
🤖 @meta superintelligence labs / recursively improving 🔁
Joined April 2013
- Tweets2K
- Following1.4K
- Followers793
- Likes11.7K
Ishita Mediratta retweeted
We will be hosting a NYC based MeetUp on Reliable AI on Nov 16.
The first one in the series will focus on Synthetic data.
Register here: luma.com/gfkzrpx3
2 x Keynotes, lightening talks and snacks.
Ishita Mediratta retweeted
There's a new meetup in NYC on Reliable AI! Organizers include @karen_ullrich, @marksibrahim.
The first is November 16, on synthetic data. I'll be giving one of the keynotes.
There's also lightning talks! Sign up and tell the NYC reliable ML community about your work.
Ishita Mediratta retweeted
Agents can often fix a mistake by retrying, then forget what they just learned! Next session, same mistake 🥲
Introducing Self-Direction: each failure becomes a small, procedural adapter the agent trains on its own reflection and reuses later 💪🏽
🌐 self-direction.github.io/
🧵1/n
Ishita Mediratta retweeted
Open weights, meet your favorite stack.
We're teaming up with @FireworksAI_HQ, @baseten and @modal for the Meta Global AI Developer Hackathon, so you can run Muse Glimmer and Muse Spark Open on the tools you already build with.
Compete for a share of $1M in prizes. Applications close Oct 25.
Limited to 20,000 selected applicants. Terms apply → bit.ly/3WDXxBe
Apply now → bit.ly/4ju9wuL
Ishita Mediratta retweeted
A few months ago we set ourselves the goal of making an open model for the claw: good for agentic, fits in a 32GB gpu. Here's what went into Muse Glimmer: 1/?
teapawt.net/AIatMeta/status/211226…
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
Ishita Mediratta retweeted
Something insane buried under the muse spark open launch today that didn't get enough attention:
it hallucinates less than claude, gpt-6 and nearly every big chinese open model.
hallucination rates below (how often it guesses wrong instead of saying "i don't know")
- muse spark open: 24%
- grok 4.7: 29%
-kimi k3: 53%
- gpt-6.1 sol: 54%
- claude opus 5.5: 59%
- deepseek v4.1 flash: 96.5%
this is extremely useful because in an agentic loop "i don't know" is cheaper and faster than a wrong answer which leads to wasted branches.
(source in comment below)
Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.
Meta and Apple’s CEOs are publicly supporting each company’s work on Muse Spark Open. theverge.com/news/1012293/me…
Ishita Mediratta retweeted
muse spark open -- weights out today! best US open-source model, and runs on the new mac studio 🍎
Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.
Our team in FAIR at Meta is hiring contract Research Engineers.
We work on agents automating AI R&D and the science of recursive self-improvement, develop new test-time learning algorithms and apply them to important problems such as alignment.
📍 On-site (Menlo Park / Bay Area) or remote (ideally US timezone).
Some of our work to give a flavor:
Self-Improving Pretraining : arxiv.org/abs/2601.21343
Self-Challenging Agents: arxiv.org/abs/2506.01716
SWEET-RL (multi-turn agent RL): arxiv.org/abs/2503.15478
Self-Taught Evaluators: arxiv.org/abs/2408.02666
Self-Rewarding Language Models: arxiv.org/abs/2401.10020
Self-Alignment with Instruction Backtranslation: arxiv.org/abs/2308.06259
AI Research Preference Models: arxiv.org/abs/2608.13940
Autodata : arxiv.org/abs/2606.25996
Recycling the Web (pre-training data quality & quantity): arxiv.org/abs/2506.04689
NaturalReasoning: arxiv.org/abs/2502.13124
Coconut (reasoning in continuous latent space): arxiv.org/abs/2412.06769
CoCoMix (LLM pretraining with continuous concepts): arxiv.org/abs/2502.08524
RL from Human Interaction: arxiv.org/abs/2509.25137
SPICE (Self-Play in Corpus Environments): arxiv.org/abs/2510.24684
Automated LLM Speedrunning Benchmark (NanoGPT): arxiv.org/abs/2506.22419
This is not a full-time role. 1-year contract. Interested? Send your resume to xianl [at] meta.
Ishita Mediratta retweeted
🚀 Happy to see AIRS-Bench, an AI R&D benchmark open-sourced by @metaai , featured in this year's @stateofai report alongside PostTrainBench and other influential AI benchmarks
🎯AIRS-Bench measures the ability of AI agents to execute end-to-end AI R&D across the full research lifecycle, from idea generation 💡 and implementation 🛠️ to experiment analysis 🧪 and iterative refinement 📈
We used it in the Muse Spark and Muse Spark 1.1 safety reports to assess the risks of models automating AI R&D work and outpacing governance mechanisms. Head over to the pdfs for harness details and frontier model comparisons:
ai.meta.com/static-resource/…
ai.meta.com/static-resource/…
The benchmark is designed to quantify the capabilities of both LLMs and harnesses in training AI models, as shown in the "Gyms for AI" slide of @nathanbenaich's report
💻 GitHub: github.com/facebookresearch/…
📜 ArXiv: arxiv.org/pdf/2602.06855
🤗 HF dataset: huggingface.co/datasets/face…
Thanks to the dozens of people who contributed to and supported the benchmark development
@alisia_lupidi @_tomwithanh @BhavulGauri @basselralomari @albertomariape Alexis Audran-Reiss, Muna Aghamelu @NickyBaldwin3 @LuciaCKun @GagnonAudet Chee Hau Leow, Sandra Lefdal, @amoudgl, @SabaNazir, Emanuel Tewolde, Isabel Urrego @mahnerak @ishitamed @EdanToledo @MartinJosifoski @rybolos @alex_h_miller @j_foerst
@syhw @yorambac 🙏
welcome to the 9th annual @stateofai report!
ai building better ai, world models, physical ai, geopolitics, cyber defense and 12-month predictions.
read it at stateof(dot)ai or watch my editor’s cut here.
a few of this year’s biggest findings follow:
Ishita Mediratta 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.
Ishita Mediratta retweeted
We are excited to announce we are partnering deeply with Muse to enable agentic checkout with Shop Pay on all Shopify stores, offering people an easy and delightful way to shop and check out with Muse.
Ishita Mediratta retweeted
the most underrated part of muse is the model that powers it. thank you to the MSL research team—some of the most technically brilliant, hardworking scientists and engineers i’ve ever known. you brought original ideas to difficult problems, invented new methods, and battled through enormous technical complexity to build the jewel at the heart of muse.
thank you to the teams working on agentic post-training, tool and computer use, memory, personalization, alignment, safety, and multimodal understanding. thank you to the people building data, evals, and infrastructure beneath it all. so much scientific ingenuity lives in those details: designing a better experiment, finding the flaw in an apparently good result, staying with a stubborn problem until a new way forward becomes clear.
and now that jewel powers a beautiful product your grandma can use without knowing a thing about the extraordinary complexity underneath. it will just work for her. all those discoveries, all that patient work, becoming something simple and useful in another person’s hands. there is something magical about that. i’m grateful to be building with you.
Ishita Mediratta retweeted
Last month I wrote about how we can build a positive and safe future for everyone: meta.com/thefutureisforevery…
Every lab has the responsibility and incentive to move at the pace required to train its models safely, and the ability to take its own actions to ensure that happens.
The reality is:
- People won't want to use agents that are misaligned with them and that don't do what they ask, so labs have a strong natural incentive to make their models more aligned.
There is a lot of debate about slowing progress on capabilities until alignment catches up. My view is that trust and alignment are quickly becoming the most important capabilities that will differentiate agents and models. Any lab that doesn't focus on alignment will fall behind.
- Labs face significant liability if their models cause harm, so they have a strong incentive to prevent this as well.
Meta delayed shipping Muse for several months to focus on safety and security. We didn't call for everyone else to do this before we would. We just did it as part of our day-to-day work because it was clearly the right thing for people and for us. I'm proud of the security foundations we've built.
- Engaging independent evaluators and advisors is industry best practice. MSL already does this today in several areas because it helps produce better work. Other labs can just do this too. In general, it would be helpful for there to be a larger and more diverse ecosystem of evaluators.
- Committing the significant majority of compute towards serving people rather than racing towards recursive self-improvement is one of the best ways to ensure we develop this technology safely. Meta has made this commitment and other labs can do this as well.
I believe the key to building a positive future for everyone is maintaining the right balance of power. This is within our power to do.
Ishita Mediratta retweeted
A cool thing about @Muse is that every Muse runs on its own Linux machine. So you and your Muse can do normal Linux stuff together — like build your own terminal!
Ishita Mediratta retweeted
1/ today we're rolling out Muse, our new personal ai assistant. Muse is always-on, wicked fast, can use a browser, connect to your apps, and is designed to be secure.
try it now: muse.ai.
Ishita Mediratta retweeted
As a test of our progress to advance the frontier of AI research, in June we entered the next generation of our autonomous AI research system, AIRA₃, in a live Kaggle competition run by NVIDIA to fine-tune a 30B Nemotron model. The challenge was to teach the model to reason better — all competitors had access to the same information and were graded externally on a private test set.
AIRA₃ placed 8th out of ~4,000 teams to win Gold, outperforming human competitors who had access to the same frontier tools.
We believe this is a reliable signal that AIRA₃ can improve a targeted capability of an AI model at a level similar to human experts.



























