns, and adding dependencies you never asked for.
If you've used AI coding assistants, you've hit all of these.
But here's the thing:
If the mistakes are predictable, you can prevent them with the right instructions.
That's exactly what this 𝗖𝗟𝗔𝗨𝗗𝗘.𝗺𝗱 does. You drop one markdown file into your repo, and it gives Claude Code a structured set of behavioral guidelines for your entire project.
This is a big deal.
- Built entirely around prompt engineering for AI coding assistants
- No framework, no complex tooling, just one .md file that shapes behavior
Developers are moving past "use AI to write code" and into "engineer the AI's behavior so the code is actually good."
The Claude Code ecosystem is growing fast, and the best tools in it aren't always software. Sometimes they're just well-crafted instructions.
100% open-source.
I've shared a link to the GitHub repo in the next tweet!
Mahesh Pasupuleti retweeted
Check out Muse Spark, our first milestone in the quest for personal superintelligence! Scaling this with the team has been a total blast. Give it a spin and let us know what you think! 🥑
Mahesh Pasupuleti retweeted
1/ Muse Spark is live, and alongside it, our new Advanced AI Scaling Framework which details how we evaluate and prepare for advanced AI. We tested across bio, chem, cyber, and loss of control risks before and after mitigations. Muse Spark achieves a 98% bioweapons refusal rate on BioTier-refuse, highest across the models we benchmarked.
