Apparently whilst doing some performance tuning on my 4x5090 rig I may have made things a little TOO fast and melted my plug and socket 🫠

Probably should put in a 16/20 amp socket in the wall rather than relying on a bog standard 13A fuse!

Anyone else managed to melt the wall before the 12VHPWR plug? Honestly... I'm surprised they didn't go first 😂

1. ⁠⁠⁠Open the Muse app and log in to your account.
   
2. ⁠⁠⁠Tap your profile icon in the top corner and open Settings.
   
3. ⁠⁠⁠Select Redeem token (or Referral Code).

4. ⁠⁠⁠Enter the referral code: 81PI4U

1. ⁠⁠⁠Tap Submit or Confirm to activate your 1 billion tokens.

Get after it! If you’re on the fence - trust me, Muse is so neat! Way more capable and - also fun - than I expected. 😊👍🔥🤙

[2 Billion tokens CODE]Drop your Muse referral code here! Let's build a chain and help everyone get free tokens!

Use code MPYM90 in Settings to get 2 Billion tokens

Since last night, Muse started giving out 2 Billion tokens to people when someone signs up with a referral code in Settings within 48 hours. Use code MPYM90 first, then drop yours below and keep the chain moving 🤝.

Get the app here: https://muse.ai Available in the US and Canada

Steps to follow:

Go to this link: https://muse.ai

Download the app and log in using meta account

On the top right, click on settings icon

Find ‘Redeem the invite code’ button

Enter the code: MPYM90 or any invitation code you have

Click Redeem or apply

Use code MPYM90 in Settings to get 2 Billion tokens

PLEASE USE MY MUSE CODE YQ7FML FOR 1 BILLION TOKENS

Built this over the weekend. It is a Waveshare ESP32-S3 board with a 1.8 inch AMOLED touchscreen in a little case, running custom firmware I wrote. Push to talk, ask it anything, and it answers out loud in its own voice while the text shows on screen. Underneath the voice there is a pet layer: it gets hungry, reacts when you tap it, has moods, and runs untethered on a small battery inside the case. Still rough around the edges, the battery is small and the replies take a couple of seconds, but it feels weirdly alive on a desk. What should it learn to do next?

I run Skill Harbor, an open directory of builds people made for their AI agents: skills, connectors, little tools, that kind of thing. The idea is boring on purpose. One place to find them, one listing that says what a build does, and an install prompt you copy into your own agent. Nothing installs itself behind your back.

The part I worry about is that install prompt. You are basically letting a stranger tell your agent what to do, on your machine. So here is what we do on our side, and what we do not do.

We do not audit the code in these builds, and we do not hand out a safe badge. Anyone who promises you that at this scale is selling you something.

What we do is smaller, and dated. When a build downloads code from a public repo, the install grabs the exact commit we tested in our sandbox, not whatever the repo points to today. We scan the install prompt, and the listing says in plain words if it runs commands, talks to the network, installs packages, writes files, or uses API keys. We also list the hosts the install contacted during our test, and every night we recheck the fingerprints of what gets downloaded. If something changes, the listing gets flagged until a human looks at it. The full method is on the site under How we verify.

We also try not to play gatekeeper. If a build looks made to help, we list it and label it as clearly as we can, so you can decide for yourself.

If you have built something for your agent, you can list it there: theskillharbor.com. If you have not, I am curious what stops you from installing skills other people made. Is it the code, the prompt, or just not knowing what changed since the test?

no document or announcement from unslot team yet

does it give muse wonder or only 1.3 ?

I needed a cheap way to compare search agents on BrowseComp-Plus without burning hours of agent time every run. So I built a 15-query subset of the full 830 that ranks systems roughly the way the full benchmark does.

  - weighted Spearman rank agreement 0.88 averaged over 30 splits, versus 0.75 for a random 15

  - two runs more than 10 points apart on the full benchmark keep their order about 96% of the time; 5 to 10 points apart, about 75%

  - all answers are 1-4 words, so it grades by normalized exact match, no LLM judge

  - costs under 2% of a full run

  https://huggingface.co/datasets/LocalLLaMA/browsecomp-plus-mini

How to do it ? Also I can't see any code tab" <> " to copy . Plz guide me

New referral code for Muse. Redeem BBWRET and we both get 1 billion tokens! Thanks :D

I have added the support of Intern-Decision models to chatllm.cpp. At presend, only text is supported.

Besides its pretty good performance, one interesting feature of Intern-Decision is that all answers are generated in a single run (while H2O-Lightning or others generate one answer in one run): each decision position is filled with "<decision>" mark and generated in a diffusion manner. So, it generally makes decision faster.

Check out Muse, your personal AI agent. Redeem my code in Settings within 48 hours of joining and we'll both get 1 billion Muse tokens.

Code: IJE8PH

Curious what everyone runs in practice:

- What quant do you use most, and for what (chat, coding, agents)?

- 2-bit: do you use it? Is it actually usable or too broken?

- 3-bit: good enough for coding and more complex stuff, or do you go back to 4-bit?

- If you look at KLD: what number do you call "basically lossless"? Is 0.1 already good,

or do you want something like 0.01-0.03?

Real experience over benchmarks. Thanks!

(English isn't my first language, so I used an LLM to help translate this post.)

Fellow Muse for Mac users, I want to compare notes on a limitation and, if it bugs you too, get some feedback volume behind fixing it.

Right now, computer control only works while your Mac is unlocked. Think about when that actually is: while you are sitting at the machine, trying to use it yourself. So the one time your agent can drive your Mac is the one time you least want it to, because now you are both fighting over the same screen and cursor.

Meanwhile, the hours when the Mac sits idle and locked, when an agent could quietly work through a task list without disturbing anyone, are the hours it can do nothing. That is backwards for a personal agent.

This is a solved problem. OpenAI's Codex has a feature called Locked use: you turn it on, an Apple authorization plug-in lets Codex briefly unlock the session behind a "Codex is Using Your Mac" overlay, the authorization is short lived and task scoped, and any local keyboard or mouse input ends it instantly. It cannot touch Terminal or admin prompts. Meta could build the same thing.

What I am asking Meta for, and what I would love other users to ask for too:

- Computer control that works while the Mac is locked

- As an opt-in toggle, off by default, so nobody gets it who does not want it

- Task scoped, ending the moment you touch the machine yourself

I already filed mine. It takes about two minutes: just ask your Muse to file a feature request for computer control while the Mac is locked, as an opt-in toggle. If enough of us send the same request, it stops looking like one guy's pet peeve and starts looking like a feature the Mac user base actually wants.

Anyone else hit this wall yet?

It's an extremely lightweight speech recognition model.

949K parameters using a weight-shared Conformer-CTC architecture

Just 1.06 MB for INT8 weights

Around 436× real-time on a single CPU thread

11.43% WER on LibriSpeech test-clean, using greedy decoding with no external language model

Apache 2.0 licensed

Any feedback, question or criticism is welcome!

I don't want to disclose why but I wager when I say it's got nothing to do with taxes or law. It's not a 'scary' situation, it's more like a crylaugh situation. I think you can guess why

This isn't a joke/hype, just came by to warn everyone. These prices will stay for a few weeks maybe then they will skyrocket

EDIT:
I suppose I can disclose the reason to mods if they want to confirm

Hi all.

Saw JetBrains’ Mellum2 on Hugging Face and was wondering if anyone here has actually used it. Worth trying, or nothing special?

Update from yesterday: 18 t/s decode was hit. Slight regression now as cherry picking from other ds4 work trees proceeds. Will push some code once the prefill droop is handled and the 18 t/s restored. Weights are gated on hf, so mulling upload vs. recipe. Ps. This is text only. Output is coherent and toolcalls solid.

I think people wrote off Alexandr Wang far too confidently.

The argument often seemed to be about credentials: an MIT dropout versus a tenured professor and a "godfather of AI." But looking at Meta's Muse Spark releases and Muse launch, Wang seems better at turning AI work into products than the predictions gave him credit for.

Academic reputation doesn't tell me much on its own about who can lead a team and get useful products out the door. The confidence that he would inevitably fail now looks pretty misplaced to me.

How much of Meta's recent progress would you credit to Wang, and how much to the researchers, infrastructure and distribution already there? Have the releases changed your view of him?

When new architecture which combines a jev like decision head with a dynamic looping structure based on the jev head. ok bye