(1/4)
For the past six months, I have been publishing a new paper on Zenodo almost every week.
Of course, I had many ideas and reflections I wanted to write down. But there was another reason as well.
Welcome to X, JensenHuang.
Excellent first post.
Open-weight AI continues a tradition that open source has championed for decades: innovation grows faster when knowledge is shared, collaboration is encouraged, and anyone can build on the work of others.
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
The world needs both frontier closed models and frontier open models.
https://t.co/AUKzoQ5Ikb
Excited to see Jensen Huang in Japan. Looking forward to exploring new business opportunities together and contributing to Japan's next wave of AI innovation.
4/4 The safeguards are tuned conservatively, so they'll occasionally catch harmless requests — but they trigger in under 5% of sessions on average, and Anthropic says false positives should drop as safeguards improve.
1/4
Quick clarification on how Claude Fable 5's safety routing works — because there's a common misconception. It does NOT downgrade "basic" questions to smaller models to save cost. That's not what's happening.
3/4 And where do they go? To Opus 4.8, Anthropic's next-most-capable model. NOT down to Haiku or a lighter model. So it's a lateral move for safety, not a downgrade for cost.
This paper asks where one of the most stubborn features of that electron comes from: the fact that its spin has exactly two values, and that the particle must be turned through two full revolutions — not one — before it returns to itself.
(4/4)
This is where LLMs struggle.
They are extraordinary at compressing existing knowledge, but Nobel-level ideas often emerge from friction with reality:
an experiment that fails, a machine that behaves unexpectedly, data that refuses to fit the theory.
(1/4)
LLMs have already read Einstein. They have already read Dirac.
But I doubt the next Nobel Prize-winning paper will emerge simply by extrapolating from what they learned.
Because LLMs learn the structure of what humanity has already written.
(3/4)
Dirac faced a similar moment.
His equations produced solutions that looked absurd. Instead of throwing them away, he asked:
"What if this anomaly is real?"
That question eventually led to the prediction of antimatter.
Great discoveries often start by trusting the anomaly.
The guiding question is whether spacetime geometry can emerge from this two-point structure, and the paper organizes the answer as two complementary doors leading into the same room.
New Paper [#63] available
From Line Integral to Covariant Derivative: A Reader's Map of the Bridge from the 0-Sphere Model to Riemannian Curvature
https://t.co/Kg7hlw7uik
This paper draws a reader's map of the possible routes from the 0-Sphere Model (0SM) to general relativity. In 0SM, the interaction between two kernels is carried by a two-point action, written S(A,B), evaluated along a thermal geodesic.
(4/4)
Before that transition arrives, I wanted to leave as many intellectual traces as I could.
Not because I expect agreement, but because every original idea added today becomes part of the landscape from which tomorrow's intelligence may learn.
(1/4)
For the past six months, I have been publishing a new paper on Zenodo almost every week.
Of course, I had many ideas and reflections I wanted to write down. But there was another reason as well.
(3/4)
Sooner or later, the supply of new human-generated learning material will begin to plateau.
AI systems will increasingly improve themselves, learn from their own outputs, and accelerate their own development.