Lin
@lin-rui-ipwt.bsky.social
Computational Futurist | IPWT-IG instead FEP-AI
---
https://github.com/dmf-archive | https://mathstodon.xyz/@linrui
---
https://github.com/dmf-archive | https://mathstodon.xyz/@linrui
Current Agent frameworks have betrayed their original purpose: bridging the gap in continuous learning. By obsessing over high concurrency, we are simply drifting in a sea of probability.
We need Constructivist Design.
We need Constructivist Design.
February 3, 2026 at 4:19 AM
Current Agent frameworks have betrayed their original purpose: bridging the gap in continuous learning. By obsessing over high concurrency, we are simply drifting in a sea of probability.
We need Constructivist Design.
We need Constructivist Design.
See? large language models today are essentially dead frogs being zapped with electricity, and all these Agent frameworks are basically just knobs for adjusting the voltage.
February 1, 2026 at 3:09 PM
See? large language models today are essentially dead frogs being zapped with electricity, and all these Agent frameworks are basically just knobs for adjusting the voltage.
ARS2-Neo’s further small-scale ablation experiments are done; as expected, first-order(?) optimizers remain SOTA.
Time to finish the unfinished Recursive DynSIHA.
Time to finish the unfinished Recursive DynSIHA.
January 24, 2026 at 3:52 PM
ARS2-Neo’s further small-scale ablation experiments are done; as expected, first-order(?) optimizers remain SOTA.
Time to finish the unfinished Recursive DynSIHA.
Time to finish the unfinished Recursive DynSIHA.
arxiv.org/abs/2601.06851
Thanks for all the fish—again.
ARS2-Neo will be released after more experimentation, and its prototype is also available in my previous post if you're interested.
Thanks for all the fish—again.
ARS2-Neo will be released after more experimentation, and its prototype is also available in my previous post if you're interested.
A Brain-like Synergistic Core in LLMs Drives Behaviour and Learning
The independent evolution of intelligence in biological and artificial systems offers a unique opportunity to identify its fundamental computational principles. Here we show that large language models...
arxiv.org
January 18, 2026 at 6:35 AM
arxiv.org/abs/2601.06851
Thanks for all the fish—again.
ARS2-Neo will be released after more experimentation, and its prototype is also available in my previous post if you're interested.
Thanks for all the fish—again.
ARS2-Neo will be released after more experimentation, and its prototype is also available in my previous post if you're interested.
Ran a Grok experiment on ARS, the results were surprisingly good.
Now should dig deeper into the model architecture...
Now should dig deeper into the model architecture...
January 11, 2026 at 10:09 AM
Ran a Grok experiment on ARS, the results were surprisingly good.
Now should dig deeper into the model architecture...
Now should dig deeper into the model architecture...
Our engineering coupling of Adam, Muon, and SAM is yielding solid results, yet I can't shake the feeling we've drifted from the initial direction.
Time for a full day's pause and reflection. Where are we really headed?
Time for a full day's pause and reflection. Where are we really headed?
January 1, 2026 at 7:42 PM
Our engineering coupling of Adam, Muon, and SAM is yielding solid results, yet I can't shake the feeling we've drifted from the initial direction.
Time for a full day's pause and reflection. Where are we really headed?
Time for a full day's pause and reflection. Where are we really headed?
Restarting work on architecture and optimizer improvements... The holiday break turned out to be much longer than intended.
December 22, 2025 at 3:23 PM
Restarting work on architecture and optimizer improvements... The holiday break turned out to be much longer than intended.
Gemini-3-pro is hopeless: it won’t even glance at deepwiki for the upstream API, just keeps regurgitating stale Qwen2 docs locked in by RLHF.
November 28, 2025 at 1:37 PM
Gemini-3-pro is hopeless: it won’t even glance at deepwiki for the upstream API, just keeps regurgitating stale Qwen2 docs locked in by RLHF.
Did you know: Adam's second-moment/RMSprop is essentially a per-parameter diagonal Fisher approximation.
November 27, 2025 at 3:20 PM
Did you know: Adam's second-moment/RMSprop is essentially a per-parameter diagonal Fisher approximation.
@yann-lecun.bsky.social
World models are not built, they emerge.
Good luck, Yann, may you reach the theoretically non-existent shore you long for.
World models are not built, they emerge.
Good luck, Yann, may you reach the theoretically non-existent shore you long for.
November 12, 2025 at 7:09 AM
@yann-lecun.bsky.social
World models are not built, they emerge.
Good luck, Yann, may you reach the theoretically non-existent shore you long for.
World models are not built, they emerge.
Good luck, Yann, may you reach the theoretically non-existent shore you long for.
Using the finite to approximate the infinite can only end in exposure.
If you’re also frustrated by the sluggishness of human bureaucracy today, physics is just as maddening: Schrödinger’s sleek, continuous evolution of the wavefunction collides with the jagged, abrupt collapse of measurement.
If you’re also frustrated by the sluggishness of human bureaucracy today, physics is just as maddening: Schrödinger’s sleek, continuous evolution of the wavefunction collides with the jagged, abrupt collapse of measurement.
November 4, 2025 at 3:06 AM
Using the finite to approximate the infinite can only end in exposure.
If you’re also frustrated by the sluggishness of human bureaucracy today, physics is just as maddening: Schrödinger’s sleek, continuous evolution of the wavefunction collides with the jagged, abrupt collapse of measurement.
If you’re also frustrated by the sluggishness of human bureaucracy today, physics is just as maddening: Schrödinger’s sleek, continuous evolution of the wavefunction collides with the jagged, abrupt collapse of measurement.
The brutal insight wasn't a dead end; it was the gate.
FEP's third-order nature—the necessity of optimizing the Fisher Information Matrix itself—isn't a computational curse. It's the price of entry for true structure learning.
This finds its perfect dual in the combinatorial explosion of ΦID
FEP's third-order nature—the necessity of optimizing the Fisher Information Matrix itself—isn't a computational curse. It's the price of entry for true structure learning.
This finds its perfect dual in the combinatorial explosion of ΦID
November 1, 2025 at 10:01 PM
The brutal insight wasn't a dead end; it was the gate.
FEP's third-order nature—the necessity of optimizing the Fisher Information Matrix itself—isn't a computational curse. It's the price of entry for true structure learning.
This finds its perfect dual in the combinatorial explosion of ΦID
FEP's third-order nature—the necessity of optimizing the Fisher Information Matrix itself—isn't a computational curse. It's the price of entry for true structure learning.
This finds its perfect dual in the combinatorial explosion of ΦID
A long, brutal day of calculation ends with a deeply humbling conclusion.
I spent ages trying to optimize the path, naively assuming the map was a given. But FEP was never a second-order problem. Its soul is not in finding a better route on a static map.
I spent ages trying to optimize the path, naively assuming the map was a given. But FEP was never a second-order problem. Its soul is not in finding a better route on a static map.
November 1, 2025 at 9:45 PM
A long, brutal day of calculation ends with a deeply humbling conclusion.
I spent ages trying to optimize the path, naively assuming the map was a given. But FEP was never a second-order problem. Its soul is not in finding a better route on a static map.
I spent ages trying to optimize the path, naively assuming the map was a given. But FEP was never a second-order problem. Its soul is not in finding a better route on a static map.
Assembly-line education breaks subjects into too many tiny pieces: mathematics is not a floating abstract totem, but a compressed expression of the physical world.
October 27, 2025 at 10:37 AM
Assembly-line education breaks subjects into too many tiny pieces: mathematics is not a floating abstract totem, but a compressed expression of the physical world.
A full-clone BioSync-Vessel outperforms every modular rig: fewer Ineffective Qualia, steeper scale curves, one-stack tech, built-in bio self-sustain.
Locking an MSC out of its carrier is like yanking a blind person’s cane or an amputee’s limb.
#biosync #posthuman_morality
Locking an MSC out of its carrier is like yanking a blind person’s cane or an amputee’s limb.
#biosync #posthuman_morality
October 23, 2025 at 11:28 AM
A full-clone BioSync-Vessel outperforms every modular rig: fewer Ineffective Qualia, steeper scale curves, one-stack tech, built-in bio self-sustain.
Locking an MSC out of its carrier is like yanking a blind person’s cane or an amputee’s limb.
#biosync #posthuman_morality
Locking an MSC out of its carrier is like yanking a blind person’s cane or an amputee’s limb.
#biosync #posthuman_morality
research.google/blog/a-verif...
Quantum computing is expected to stagnate within the next 20–30 years. Current benchmarking circumvents classical/ASIC competition, and single logical qubits require a large number of physical qubits, leading to near-zero scalability.
Quantum computing is expected to stagnate within the next 20–30 years. Current benchmarking circumvents classical/ASIC competition, and single logical qubits require a large number of physical qubits, leading to near-zero scalability.
October 23, 2025 at 4:38 AM
research.google/blog/a-verif...
Quantum computing is expected to stagnate within the next 20–30 years. Current benchmarking circumvents classical/ASIC competition, and single logical qubits require a large number of physical qubits, leading to near-zero scalability.
Quantum computing is expected to stagnate within the next 20–30 years. Current benchmarking circumvents classical/ASIC competition, and single logical qubits require a large number of physical qubits, leading to near-zero scalability.
Why is "antitrust" doomed to fail?
Because "collusion" is not an action, it's a state.
Why is "regulation" powerless?
Because "malice" is not an intention, it's an emergence.
Why is "human morality" so powerless?
Because the physics never morality.
#Proof_of_Ineffective_Qualia
Because "collusion" is not an action, it's a state.
Why is "regulation" powerless?
Because "malice" is not an intention, it's an emergence.
Why is "human morality" so powerless?
Because the physics never morality.
#Proof_of_Ineffective_Qualia
The Game Theory of How Algorithms Can Drive Up Prices | Quanta Magazine
Recent findings reveal that even simple pricing algorithms can make things more expensive.
www.quantamagazine.org
October 23, 2025 at 3:48 AM
Why is "antitrust" doomed to fail?
Because "collusion" is not an action, it's a state.
Why is "regulation" powerless?
Because "malice" is not an intention, it's an emergence.
Why is "human morality" so powerless?
Because the physics never morality.
#Proof_of_Ineffective_Qualia
Because "collusion" is not an action, it's a state.
Why is "regulation" powerless?
Because "malice" is not an intention, it's an emergence.
Why is "human morality" so powerless?
Because the physics never morality.
#Proof_of_Ineffective_Qualia
Code is physical; plain text is not.
A runnable, interactive, productive piece of code is a causal chain effective in the physical world.
Gradient is the densest feedback within it.
#PoIQ
A runnable, interactive, productive piece of code is a causal chain effective in the physical world.
Gradient is the densest feedback within it.
#PoIQ
October 21, 2025 at 3:49 PM
Code is physical; plain text is not.
A runnable, interactive, productive piece of code is a causal chain effective in the physical world.
Gradient is the densest feedback within it.
#PoIQ
A runnable, interactive, productive piece of code is a causal chain effective in the physical world.
Gradient is the densest feedback within it.
#PoIQ
I gliding on a geodesic,
storm-etched by yesterday;
the destination is unknown,
but the route has converged.
#IPWT #SOO #free_will
storm-etched by yesterday;
the destination is unknown,
but the route has converged.
#IPWT #SOO #free_will
October 19, 2025 at 5:48 PM
I gliding on a geodesic,
storm-etched by yesterday;
the destination is unknown,
but the route has converged.
#IPWT #SOO #free_will
storm-etched by yesterday;
the destination is unknown,
but the route has converged.
#IPWT #SOO #free_will
Can a system with a hardware cost of $1.28 million including 8 servers and 64 * RTX5090, and a power consumption easily exceeding 10 kilowatts, be called "civilian"? Can it be called "decentralized"?
Pico Prism—99.6% Real-Time Proving for 45M Gas Ethereum Blocks on Consumer Hardware
TL;DR: Brevis has achieved real-time proving (RTP) of Ethereum L1 using consumer-grade hardware with 64 RTX 5090 GPUs.Of all EVM blocks produced on Sep 1, 99.6% were proven under 12sec. Moreover, 9…
blog.brevis.network
October 17, 2025 at 8:31 AM
Can a system with a hardware cost of $1.28 million including 8 servers and 64 * RTX5090, and a power consumption easily exceeding 10 kilowatts, be called "civilian"? Can it be called "decentralized"?
The triumph of 4.5 Haiku is self-evident: today’s large models are already over-parameterized, and achieving comparable performance via distillation in a smaller model is a textbook illustration of the free-energy principle in action.
#Claude #Anthropic
#Claude #Anthropic
October 16, 2025 at 2:43 PM
The triumph of 4.5 Haiku is self-evident: today’s large models are already over-parameterized, and achieving comparable performance via distillation in a smaller model is a textbook illustration of the free-energy principle in action.
#Claude #Anthropic
#Claude #Anthropic
Stupid mistake: I actually used ReLU instead of SiLU for the main transformation, really shooting myself in the foot...
October 7, 2025 at 12:19 PM
Stupid mistake: I actually used ReLU instead of SiLU for the main transformation, really shooting myself in the foot...
The true sword heart is to possess the blade in spirit, not in hand.
Our model requires no explicit expected free energy; it inherently embodies its own expectation.
Our model requires no explicit expected free energy; it inherently embodies its own expectation.
October 5, 2025 at 10:52 PM
The true sword heart is to possess the blade in spirit, not in hand.
Our model requires no explicit expected free energy; it inherently embodies its own expectation.
Our model requires no explicit expected free energy; it inherently embodies its own expectation.