High-functioning millennial ■ AI 4 XR @ Meta ■ Prev: NVIDIA Research, XR, Omniverse ■ VFX⎯Mathematics⎯Electronic Elsewheres ■ Opinions mine, you can't have them
Joined July 2009
- Tweets 38,108
- Following 1,631
- Followers 5,291
- Likes 37,561
3,089 Photos and videos
Pinned Tweet
1 Jul 2023
All my socials are here: omershapira.com/about-me/
Tangentially: what the actual fuck
2
22
15,408
May be obvious to people who studied baroque, but I’ve always been fascinated by the way Johnny Marr creates polyphony in SOME GIRLS ARE BIGGER THAN OTHERS, but it didn’t hit me until I started practicing Bach’s JESU, JOY OF MAN’S DESIRING that they’re doing similar things
1
178
Funny how a Neal Stephenson story “World’s hottest SWE model was post-trained by Rio de Janeiro’s municipal government” lasted a few hours until it turned into a Martha Stewart recipe
23h
The Rio 3.5 model broke the internet this week. The plot twist? It’s essentially our open-source model, Nex N2 Pro, wearing a different hat.
🤯 We analyzed the weights, and the recipe is exact: Rio 3.5 ≈ 0.6 * Nex N2 Pro 0.4 * Qwen 3.5
It even literally introduces itself as "Nex N2 Pro" if you ask it without initial system prompt!
😂 We are flattered that the City of Rio used our work to achieve SOTA performance. Thanks for the ultimate benchmark validation.
🤝 But in the open-source world, attribution matters.
👇 Full mathematical proof & verify script in the first reply!
1
352
SIMD Crawford 🟣 retweeted
Jun 13
U.S. Aerospace companies have long known they can’t hire non-US persons to work on tech that affects national security. The companies build technology firewalls between U.S. and non-US divisions. It’s kind of shocking that AI companies didn’t know they’d have to do this, too.
Jun 13
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.
The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Access to all other Claude models is not affected.
We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.
Read our full statement: anthropic.com/news/fable-myt…
66
54
841
80,153
SIMD Crawford 🟣 retweeted
Jun 6
Apocalyptic bird nest.
A Russian glide bomb knocks down a tree in Donbas. From the shattered branches rolls out a tiny bird’s nest.
Made of drone fiber-optic cable.
Source: Oleg Malchenko
146
3,940
28,391
1,022,821
Jun 6
If you, like me, have been pushed cockatiel videos on Instagram, you probably noticed many of them sing a variation on one song, something to do with “cookie” (look it up).
Today I discovered that this song was INVENTED BY A COCKATIEL AND SPREAD TO OTHERS VIA SOCIAL MEDIA.
1
9
774
Jun 6
Cockatiels don’t have access to instagram. Their owners do. These birds have independently discovered a variant of Snow Crash that works equally well on Humans and Cockatiels.
2
2
726
SIMD Crawford 🟣 retweeted
Israel and Lebanon are very similar and there is no reason why they should not be friends and allies. Together, the sky will be their limit. They can also pull the region out of its backwardness and misery.
📊 U.S. billion-dollar startup companies founded by immigrants from different countries, normalized by population (founders per 100 million inhabitants).
Israel 🇮🇱 and Lebanon 🇱🇧 stand out as the top 2 countries in producing founders of U.S. billion-dollar startups. Their exceptional performance highlights a strong culture of entrepreneurship, ambition, adaptability, and innovation that far exceeds what their population size would suggest.
Source : National Foundation for American Policy.
6
10
82
6,845
SIMD Crawford 🟣 retweeted
I’ve been capturing 3D human motion for 30 years and today is maybe the biggest day in that history. We are presenting MAMMA at CVPR (oral session 2A). MAMMA is a markerless multi-camera system that has accuracy similar to marker-based systems.
52
258
2,123
230,311
Jun 4
Missing CVPR this year. But it's ok, if really I needed to be in a trade hall full of 30-somethings on amphetamines standing in front of canvases praying for an industrialist to find them, the Armory Show is just around the corner
3
298
Jun 3
Tired of your useless GPU hoarding its' precious RAM? Show that bitch who's boss by using VRAM as swap space. Breathe a sigh of relief as the gmail tab takes 15ms less to load
github.com/c0dejedi/nbd-vram
1
10
864
Jun 3
Jun 2
C devs love Undefined Behaviour so much,
They even made a function to introduce UB on purpose.
10
501
Jun 2
This is the take I agree with. Memory bandwidth half of M5 Max could matter less depending on how the 128GB unified memory is scheduled.
M1 and GH were both compelling evidence that unified memory solves more pain to devs than it introduces to the market.
Jun 1
I did not expect Nvidia to challenge Apple's unified memory monopoly with a Windows on Arm chip, but the RTX Spark is exactly that. 128GB of LPDDR5X with native CUDA means we can finally run vLLM, Triton, and TensorRT-LLM on a laptop without fighting Metal. This is so good??
1
3
1,142
Jun 2
u gotta wonder what the downstream effects of DDR5 loosies being effectively extinct will be in a year.
1
306
Jun 1
Even my best attempt at putting jokes in past NV keynotes pales in comparison with this
1
12
3,416
May 30
In A MATHEMATICIAN'S APOLOGY (1940) G. H. Hardy, depressed from waning physical condition and with a general sense of doom about World War II, writes about the futility of assigning value to a Mathematician's human endeavour.
1
1
613
May 28
gulp
For over a decade, we’ve accepted that end-to-end backprop is the only way to train deep networks. But holding the entire network in memory all at once is why AI training is hitting a resource wall.
We found a new way to break the network into blocks and train them independently. The trick? Treating the network’s forward pass like a diffusion model denoising a signal.
This reinterpretation slashes the memory needed to train deep models. In our #ICLR2026 paper (arxiv.org/abs/2506.14202), we matched end-to-end performance across ViTs, DiTs, and LLMs. We did this while training just one isolated block at a time.
1
2
563
May 27
*click* Idea for a program: DOOM, except it’s a battle for child custody
1
489