Obsessed by automation. Unstoppable. I tame and teach machines 🇪🇺🇮🇹 "Through the fire, and the fury, with a heart made of steel". OSS/Acc.

Milan, Italy
Joined May 2011
Jacopo Nardiello retweeted
-----BEGIN EC PRIVATE KEY----- 4 -----END EC PRIVATE KEY----- 😬
82
623
66
9,197
197,451
Jacopo Nardiello retweeted
We’re releasing more open models to help teams across Europe build AI with their own data, on infrastructure they control. Announced at GTC Berlin, the models help developers, researchers and enterprises build specialized AI agents, train robots, design proteins and develop quantum computers. See what’s new: nvda.ws/4iwBMvF
32
50
30
382
27,113
Jacopo Nardiello retweeted
Running GLM-5.3-Flash on 2× DGX Spark with NVIDIA Flare ✨ - 1M context window, 2.32M-token shared KV pool - Prose decode, single stream: 117.2 tok/s - Prose decode, 2 concurrent: 167.4 tok/s aggregate - Prose decode, 4 concurrent: 215.8 tok/s aggregate - 32K prefill: 2,327.9 tok/s - Available memory after loading: ~13 GiB on head /worker each Checkpoint: GLM-5.3-Flash-NVFP4-MXFP8-CSF-QAD github.com/NVIDIA/flare-loca…
NVIDIA just released Flare, a local inference server built for DGX Spark and RTX Blackwell GPUs - Built for long chats, agents and tool calls - Serves from Rust, no Python - Plans memory up front - FP8 KV cache by default Interesting timing... right around when DeepSeek open-sourced their inference engine. Link below.
64
27
18
265
36,851
Jacopo Nardiello retweeted
Leaving this here, PHP 8.5.12 security patch is incomplete. ( needs further validation )
Readers added context they thought people might want to know
Unvalidated AI output (author admits in replies: omp with glm5.3 + ds4), not a lab result. Its screenshot says both CVEs are fixed; the 'surviving paths' need unserialize() of attacker data or an app LFI. Patch per php.net: php.net/archive/2026.p…
76
364
145
2,565
466,119
We gonna hunt down 11yo vandals looting stores with our chinese robot army.
I need few of these for some experiments in Milan to defend the center of the city.
101
Jacopo Nardiello retweeted
yesterday was fun: - Revolut down for 9+ hours - 8 intermediate CA private keys discovered - PHP CVSS 10 vulnerability announced/leaked - TiDB whole infra/data exfil
18
57
7
870
33,242
Jacopo Nardiello retweeted
deepseek posted way more alpha on their zhihu page, if you only read twitter you are reading the laowai approved stuff "Based on this framework, we have implemented a high-throughput, low-latency large-scale inference service for DeepSeek V4.1 Flash. Currently, the service throughput of a single cluster is 155 trillion tokens/day, and the overall revenue is $38 per card per hour (based on off-peak prices)." it costs roughtly $5-10 per gpu hour so depseek is claiming 80% serving margin at THEIR prices lololololol
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source! Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable. 🔗 github.com/deepseek-ai/DeepI… 1/4
105
141
31
2,286
135,073
Jacopo Nardiello retweeted
DeepSeek infer is the cloud version of DwarfStar, conceptually. Vertical inference engine for one model and specific hardware. Good move. And well, a lesson to the ones that criticized this approach as well.
24
26
1
797
30,123
Jacopo Nardiello retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source! Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable. 🔗 github.com/deepseek-ai/DeepI… 1/4
1,551
4,768
2,295
42,177
6,400,310
Own your weights and open models.
SITUATION DETECTED: The US government is banning DeepSeek from all civilian federal agency IT systems under national security.
4
360
Jacopo Nardiello retweeted
A fly that died years ago is flying around a harbor that doesn't exist. Here's what's going on. Scientists took one female fruit fly and sliced its brain into about 7,000 pieces, each roughly 2,000 times thinner than a human hair. They photographed those slices 21 million times under an electron microscope. Then hundreds of people spent years tracing every connection by hand until they had the whole thing, all 139,255 neurons and 54.5 million synapses. That map is called FlyWire. Odyssey hasn't said exactly which map it used, but FlyWire is the one most of these projects build on. Earlier this year a startup called Eon ran that map as a simulation with zero training and got a virtual fly that walks and grooms on its own. Now Odyssey has put the brain somewhere new. The harbor in this clip isn't sitting on a server as a 3D model or a game level. Odyssey-3 is a 14-billion-parameter AI that learned how water and light behave by watching a ton of video, and it invents each frame as the fly steers through it. This is kind of a big deal for a nerdy reason. For as long as people have tried to build a mind, there have been two camps. One says copy a real brain. The other says train a giant model on tons of data. Those camps have mostly fought each other, and in this clip one is living inside the other. And here's a fun bit of math. The fly has 54.5 million synapses. The AI dreaming up its world has 14 billion parameters. Line those up and the world is about 250 times bigger than the brain living in it. That's what happens when one approach rides on GPUs and the other rides on grad students with microscopes. The trained side ballooned in a couple of years. The copied side took the better part of a decade to get one fly. What really got me is how little it takes to fool a fly. A fruit fly sees through about 800 tiny lenses per eye, so its whole view of the world is around 1,600 pixels. Each of your eyes sends about a million nerve fibers to your brain. Keeping a fly convinced means the AI only has to look right across 1,600 pixels, which is a pretty low bar, and it's clearing it. The fake world will look good enough for human eyes long before anybody can slice up a human brain and rebuild all 86 billion neurons.
We placed the fruit fly brain inside Odyssey-3, a frontier world model. The lil’ guy is now flying around totally imagined, simulated worlds all by itself 🫶
74
246
23
2,546
597,121
Jacopo Nardiello retweeted
Si vous avez téléchargé aujourd'hui la dernière version de CLI d'OVHcloud, merci de la mettre à jour à nouveau très rapidement. Sinon effacer. Il s'agit de repo: github.com/ovh/ovhcloud-cli La version CLI de 14h47 a été compromise par un hacker. Nous l'avons détecté et bloqué à 16h21. Merci au client qui nous a remonté le souci. La version compromise est v0.11.0 v0.12.0 v0.13.0 v0.14.0 v0.15.0. La version v0.10.0 et la future 0.16.0 est clean. On vient de finit les premières vérifications et aucun autre repo n'est affecté. Nous avons aussi retracé l'incident. Le hacker a utilisé le compte github d'un des développeurs et a ajouté un binaire dont le job est récupérer les secrets sur le disque. La qualité du binaire est médiocre mais il peut aller jusqu'au bout. Si vous êtes concernés, merci de changer tous les secrets sur votre disque. On continue les recherches sur le binaire et on va publier les détails additionnels sous peu.
58
420
40
924
149,654
Jacopo Nardiello retweeted
Full GLM-5.3 (753B) TP4 on 4× DGX Spark on @mmastrac's gr8 engine: recipe v0.9 🐉 • Prose decode: 40.6 tok/s at c1 with thinking on (RigMark) • +40 % prose over upstream gr8 on the same boxes (29.0 → 40.6) • Code 60.6 / structured 75.5 tok/s at c1 • Aggregate: 47.0 / 59.9 / 76.6 tok/s at c1 / c2 / c4 • Prefill: ~1.38k tok/s at 8k and 32k, 1.23k at 128k • 🔥speedbench: 49.4 tok/s average over 73 prompts (prose 42.9, code 53.4, structured 58.2, agentic 50.3, long context 48.9) I'm also releasing speedbench. RigMark runs one prompt per workload, so at T=0 it measures how fast one particular text decodes. Any change that alters the output can move it ±2–3 % by luck alone. Once it showed +3.5 % where the real gain was ~1.5 %. speedbench uses 73 prompts across prose, code, agentic, multilingual and long context, pairs the A/B runs and gives a confidence interval. It works with any OpenAI-compatible server. Still early and I'm still testing it, so feedback is welcome. Recipe, measurements and credits: github.com/knapcio/GLM-5.3-g… speedbench: github.com/knapcio/speedbenc…
5
4
2
47
2,370
Jacopo Nardiello retweeted
DeepSeek V4.1 Flash updated to 1017 in OpenCode Go and Free No changes needed on your end
467
674
377
12,735
664,822
Jacopo Nardiello retweeted
US: Tesla Roadster. EU: tethered bottle caps. I made that joke too. But let me be serious for a moment. Made in Europe. Europe built it. Others told the story. The Web was born at CERN. Linux in Helsinki. Python in Amsterdam. ARM, the chip in your phone and your Mac, in Cambridge. DeepMind in London. The first commercial microprocessor was designed by an Italian, Federico Faggin. Every leading-edge chip on Earth is printed with ASML machines from the Netherlands. Mistral trains frontier open models in Paris. Hugging Face: French founders, although it is headquartered in New York. llama.cpp, the engine of local AI, started in Bulgaria. The first mRNA vaccine came from Mainz. Europe doesn’t lack innovation. It lacks a megaphone. 🤷🏻‍♂️ Innovation without narrative becomes someone else’s success. This isn’t Europe vs. anyone. I love innovation wherever it happens. I just want Europeans to feel proud and optimistic about the future again. Did I miss something relevant here?
64
31
8
246
14,049
Jacopo Nardiello retweeted
Un gruppo di coloni israeliani ha aggredito una delegazione diplomatica europea in Cisgiordania ilpost.link/5wzbcUOQzQ
42
100
17
289
20,502
Jacopo Nardiello retweeted
I've been using GLM 5.3-Full on 4 sparks for days now and it's truly trading blows with Opus 4.8. It's a model where even if everything stopped right here, it would by a huge margin pay for itself in productivity gained.
The GLM 5.3 setup on four sparks is legitimately a near frontier powerhouse. @CK2084 build a solid engine, we tuned it and @zb
5
3
1
53
3,302
This guy is singlehandedly exposing the truth in the west bank. The hero we need. You should follow him, singlehandedly exposing the lies of politicians and mainstream media. And I have been the most pro-Israel person on the planet: not anymore.
This is @infinite_jaz rushing to help Heidi under fire from Israeli settlers slings
1
1
4
276
In the video: american journalist kidnapped and beaten by Israeli settlers, or, as per official position from the Israeli gov by “unknown” masked men. Spoiler: these are settlers, known to authorities and actively working with the IDF in the region. Shame!
52