很好的一期 youtube.com/watch?v=goiVFy4j…
而数据中心建设受到政治阻力根本原因,简单的说,其实就是AI发展得太快,社会还没有准备好。
这是个新问题,之前从来没有碰到过。不应该拿之前的例子打比方,那些都不准确。
而这些问题也都是有解的,只是要处理得妥善一些。
11
I think that if it is confirmed Jensen Huang violated U.S. export controls by selling advanced chips to China, a relatively sensible approach would be to place the main responsibility on Nvidia as a company, imposing substantial fines, while refraining from pursuing personal criminal charges against him. The decision on whether he should remain as CEO can be left to Nvidia’s board.
48
Replying to @r2d2beeps @cherylnatsu
If it’s really as you say—that other people’s knowledge all comes from public information—then you should just train on public information yourself. What you call “distillation” isn’t merely copying an answer; it’s extracting the detailed reasoning process and steps someone else used to arrive at that answer. That is not public information. It’s high-quality know-how that others obtained by investing substantial effort and computing power. Denying the value of this information is essentially denying all intellectual property.
Gary Shi retweeted
This letter was sent to Nvidia's board today.
A first for @atreidesmgmt
Absolutely share @altcap belief that this letter outlines minimum acceptable standards.
At the end, Jensen Huang looked broken and shattered, his legendary energy depleted. After weeks spent denying the allegations of infidelity and fighting to remain in his job as Nvidia CEO, his surrender late Wednesday night was unconditional.
Huang’s usual take-charge demeanor had evaporated, according to a source, and he’d started using phrases like “whatever” and “I don’t care.” Read more: bloom.bg/4iSdjA6
📷️: Jose Sarmento Matos/Bloomberg
The problem of Apple, is that they make a HUGE deal of EVERYTHING: any change they make must be awesome, must be amazing, must be revolutionary. The fact is, most of these new features are not (still remember the Mac touch bar?).
For touch screen, it's a nice feature to have. On Windows and Chromebooks, I do use them occasionally, like a couple of times a day. And on laptops that don't have them, I do miss them, sometimes. There are cases when they do come handy. I have no reason to hate that feature, if it comes cheap. And you know what, touch screens on laptops have been cheap for quite some years!
I don't know about you, but the part I don't like the most about Apple, is their EGO. It's okay to have a big ego when they offer something really revolutionary. But hey how long since that be?
youtube.com/watch?v=hdMszZ2a…
42
Replying to @StockWe_Com
讲到底,一个公司,在一个时代,能提供怎样的独特价值。在AI之前,Apple的硬件、生态、供应链、隐私政策有足够的护城河,但AI带来的价值一不小心就把Apple多年经营的供应链地位给掀了,这说明什么?本质上,它提供的价值已经远没有AI模型、应用和硬件产业链高。想回到原有地位,除非这点发生改变。
Replying to @fi56622380
I agree with every single word in part III. But the arguments presented here are much softer, frankly speaking.
The fundamental issue is just demand. The price is simply the equilibrium point between the supply and demand for tokens—a dynamic largely independent of whether a model is open-source or closed-source, and now we are largely compute-constrained. If there is a further shortage of compute, the cost of open-source models will rise just as much as that of closed-source ones.
As for the idea that open-source models might capture market share and compress the profit margins of closed-source providers: in the foreseeable future, open-source and closed-source models will serve entirely different markets. For open-ended problems, the challenge posed by open-source models is quite limited. Meanwhile, industry-specific, customized applications are almost invariably built upon open-source foundations—a segment that closed-source providers wouldn't target anyway—so this actually serves to expand total demand. While there will be some impact, it is unlikely to significantly alter the level of investment directed toward frontier model training.
The observation that "capability breakthroughs drive leaps in demand" is profound. Even in the realm of software engineering, there is still immense room for growth. While AI can handle coding tasks, we are still a long way from a "human-out-of-the-loop" scenario—where AI could fully manage the entire software engineering lifecycle, effectively replacing an entire team or company. Achieving this would generate an increase in demand at least an order of magnitude greater than what we see from AI-assisted coding alone. Beyond that, we move into the realm of "what's after software is still software".
Even for bounded tasks, the ability to complete them end-to-end with minimal human intervention is enough to unlock a massive surge in demand for compute.
The potential TAM is far from exhausted. Currently, the market is worth a little over $100 billion, while the annual global income for software engineers is roughly $1.5 trillion (which already represents about 10%). If we look at the total income of all knowledge workers—estimated at $35–40 trillion globally, with about $10 trillion in the U.S. alone—applying a similar ratio suggests a market potential more than ten times larger than what we see today.
So, to return to the beginning: demand is the most fundamental, and it is driven by the growth of capabilities. That is where we should focus on.







