Neutral net 🕸️

in the next 2 minutes, I'll walk you through why every single AI company, including Nvidia, was unnecessarily sacrificing model intelligence, speed, and often both... and how we fixed it the tldr: - error(layer a) + error(layer b) < error(layer a) alone - quantizing MORE of the model can result in the same or HIGHER quality model, if you know which layers to quantize. let's take GLM5.2 architecture as an example. tokens here go through: 1) normalize 2) attention 3) normalize 4) moe repeated 78 times moe: > just a big mlp, split into 257 separate smaller mlps (experts) > 8 are routed and one is shared (only 3.5% of the mlp is activated) so we say, let's make the matrix multiplications faster. we quantize the matrices to nvfp4 but what do you quantize: every single one of the 257 experts, the attention projections, both? Nvidia, and every other quant scheme, settles on a somewhat intuitive solution: """ a) there is only one router per layer, only one shared expert, and only one output projection per layer. b) every single token goes through these, so they must be pretty important, and if we quantize these, errors will compound across the xN number of tokens input. c) so let's leave the shared experts and projections unquantized. d) let's quantize ONLY the 256 routed experts """ -this is literally nvidia/GLM-5.2-NVFP4 intuitive. simple to reason about. wrong. empirically: if you have a model that processes its input through layer a, then layer b, it is possible that a scheme quantizing BOTH layer a and layer b performs better than a scheme quantizing just one of these. by way of example, our config quantizes every single out projection (all 78) of the attention, and 43 of the 78 shared experts. Nvidia quantizes none of these. we score the same on benchmarks (quality). we get a free 20% higher throughput. i leave the math proof as an exercise for the reader in the paper... and some pictures to click through if you'd rather skip the (43 pages of) math

Ask and ye shall receive. Heres our paper on how we made a SOTA quantization method using Fourier Analysis on Groups 🧵 We achieve 20% higher throughput on GLM 5.2 compared to the existing configs while matching downstream quality.

Today, we’re announcing Bonsai 27B: the first 27B-class model to run on a phone. Bonsai 27B is the new multimodal flagship of the Bonsai family. Based on Qwen3.6 27B, it brings a new capability tier to local AI: multi-step reasoning, structured tool use, long-context workflows, and coherent agentic loops. Until now, models in this class have been impractical to deploy locally. A 27B model occupies roughly 54 GB in 16-bit precision, and even a strong 4-bit build is around 18GB - too large for a phone and for most laptops. Bonsai 27B changes that. It comes in two variants: • Ternary Bonsai 27B: 5.9 GB, 1.71 effective bits per weight, optimized for laptop-class quality. • 1-bit Bonsai 27B: 3.9 GB, 1.125 effective bits per weight, optimized for phone-class footprint. Everything is open-sourced today under the Apache 2.0 license.

@lukepierceops Wait till the maintenance debt kicks in for the in house system

Starbucks spends $400 million a year on software. Yesterday they announced they're moving off IBM and Microsoft to build their own custom systems in-house. IBM dropped 3% and Salesforce dropped 4% on the news. And honestly this is, unequivocally, the biggest signal I've seen since OpenAI and Anthropic launched their consulting arms back in Q1. The largest companies in the world are done paying for software that half fits how they work. We saw this coming about a year ago. Moved everything we build off Airtable and low-code tools and went fully custom. Already paying off, and it's only going to compound from here. This is the opportunity right now. You get all of a company's data into one system. You build out a single operating system for the entire business. You cut out bad, redundant processes. Then you layer AI on top of it, under the correct processes. That's the core of AI consulting. Helping companies actually operate better. There are a lot of fly-by-night offerings circulating right now when it comes to Ai Services. For example, 'second brains'. Throwing scattered data into a second brain while the processes underneath stay broken does nothing. The companies who will absolutely destroy their competition over the next 5 years are rebuilding how they work from the ground up. Starbucks is showing you what other companies will be doing over the next several years. Your job is to position yourself to facilitate that process for as many companies as you can.

@_svs_ Last line is the most important

the whole 'read the code' debate headline also misses one more nuance - read the code when? glance at it as the llm works or review every changeset? i do read what the llm is doing. i don't read the code as it lies on disk anymore. I'm probably doing this wrong as can't handle more than 4 concurrent sessions this way but it works for me so.... also bugs are not very serious for me. ymmv.

@paulg Dual nature of students, the moment you look at them their score changes

A Brown professor gave his students a take-home midterm exam. After suspecting many cheated using AI, he made the final in-person. The orange dots are the midterm scores and the gray dots are the final scores. Looks like all but 3 cheated on the midterm.

@siddharthshahx By volume or by revenue?

TIL: India is currently the second largest enterprise market for Anthropic

@Entity3Self @TrisH0x2A It can be used by garbage collectors to find smart pointer cycles

@TrisH0x2A 🤷 I've heard of the algorithm, but never of a situation where the algorithm would be needed in practice

Floyd's cycle detection algorithm finds a loop in a linked list using just two pointers and O(1) extra space one pointer moves one node at a time while the other moves two if a cycle exists they are guaranteed to meet after they meet reset one pointer to the head and move both one step at a time the node where they meet again is the exact start of the cycle one elegant proof from modular arithmetic powers one of the most famous linked list algorithms ever written

@martinmbauer @blackstar31416 And for gluons?

@blackstar31416 Resistance for W and Z bosons, not for photons

It’s a pity the Higgs boson’s deep impact hasn’t been communicated better. It wasn’t profound because it was unexpected, but because it gave overwhelming evidence that the vacuum of our Universe behaves, in a deep sense, like a superconductor. Permeated by a field that breaks a fundamental symmetry of nature and shapes the forces within it. Democritus imagined atoms more than 2,000 years before they became experimentally real to us. That didn’t make the discovery any less revolutionary. The Higgs was that kind of moment.

@rashi_kakkar Which?

Me: I really want to spend more of my free time reading and writing … Also Me: Downloads yet another text based social media app on my phone 😭🫣

@sanjeevsanyal Penal interest was already there

EPFO radically simplifies withdrawal rules: Full withdrawal in case of VRS, disability or emigration; full withdrawal at 55 years; 75% for unemployment. Settlement to be done in 3 days (with 12% penal interest on delays by officials & recovered from EPFO commissioner's salary). economictimes.indiatimes.com/news/india/new…

@paraschopra They are the same problem

The two big unsolved problems in ML today are: - Non-stationarity: future inevitably is different from the past, but your model remains stuck in regularities it learned from the past. How do you adapt? - Low coverage regime: your model spends its capacity on frequently occurring patterns, but rare ones do crop up every now and then but it hasn't learned them. How do you deal with them? These two problems come in many guises (continual learning, OOD generalization, hallucinations, sample efficiency and so on) but the common core is this.

@paulg The two parts in your first sentence are contradicting?

Students need to learn how to use AI, but they also need to learn how to think for themselves. So in schools there should be some kinds of work where students are expected if not required to use AI, and others where it's banned, and no mushy middle ground in between.

@LostMemeArchive The 8-bit signed integer might have overflowed with 127 USB devices

In 1998, Microsoft demoed Windows 98 live on stage. They plugged in a scanner to show off USB support. The computer immediately blue-screened in front of Bill Gates. He just smiled and said: “That must be why we’re not shipping Windows 98 yet.”

All day using GLM 5.2. Didn't miss much. First open model that passes the bar as a daily driver. Things are not going to be the same. Damn, now I want to buy some serious hardware.

@shmidtqq He's creating a fable 😂

Anthropic CEO: "If my revenue is not $1 trillion, even $800 billion, there's no force on earth, no hedge on earth, that could stop me from going bankrupt." In a 3-hour podcast, Dario Amodei does the math on his own bankruptcy. Revenue 10x a year. 90% of code written by the model. A country of geniuses by 2028. He can't tell you if it ends in trillions or zero. The most honest voice in AI, or the biggest bubble admitting it?

@stats_feed Sitting (sedentary lifestyle)

What is a silent killer that people dont realise is slowly killing them?

@paraschopra It could if it were entangled

I used to believe in panpsychism, but now I’m not too sure. Take an electron, for example. It simply doesn’t have enough representational capacity to encode complex feelings (and, frankly, no need). Where and how would it represent pain, pleasure and all the gamut of emotions and qualia that make up our world? Even if it encodes basic valence somehow, how is the readout happening and is it influencing behaviour. In retrospect, panpsychism seems significantly handwavy. The binding problem - how simple qualia bits combine to produce our rich experience - is a showstopper.

@shikhr_ No it'll go on 4.9999... 😂

After Opus 4.9 Anthropic has no choice but to release Opus 5

@ians_india Same energy

Chandigarh: On CBSE's On-Screen Marking system, CBSE Regional Head, Rajesh Kumar Gupta says, "Regarding your question about the website being hacked, I completely deny it. I am rejecting this allegation outright. Because exams are being conducted offline so there are no questions of website being hacked..."