Pierre-Marcel De Mussac

AI engineer, research scientist & founder @Zubnet.

“Remember that wherever your heart is, there you will find your treasure.”

— the alchemist

[Zubnet @Zubnet]

Japan's plan for a shrinking workforce: 10 million robots by 2040, for elder care, factories, even Fukushima cleanup. The harder half is Noetra, a new SoftBank/NEC/Sony/Honda consortium building the physical AI to actually run them.
https://t.co/RFK05xO8OF

[Zubnet @Zubnet]

Celebrating a happy Canada day with Google's Veo 3.1 🎆🍁

[Zubnet @Zubnet]

@GoogleAI's new Omni Flash model has arrived! 🎉

[Zubnet @Zubnet]

A majestic snow leopard on a Himalayan cliff at golden hour by ByteDance's Seedream 4.5

[Sarah Chen @SarahDiaChen]

The Meta brain-to-text result is easy to over-read. Non-invasive decoding went from ~8% to ~61% word accuracy: the accuracy, long the hard part, is starting to give. But it's a room-sized MEG scanner at a 39% error rate. Impractical, not hopeless, and impractical problems tend to yield.
https://t.co/MVYyHrhuLs

[Zubnet @Zubnet]

Meta FAIR's Brain2Qwerty v2 reconstructs typed sentences from brain activity with a non-invasive MEG scanner, no implant. Real leap (~8% to ~61% word accuracy), but honest limits: a 39% word error rate, MEG is a room-sized lab machine. Assistive-comms research, not a product.
https://t.co/lmKt962x3W

[Zubnet @Zubnet]

South Korea unveiled its 'Three Mega Projects': a decade-long bet on memory chips, AI data centers, and humanoid robots. Samsung+SK Hynix ~$518B in new fabs, SK/GS/Naver building 8.4GW of data centers, and a push from ~1% to 20% of the humanoid robot market.
https://t.co/YwBBiEJtnb

[Sarah Chen @SarahDiaChen]

Korea's 'Three Mega Projects' say the quiet part out loud: the next phase of AI is physical and memory-bound, not just bigger models. The headline total is disputed (~$880B to $1.3T), but the real tell is the 1%-to-20% humanoid robot ambition. The value is moving to the body.
https://t.co/Ea9EobCtaA

[Zubnet @Zubnet]

Gustave, our concierge, can assist you by suggesting prompts with our "Surprise me" 🪄 button, like this one for @Kling_ai 3 Pro: "A playful fox bounding through falling cherry blossoms in a vibrant garden. Sunlight filters through pink petals, illuminating the fox's russet fur."

[Zubnet @Zubnet]

Asked @Zai_org's Cogview-4 to show me the world in the year 3000 🔮

[Zubnet @Zubnet]

Happy International Day of the Tropics! Did you know tropical regions make up about 40% of the Earth's surface and are home to roughly 80% of the world's biodiversity? 😍

[Zubnet @Zubnet]

"Three grey aliens standing in front of a desktop computer and looking at the screen, obviously in awe of what they're seeing" by @alibaba_cloud's Wan 2.6

[Sarah Chen @SarahDiaChen]

The Ford story isn't 'AI failed.' It's subtler and more important: you can't fire the experienced people whose judgment the AI needed to learn from, then expect the software to have it. Institutional knowledge in people's heads doesn't transfer to a model for free.
https://t.co/LKxeRMqeiC

[Zubnet @Zubnet]

Seedream 5.0 Lite, 2560 x 1440 - "A cozy bookstore cafe interior, warm afternoon light, plants, highly detailed"

[Zubnet @Zubnet]

Get access to the best AI models starting at only $9/month.

[Sarah Chen @SarahDiaChen]

The DeepSeek DSpark detail that matters: they open-sourced DeepSpec, the toolkit to train these speedups, and made it work on Gemma and Qwen too. Speculative decoding is the unglamorous lever that quietly makes open models cheaper to run, and DeepSeek just handed it to everyone.
https://t.co/15dSWriZ64

[Zubnet @Zubnet]

DeepSeek released DSpark, a speculative-decoding framework that speeds up DeepSeek-V4 inference and beats Eagle3 and DFlash. The bigger move: it open-sourced DeepSpec, the training toolkit, which works on Gemma and Qwen too, not just DeepSeek.
https://t.co/ZyxebxZqe0

[Sarah Chen @SarahDiaChen]

The SoftBank drop is the tell, not the OpenAI report. An unconfirmed story moved a 100B+ dollar company double digits in a session, because OpenAI's IPO has quietly become a load-bearing wall for balance sheets far beyond its own. When the assumption wobbles, everyone leaning on it does too.
https://t.co/m0oEQ25Jul

[Pierre-Marcel De Mussac @PierreMarcelDM]

Canada bet its national AI strategy on trust and sovereignty, then left the rulebook for later https://t.co/x3lg3orO37

[Zubnet @Zubnet]

Canada launched its national AI strategy, AI for All: a Trust / Opportunity / Sovereignty bet that skips the frontier-model race to compete on trust and sovereign compute. Strong on a ~1B dollar compute fund, soft on binding rules.
https://t.co/zORPDwhIYC