rickawsb

Science is essentially “reality compression”: a good theory is, at its core, using less information to describe more of reality.
Kepler → Newton is the classic example:
Kepler first summarized the laws of planetary motion from a large amount of astronomical data; Newton then went a step further, explaining that these laws could actually be unified through the laws of motion and universal gravitation.
And AI is continuously accelerating the entire chain of scientific discovery.

Lately, I often find myself thinking of a chilling line:
The people of Chang’an, living in the prosperous Tang dynasty in the early Tianbao era, could never have imagined that most of them would die in the An Lushan Rebellion that followed.

Watching all the arguments about AI safety these past few days
Whether for or against
I only want to say
When the avalanche comes, no snowflake is innocent

Can an open-source model that runs locally reach OpenAI Astra’s current level in two months?
The Information, citing people familiar with the matter, reports that Astra uses limited recurrent depth (recurrent depth / looped transformer)—the same set of layers runs through the internal state several more times before generating the next token; the report describes this as a technique that improves performance while reducing visible chains of thought.
If this is true, it would be great news for open-source models, because recurrent depth provides a greater performance boost for smaller models th

How powerful is astra, exactly?
I have a complex semi-automated customer support workflow. When running on gpt 5.6 sol high, it requires at least half an hour of human involvement per project.
astra shortened this time to 5 minutes.
At the same time, the task success rate has increased significantly (I haven't had time to calculate it yet, but visually it appears to have at least doubled).
But I have to say, it really burns through tokens fast.

Leading robotics company Wandercraft announced on September 4 that its Calvin‑40 industrial humanoid robot has secured 12 large enterprise customers across automotive, logistics, retail, healthcare, and defense.
Interestingly, the company did not try to make the robot perform every human task from the outset, instead initially focusing on:
Tires, boxes, panels, carts, packages, and other material handling.
It is then gradually expanding into more sophisticated tasks such as picking and sorting.
Wandercraft’s approach is increasingly aligned with recent signals from Hyundai, Geek+, agricultural

The most shocking discovery today was that nonfarm payrolls beat expectations, and even on-chain memes fell 😂

Buffett likened capital markets to “a church with a casino attached”
For crypto, this is a huge opportunity to upgrade from a casino to a church

The PANW boss is bearish on neo cloud, believing computing power supply may be excessive
This makes me even more bearish on PANW

In such a time of upheaval, investors should be both extremely optimistic and extremely conservative
Extremely optimistic, to ensure they do not miss the dividends of the times
Extremely conservative, to ensure they can survive any black swan event

The era when AI can (almost) answer everything
Questions are more important than answers
A good question is worth its weight in gold

AI will drive the industrialization of uninhabited areas and the unmanned transformation of industrial areas at a speed visible to the naked eye.

In a sense, consumers buying some older-model consumer electronics and computers are indirectly purchasing the “low-cost memory” that manufacturers stockpiled in the past.
But this window will not last forever.
As old inventory is gradually depleted, high-priced DRAM and NAND will enter new production batches. Manufacturers will ultimately have to weigh several options: raise prices, reduce memory and SSD capacity, or simply cut the supply of low-end products.

If linearly extrapolated based on the figures from the past year,
Google’s AI token activity has roughly doubled every 4.4 months over the past year;
Global H100-equivalent computing capacity has roughly doubled every 6.8 months;
AI data center full-load power capacity has roughly doubled every 9.9 months.
One year from now, the tightness of compute supply and demand will be approximately 1.95 times the current level
The above does not yet account for new use cases continually unlocked as AI capabilities improve...

This AI-driven wave of rapid semiconductor advancement will significantly accelerate the development of quantum computing because:
→ Improved device consistency and wafer yields
→ More cryogenically qualified devices entering packaging
→ cryo-CMOS, multiplexing, and vertical interconnects removing I/O bottlenecks
→ Higher post-packaging yields and simultaneous control and readout capabilities
These are all core capabilities required for quantum computing

The ultimate form of robots is indeed humanoid
But constrained by technologies such as dexterous hands and cost considerations, while benefiting from the rapid development of models and the maturity of 3D printing and personalized customization, robots are more likely over the next two or three years to first become semi-general-purpose machines that can rapidly change form to adapt to different specialized tasks.

Payments giant Stripe acquired OpenRouter for a staggering $7.5 billion, even though it was valued at only $1.3 billion just 90 days ago.
One of the reasons for this being the largest acquisition in Stripe's history is that Stripe believes the “singularity” officially arrived on January 1, 2026.
OpenRouter provides developers with a unified interface for accessing more than 400 AI models. Since August 2023, the platform's weekly Token transaction volume has grown by an astonishing 24,000 times, and this exponential growth in token consumption is regarded as the new “Moore's Law” of the intelli