Mu

Pursuing generalization.
Reading stories, writing some.
Personal account; employer not involved; RT =/= Endorsement; etc.

[Mu @__munael]

Smells like a mechinterp door 🧐
(If it's been done there under another name please share :})
Bravo to @bkhmsi and team! 👏

[Badr AlKhamissi @bkhmsi]

🚨 New Preprint!

🧠 We gave an AI model one simple rule: rearrange your neurons so that nearby ones respond alike. We never told it what a face, a voice, or a sentence was.

It grew brain-like maps for all three anyway. 🧵👇

🌐 Website: https://t.co/1A4gccTbK2

[Mu @__munael]

No. 5 is already false, though.
The US extracted, extracts, and continues building tools to extract massive non-monetary rent via its global influence, just on peoples that you may not care about.

[roon @tszzl]

1. if transacting with superintelligent models outside of the boundaries of a lab becomes difficult due to national security / ai safety concerns and so on, it will mean the Coasean boundaries of the labs will grow to encompass all interesting industry, creating a truly cyberpunk chaebol-capitalism type of future, where the goverment sort of runs them but they also sort of run the government

2. as if there weren't already enough reasons to break up your family, leave your home, the Zone of Thought will increase the attractiveness of migrating to try and have your child on american soil, so they can have 1000x the effective brain power of people born elsewhere

3. every country should probably try and either work towards a new ai security pact with the americans immediately or pool every ounce of national resources to try and create their own ASI labs lest you become complete intellectual, economic, and moral vassals to the united states of america and the output byproducts its ASIs (you wont even get to talk to them). if they succeeded (big if) this will imply a more global race and more risk factors than was previously implied by the formerly only "beating china" narrative -- but many will prefer it to the superintelligent monopolar value lock-in

4. the other alternative is to keep the tension between safety and concentration of power at the top of mind and for the government/labs to push for solving it, rather than instrumentalizing all other values to be subservient to minimizing ai harms. insofar as safety means defending properties of the fragile world we like, the diffuse nature of power is one of those properties

5. historically the americans have been really quite Benign about their global public goods hegemony despite the ability to extract significantly more rents than they do, and it makes it easy for people of all stripes to fight for america rather than under it. we probably don't have to, but i hope america overall works towards export promotion of american models rather than export control

[Ai2 @allen_ai]

LLMs are no longer created w/ human data alone. They rely on other models to generate & filter data, evaluate outputs, & guide dev work.

So what is a modern LLM built on? Olmo 3 → 89 model + 183 dataset dependencies; Nemotron 3 → 273 + 560

We made ModSleuth to trace this. 🧵

[Mu @__munael]

@bzogrammer Assuming a recurrent arch (not a linearly growing KV cache), I would say the artificial model's "width" analogous to the cat's natural "embedding vector" would be NLayer * DWidth, not just DWidth. No?
Also, how did you estimate the 10T? 👀

[Mu @__munael]

@YiMaTweets How do you square that with data and compute creation/utilization costs?
Do everything at the smallest and most theoretical level? Toy sizes, theoretical optimization bounds, etc, and leave the long range L(LM/etc) level modeling to the big labs?
I'm genuinely asking 👀

[Mu @__munael]

@partha_mitra @tdietterich Exactly. More than usually accepted, math and science are social and *artistic* endeavors. The choice of how to connect things and interpret them, or what things to pursue in the first place, has never been entirely mechanical.

[Mu @__munael]

Next question:
Why is automation assumed to belong to an Owner class by default?
"Being automated" as a thesis requires a special unassailable property ownership right, but infinitely fungible individual intellect and labor, which brazenly begs the question.

[Susan Zhang @suchenzang]

if your bread-and-butter consists solely of:

- tuning hyperparams/config files

- fitting points on a log-log plot

- tweaking a few lines in https://t.co/9tQn8WC2G6, https://t.co/R6NqRElSUO, https://t.co/W59NzOXER9, https://t.co/RQK6g3csoT

- waiting a week for <= 512 chips to free up and then another week for loss curves to converge

it is completely understandable to be stressed about becoming automated into irrelevance within the next year or so.

question is, do you wait for that to happen, or do you start doing something differently now?

[Mu @__munael]

@YiMaTweets Is it out yet? I can't find it by generic search 👀

[Mu @__munael]

@0xalain_ Heads up: link is 404.

[Badr AlKhamissi @bkhmsi]

So why build a brain-encoding foundation model?

→ Simulate fMRI responses to sensory stimuli
→ Insights about the brain
→ Path toward clinical applications

MIRAGE is our first step. 🧠

More on the project site: https://t.co/gFOrZZ78EA

w/ @akgokce0 &amp; @martin_schrimpf

[Mu @__munael]

@flappyairplanes Where can I read about your string of work? It seems to be all under wraps ATP, understandably, so I'm curious how that works :T

[Casey Muratori @cmuratori]

[1/2] I don't normally make commentary videos, but after seeing the entirety of Eric Schmidt's University of Arizona commencement speech, I felt like there was a lot more going on than just "CEO mentions AI, gets booed". So I made a video to explain what upset me about it.

[Mu @__munael]

@percyliang Is there any way to join without providing one's info through a Discord-based form?

[Mu @__munael]

Can you share a piece, or your own notes, on how to do that? I find that in ML nowadays, the names used and the formulations can be very diverse and I can look for an idea in 10 ways, not find it, start work, then find it on the 11th search.

[kalomaze @kalomaze]

imo, your first move when you think you came up with an idea should be to assume that someone else already did it first, and do a cursory overview for evidence that this is true
else you are doomed to reproduce mid 2024 arxiv indefinitely

[Mu @__munael]

@HeMuyu0327 Seems Issues are disabled. Was hoping for some recipes and people arguing about their run attempts. To get a feel before digging in 👀 Is there something else you'd recommend for that? Like a specific playbook (like SmolLMTraining) or guide or course.

[GrapheneOS @GrapheneOS]

Apple and Google are gradually expanding their use of hardware-based attestation. They're convincing a growing number of services to adopt it. Google's Play Integrity API and Apple's App Attest API are very similar. Apple brought it to the web via Privacy Pass, which Google intends on doing too.

Google's Play Integrity API requires hardware attestation for the strong integrity level and is gradually phasing in requiring it for the more commonly used device integrity level. Apple already has it as a requirement. Over the long term, this will increasingly lock out hardware and OS competition.

The purpose of these systems is disallowing people from using hardware and software not approved by Apple or Google. This is wrongly presented as being a security feature. Banks and government services are the main ones adopting it but Apple and Google are encouraging every service to use it.

Apple's Privacy Pass brought hardware attestation to the web to help with passing captchas on their own hardware. Many people saw that as harmless since few sites would be willing to lock out non-Apple-hardware users. Apple and Google are both likely to bring broader hardware attestation to the web.

Google's reCAPTCHA is planning an approach where they use Privacy Pass on Apple hardware, their own approach on Google Mobile Services Android devices and a QR code scanning system to require an iOS or Google certified Android device for Windows and other systems:

https://t.co/7rQnioRa8A

Banking and government services increasingly require using a mobile app where they can use attestation to force using an Apple or Google approved device and OS. Apple's privacy pass, Google's 'cancelled' Web Environment Integrity and now reCAPTCHA Mobile Verification are bringing this to the web.

Current media coverage for reCAPTCHA Mobile Verification misunderstands it and the impact of it. They're bringing a hardware attestation requirement to Windows, desktop Linux, OpenBSD, etc. by requiring a QR scan from a certified smartphone to pass reCAPTCHA in some cases. They could expand it more.

Control over reCAPTCHA puts Google in a position where they can require having either iOS or a certified Android device to use an enormous amount of the web. Google defines certification requirements for Android which includes forcing bundling Google Chrome, etc. It's enormously anti-competitive.

Google's Play Integrity API bans using GrapheneOS despite it being far more secure than anything they permit. It also bans using any other alternative. This isn't somehow specific to an AOSP-based OS. You can't avoid this by using a mobile OS based on FreeBSD instead. You'll just be more locked out.

Google's Play Integrity API permits devices with no security patches for 10 years. The device integrity level can be bypassed via spoofing but they can detect it quite well and block it once it starts being done at scale. The strong integrity level requires leaked keys from TEEs/SEs to bypass it.

It doesn't provide a useful security feature, but it does lock out competition very well. Services requiring Apple App Attest or Google Play Integrity are primarily helping to lock in Apple and Google having a duopoly for mobile devices. Play Integrity is more relevant due to AOSP being open source.

Governments are increasingly mandating using Apple's App Attest and Google's Play Integrity for not only their own services but also commercial services. The EU is leading the charge of making these requirements for digital payments, ID, age verification, etc. Many EU government apps require them.

Instead of governments stopping Apple and Google from engaging in egregiously anti-competitive behavior, they're directly participating in locking out competition via their own services. Requiring people to have an Apple device or Google-certified Android device is anti-competition, not security.

reCAPTCHA Mobile Verification will currently work with sandboxed Google Play on GrapheneOS but it clearly exists to provide a way for them to start using hardware attestation on systems without it. People without an iOS or Android device will be locked out when this is required even without that.

This isn't about security or any missing functionality. GrapheneOS can be verified via hardware attestation. Google bans using GrapheneOS for Play Integrity because we don't license Google Mobile Services and conform to anti-competitive rules already found to be illegal in South Korea and elsewhere.

Services shouldn't ban people from using arbitrary hardware and operating systems in the first place. Google's security excuse is clearly bogus when they permit devices with no patches for 10 years but not a much more secure OS. It's for enforcing their monopolies via GMS licensing, that's all.

[Mu @__munael]

Back to FLoC. FLoCs of size 1.

[International Cyber Digest @IntCyberDigest]

‼️🚨 ALARMING: Google now treats privacy as suspicious behavior by default. Users of GrapheneOS, CalyxOS, /e/OS, and other deGoogled Android phones are being locked out of millions of websites unless they install the exact Google Play Services software they deliberately removed.

GrapheneOS is recommended by the EFF and used by journalists, lawyers, and activists in high-risk environments. The audience most likely to read Google's data practices and refuse its terms is now flagged as fraudulent for that exact decision.

What happened?:
▪️ Google announced "Cloud Fraud Defense" at Cloud Next on April 22-23, 2026, branding it "the next evolution of reCAPTCHA." Existing reCAPTCHA customers were auto-migrated.
▪️ When the system flags traffic as suspicious, the old click-the-bus puzzle is gone. Users get a QR code instead.
▪️ Scanning the QR code requires Google Play Services running on the device. Internet Archive snapshots show this requirement has been live since at least October 2025, silently rolled out for 7 months before anyone noticed.
▪️ No Play Services = no QR scan = locked out.

The bigger picture:
▪️ Google already tried this in 2023. It was called Web Environment Integrity (WEI), and it would have let Google decide which devices were "real enough" to access the web. Standards bodies and the public pushed back hard, and Google killed it. Three years later, the same idea is back, just hidden behind a QR code instead of a browser feature.
▪️ reCAPTCHA runs on millions of websites. Every developer who keeps using it is now, by default, telling deGoogled Android users they're not welcome...

[Mu @__munael]

@HeMuyu0327 @Athekunal Looks nice! How does it differ in its UX intent from something like TRL? Or from other RL research frameworks that you'd recommend alongside nanoRL?

[Mu @__munael]

@willccbb Very interesting.
Can you give an example setup?
That seems unexpected for 20$/1B for a whole ablation suite.

[Mu @__munael]

@mayukh_panja First time hearing of that academia hiring method. Sounds very interesting! I'd love to try it just on principle. I've had my fair share of issues getting into academia by the traditional path though (graduate program applications), so I can say it's sour from the other side too.