I fear not the man who has made 10,000 apps once, but I fear the man who has made one app 10,000 times
Dave W Plummer@davepl1968
Hey look, it's your old friend Task Manager... on the Mac!
English
Atmadeep Banerjee
284 posts
Atmadeep Banerjee
@abanerjee99
Building @sabi. Prev: mapped brains @HarvardVCG, taught satellites to see @PixxelSpace, drones @join_ef. Future is wonderful.
Palo Alto, CA Katılım Eylül 2019
285 Takip Edilen2.3K Takipçiler
@jalajupadhyay @thegautamkamath Celeb Profs with big labs? It's clearly for 20 last author pubs and not 20 first author pubs.
English
@thegautamkamath Would have preferred smaller threshold for max submission! Who can write 20 papers in a year let alone a single conference?!!
English
ICLR 2027 has authorship quotas:
- No more than 20 submissions per author
- No more than 1 submission where no author has a paper previously accepted to a major ML conference
ICLR@iclr_conf
The ICLR 2027 CfP is up! iclr.cc/Conferences/20…
English
Genuinely why do we put up with this format? Why not make OLED sheets that scale to 1000 in. diagonally and simply replace projection. A massive 16k screen like in the Vegas Sphere blows any projection tech out of the water.
Eric 🐇@Mammoth
Wow
English
Time to make the new impossible chicken
trash@trashh_dev
people who do local ai are the new vegans
English
Enough with the O1-As. Time to bring in the O1-Bs.
“paula”@paularambles
they’re doing the thing
English
@francoisfleuret At Q4 40B is 20GB not 20Gb. This is saying 2.6bpe so about 13GB active. The full 750B at 2.6bpe is about 243GB. At 128GB unified memory he is caching about half the model in memory and half in SSD. Cache misses are likely very forgiving.
English
I do not get it:
- GLM-5.2 is 750B parameters, 40B active= 20Gb,
- A top SSD can do *in theory* 15Gb/s.
how can you do better than <1 token/s ?
antirez@antirez
GLM 5.2, Q2_K routed experts (effectively ~2.6 bits) running with SSD streaming on an M5 Max 128GB computer.
English
@SwayStar123 Apparently a finetune of Kimi K2 thinking
English
Lol, the text diffusion research team at DeepMind has existed before Inception was even founded.
Also bizarre to compare a 1T parameter closed source model to a 26B open weights model 🙃
Inception@_inception_ai
Welcome to the diffusion era. We bet on parallel generation years ago, when it was a contrarian idea. It's great to see the industry arrive. Mercury 2 continues to lead the Pareto frontier for quality, speed, and cost among publicly available diffusion LLMs.
English
Imagine if Ant had IPO'd around Fable release and K3 came out today. Recession in the Bay Area.
English
Resisting the urge to ask Claude for a status update on experiments (it will burn $50 in credits)
English
Linux is so funny sometimes. Tried to print 2 photos together with GwenView and apparently you cannot do that. A bug report (bugs.kde.org/show_bug.cgi?i…) was filed in 2008. Last activity Jan 2026. For 18 yrs people havent been able to come to a consensus about how to solve this.
English
Atmadeep Banerjee retweetledi
Impressive results from @alephneuro producing rich ultrasound images using contrastive bubbles in arteries. (previous SOTA was 25 microns with ultrasound and contrastive bubbles [1])
It's quite amazing how prescient and remarkable MindEye[2] has been in laying the groundwork for non-invasive neural decoding, spawning many of the most interesting works in the field.
In 2023, @abanerjee99 and @humanscotti had the unique foresight to use large fMRI datasets and train deep learning models to decode what a person was looking at, purely from neural signals.
The consensus in the field at this time was that the only way to do meaningful decoding is via drilling a hole inside the skull and putting a chip inside the brain, which, if it had its way, would be an extremely morbid future for humanity.
When we first saw the MindEye reconstructions, it was clear we would finally be able to build a cap that could decode thoughts and bring telepathy to humanity at scale. That's also exactly when @abanerjee99 and I decided to build @sabi together.
----
1. Transcranial ultrafast ultrasound localization microscopy of brain vasculature in patients
pmc.ncbi.nlm.nih.gov/articles/PMC76…
2. Reconstructing the Mind’s Eye: fMRI-to-Image with Contrastive Learning and Diffusion Priors
arxiv.org/abs/2305.18274
Aleph@alephneuro
We think this problem is newly tractable. In 2023, MindEye showed astonishing reconstructions of seen images from fMRI brain data. On the left: the image that was shown to a subject. On the right: the image predicted from the subject’s brain activity.
English
Incredible work. Mindeye keeps spawning the most goated research.
Aleph@alephneuro
We recently obtained the highest-resolution 3D images of the human brain ever taken from outside the skull. This is the first look. Introducing Aleph, a research lab building brain interfaces for the telepathic future. (1/n)
English
@secemp9 You cannot buy a "single GB300". GB300s I think are only sold at rack-scale. GB200 you can buy as a single unit of 8xGB200s for around $600k last time i checked.
English
so let me get this straight, you claim 2x tinybox or 1x tinybox "pro" is around $150k while the GB300 is both cheaper and more powerful?
the tiny corp@__tinygrad__
I have on good authority that GLM 5.2 is running at 120 tok/s across two networked Blackwell tinyboxes. $150k and that setup can be yours, either 2x tinybox or 1x tinybox pro. Never pay the cloud again.
English
@mihirneal They don't know how. They don't have the technology.
English
