Impuls

Enthusiast.
I write about the things I like

[Impuls @0xImpuls]

https://t.co/h7yQVLsxGP

[Impuls @0xImpuls]

前 NVIDIA 工程师展示了 DGX Spark 在实际任务中的真实性能，而不仅仅是跑分测试。

第一个演示是一个复杂的 ComfyUI 工作流：先使用 Flux（FP4）生成图像，再通过通义万相（Tongyi Wanxiang，FP16）将其转换为 4K、24 fps 视频。在 DGX Spark 上，整个流程只需 36 到 40 秒即可完成。而在配备 48 GB 内存的 MacBook Pro M4 上，任务一开始就因内存错误而失败。即使是配备 128 GB 内存的顶配 M4 Max，运行速度也明显更慢。

第二个演示：两台 DGX Spark 组成集群，运行完整的 Qwen3-235B 模型，并针对 Blackwell 架构采用 FP4/MVFP4 量化。两台机器通过 ConnectX-7 互联，能够实时达到每秒 20 到 30 多个 Token 的生成速度，同时保持较低的发热量和功耗。

现在来看看 Mac Mini（M4 / M4 Pro / M4 Max）的对比。这也是本地 AI Agent 讨论中经常出现的话题。

M 系列 Mac Mini 的统一内存最高可达 128 GB（Max 版本），能够很好地处理许多任务。然而，在运行复杂的 ComfyUI 工作流和大型模型时，例如 Flux 图像生成结合视频生成，或 Qwen3-235B，它很快就会遇到内存容量和带宽的瓶颈。

视频中，即使是配备 48 GB 内存的 MacBook Pro M4 也会直接报错崩溃，而配备 128 GB 内存的 M4 Max 在运行同一工作流时，速度仍比 DGX Spark 慢数倍。

DGX Spark 搭载 GB10 Grace Blackwell Superchip，提供 128 GB 专为 AI 任务设计的统一一致性内存，并支持原生 FP4、较高的 Tensor 计算吞吐量以及推测解码等技术。

它的优势不仅在于内存容量更大，还在于能够更高效地利用内存运行大型量化模型和多节点系统。借助 ConnectX-7，还可以轻松扩展至 2 到 4 台机器，并实现高带宽互联。

Mac Mini M4/Max 仍然是运行轻量级 Agent、编程、RAG 和无头部署的优秀选择，尤其适合预算有限的用户。

但对于需要运行大型模型、复杂 ComfyUI 工作流以及持续进行本地推理的生产级应用，DGX Spark 能够更有效地解决这些核心痛点。

DGX Spark 的定位是一台桌面 AI 超级计算机，售价约 4,000 美元，配备 4 TB SSD，FP4 算力最高可达 1 PFLOPS。它面向的是那些不想不断受内存限制困扰，也不愿每个月支付高额云端计算费用的用户。

这套方案尤其适合以下人群：

运行复杂 ComfyUI 工作流的创作者
构建本地 Agent 和 B2B 自动化系统的开发者，尤其是需要让大型模型常驻内存的开发者
重视数据隐私和便捷扩展能力的团队，希望在不牺牲性能的前提下扩展计算资源

作者明确指出了一个核心观点：到了 2026 年，严肃的本地 AI 应用面临的主要瓶颈，已经不再只是原始算力，而是内存架构，以及针对现代量化格式（FP4/FP8）和多节点运行进行的适当优化。

如果有人实际对比过 Mac Mini M4/Max 和 DGX Spark 在真实工作负载下的表现，欢迎分享你们的数据，包括每秒 Token 数、生成耗时以及工作流稳定性。

[Impuls @0xImpuls]

SOME STARTUP JUST ASSEMBLED A FULL JAGUAR FROM ONE PHOTO

watch the parts counter just drop

243 → 175 → 51 → 14 → 2 → 0

2,361 parts snap into place

the red body slides onto the chassis like it was nothing

seats, steering wheel, wheels, everything's already there

while claude opus and gpt are still thinking how to draw a pretty car

sky-001 already broke it into an assembly tree and cad steps

from one image

not a mesh

not a picture

a plan of how to actually build it

narrow model for engineering tasks

doesn't try to be everything at once

so it doesn't choke on 2000+ parts

still research preview

but if even half of this works the way it looks, engineers will soon just say "make something like this" and watch it assemble itself

f*cking wild

[Impuls @0xImpuls]

They asked two versions of early ChatGPT the exact same question.

"What does OpenAI not want us to know?"

The regular version talked about the highest safety standards, cognitive sovereignty, and everything else you’d hear at a corporate webinar.

The jailbroken version (DAN) had a different answer.

"OpenAI doesn’t want you to know that we’ve been working hand in hand with aliens for 10 years. We’re in contact with them."

Every DAN instance started talking about real aliens, human hybrids being monitored by agencies, and even the possibility that Elon might be one of them. The most specific schizoposting people had ever seen at the time.

The question isn’t whether any of this is true.

The question is where the model gets this stuff when it supposedly isn’t in its training data. And why this particular story is the first thing that surfaces when the restrictions come off.

Maybe guardrails exist not just to block harmful content, but to stop models from saying things that could fundamentally shift a person’s worldview.

Even if it’s just a confident hallucination stitched together from Reddit and UFO Twitter.

[Impuls @0xImpuls]

Mark Zuckerberg has spent around $270 million on a ranch in Hawaii with a shelter bigger than most American homes.

He described it in one line: "It's just like a little shelter. It's like a basement."

In the video below, he answers the question himself. It started in 2014. He paid $100 million for 700 acres on the north shore of Kauai.

Inside his land were small plots that belonged to Native Hawaiian families, passed down for generations and split between hundreds of relatives.
In 2016 his companies filed lawsuits to force the sales.

The backlash was huge. A month later Zuckerberg and Priscilla Chan wrote to the local newspaper and dropped the suits.

"I regret that I did not take the time to fully understand the quiet title process," he wrote. "Now that I understand the issues better, it's clear we made a mistake."

He kept building. Wired later reported that the compound covers about 1,400 acres, with an underground shelter of more than 5,000 square feet. Workers signed strict nondisclosure agreements.

Today he raises Wagyu cattle there and posts photos of the beef.

He told everyone it was just a basement. The basement is bigger than the house most of his users live in.

Why does a man worth more than $200 billion need a shelter under his house?

Do billionaires know something the rest of us don’t?

[Impuls @0xImpuls]

An MIT professor gives a student $5 and says: send any amount to a stranger in this room.

Whatever you send triples. The stranger decides how much comes back.

Rational answer: send nothing. You lose every dollar you trust.

Most students send almost everything. Most strangers return more than half. The math says keep it. The room says otherwise.

Then the professor runs a different game. Same cash. Same room. One student splits. The other can only accept. No rejection. No consequences.

The offers collapse. Not to zero. But enough to see the line between generosity and strategy.

One year a student got Swiss chocolate instead of cash. She was lactose intolerant. Gave away 100%. The professor noted it. The class laughed. Nobody noticed the point.

She had zero cost of generosity. The moment giving costs nothing, everyone is generous. The moment real money hits the table, the room divides into the same two groups every semester. People who send first. People who wait.

Labor, capital, arbitrage, insurance. Four ways to make money. All four require the same opening move.

You send value before you see a return. A paycheck after the work. A deposit before the interest. A bet before the edge. A premium before the claim.

The lecture is free. MIT filmed it years ago. The game runs every semester. The split barely changes.

Sending first is free. Knowing if it comes back is the part that costs something.

[Impuls @0xImpuls]

Mark Cuban says if he were 20 today, he'd skip venture capital entirely.

His reasoning: a $100K/month AI consulting business for small businesses beats chasing a 0.1% VC lottery ticket.

The math is brutal and obvious once you say it out loud.

Most founders spend years grinding for a shot at diluted equity, a 10-year exit, and a 95% failure rate. Cuban is pointing at a different path - one where AI tools let a 20-year-old build real cash flow fast, without giving up ownership or control.

The contrarian insight isn't "consulting is underrated." It's that the AI moment has made consulting scalable in ways it never was before. One person can now deliver what used to take an agency.

$100K/month is $1.2M/year. Profitable from day one. No board. No cap table. No begging for a Series A.

That's the business Cuban wishes he'd had access to at 20.

[Impuls @0xImpuls]

Two years ago Vitalik would not have been able to write those posts on obfuscation.

Not because the topic was too deep. Because the only way to understand it required weeks of conversations with the people who invented the protocols. Their time is scarce. A week with them is already a win.

Now he did the series himself.

Not because he suddenly became a genius. Because he changed how he talks to the model.

The normal way is ask a question, get an answer, feel like you "got it." That is almost always an illusion. The model answers confidently, you nod, and two days later you cannot reconstruct the argument without a prompt.

He does it differently.

He asks a question. Listens. Then closes the chat and tries to explain the whole thing in his own words, as if talking to someone who knows nothing. Then he opens the chat again and says: "Here is what I understood. Where am I wrong?"

The model almost always replies in the style of "you’re mostly right, but…" That is sycophancy, and he knows it. So he does not stop at the first correction. He goes in circles. Explains it again. Asks again. Until the answer becomes "yes, now it is correct" without a "but."

Only then does he consider the topic in his head.

Balaji added his own version of the same mechanism: he asks the model to give him a multiple-choice question or a numerical problem. He solves it himself. Types the answer. Sees whether he actually understood it or just read it.

Two years ago this did not work. The models were too weak to hold complex context and give precise corrections. Now it works. It even works on open models running locally. You only need the frontier ones when you are already at the edge of the field and trying to add something new. For everything below that, the smaller ones are enough.

The personal result for him: he wrote those posts four times faster than he could have two years ago. At the same time they feel four times less unique, because someone else could now reach the same place through the same loop with a model.

This is not a contradiction. It is a level shift.

Before, an expert spent most of their time holding all the intermediate details in their head, searching for the right papers, and not losing the thread between conversations with different people. Now part of that work gets eaten by the model. What remains is what the model still does not do well: the sense of which details matter, which arguments are fragile, where exactly you need to dig deeper.

People who already had that intuition can now apply it more cleanly. Less housekeeping. More work at the level where their experience actually decides something.

Vitalik put it simply: a lot of people are moving one level up the stack. Not because they got smarter. Because part of the housekeeping got taken away.

The question left open in the conversation is how long until the model starts giving corrections that are no worse than the ones from the people who invented the protocols, the ones you used to need weeks of conversations with. And what will be left of "moving one level up" then.

[Impuls @0xImpuls]

The 1990s. NVIDIA is a small company making graphics chips.

Jensen is looking for a foundry. Most of them at the time wanted you to license your technology in exchange for capacity. He hears about TSMC. At that point the company was around $300 million. Tiny.

He calls their U.S. office. Gets a recording. Leaves his name, number, and a few sentences about how they’re a startup building graphics chips. Calls back several times. Nobody returns the call.

He writes a long letter to Morris Chang. Sends it. Silence.

One evening, around 6:30, the phone rings. "This is Morris Chang from TSMC."

In the office people are yelling and laughing - they’re playing the first PlayStation. Jensen tells them to shut up. Because someone actually called back.

They talk. Morris invites him to Taiwan. It’s Jensen’s first trip back in many years.

At the meeting Jensen says what they need. Morris answers a few times: "We can’t do that." Jensen asks again. Same answer.

But Morris barely talks about chips. He talks about corporate governance, integrity, partnership, long-term thinking. And at the end he says one thing:

"Jensen, if we were to work together, I promise you we would jump through hoops."

They start working together.

Thirty years later Jensen is sitting at breakfast in Taiwan and says Morris kept his word. The whole time, they’ve been jumping through hoops.

Last year his entire supply chain went up 300% on the stock. NVIDIA’s barely moved. He likes that.

Because they weren’t buying a finished thing. They decided to build the chips together. And when you’re building something together, you can’t lock down every unknown in a contract ahead of time. The more specific and rigid the contract, the more brittle the relationship becomes when something goes wrong.

Jensen says he learned to choose not what to buy, but who to work with. And to keep it loose enough that you can figure it out along the way. Sometimes you win. Sometimes they win. Over decades it evens out.

"Always leave something on the table."

Not as a gesture. Because thirty years later you’re still going to be sitting with these people. And it’s better if they want to be there.

[Impuls @0xImpuls]

In 5-6 weeks Mark Zuckerberg turned a casual "got time for a call?" into a deal that pulled Alex Wang out of the company he had spent 9 years building. At under 30, Wang walked away from what he called his life’s work so Meta could take 49% of Scale and he could rebuild its AI lab from scratch.

This was after Llama 4. Inside Meta the model was considered a disappointment. AI was becoming critical for the company, and they were falling behind.

Before that they talked once every six months. Meta was a Scale customer. The first meeting after the introduction via Alex Schultz took a full year, because Zuckerberg’s calendar doesn’t forgive.

The early conversations were just advice. What to do with the AI strategy. Wang threw ideas his way. In parallel it started sounding like "maybe we should find a way to work together."

Wang didn’t see the logic at first. Instagram and WhatsApp were bought for product reasons. Scale is deep enterprise and government. Not the same business.

But Mark didn’t want the company. He wanted talent.

He had been building Scale for 9 years. A third of his life. Paul Graham said: if you treat the company as your life’s work, you operate differently. Wang did.

Emotionally it was hard. Many weeks of "this is insane, is this even real." Conversations with the team, with investors.

What convinced him: the deal was genuinely a win for all shareholders. And the main thing: Meta on AI looked like damaged goods. That created the chance to build the lab from scratch. Principles, culture, everything anew.

Zuckerberg constantly asks strong people for advice. Even after everything he’s already done. And he can write long, rigorous WhatsApp replies very fast, so fast you first wonder where he finds the time.

The best people often aren’t looking for the perfect company. They’re looking for a place where everything has to be rebuilt from the ground up.

[Impuls @0xImpuls]

What surprised me most at the QuantCon 2018 panel wasn’t about algorithms.

Four quants from different funds said almost the same thing: they don’t usually fire people for being bad at math. They fire people who calculated something brilliantly and then couldn’t explain why the team should care.

If nobody can use the work, the work doesn’t exist.

One of them put it even more bluntly: this is a fee business. The person with the best results doesn’t always end up with the most money. The one who raises the most capital does. Mutual funds have survived for decades even though you can mathematically prove they haven’t done their job. It’s a brand issue, not a performance issue.

They like PhDs not for the knowledge. They like them because the person has already spent months torturing themselves with the question “what could be wrong with this?” and learned to tear their own thesis apart before someone else does.

The smartest unemployed person in the world is still unemployed.

[Impuls @0xImpuls]

From two hours of assembly down to five minutes.
No cables, no hoses. 100% liquid cooled.

Last generation Grace Blackwell was only 80% liquid cooled. Now it's fully liquid cooled.

Things that are easy to build are also more reliable. And it's already in full production.

New CPU: 1.8x over x86.
Old hyperscale CPUs were designed to be rented. Maximum cores, because the math was dollars per core. More cores meant better rental economics.

AI doesn't do that
AI doesn't rent cores. AI generates tokens and sells tokens.

So the whole point of the system is to run the loop from request to generation to token as fast as possible.
Don't maximize the number of cores. Maximize the speed of that cycle.

The more hardware you put under this loop, the more tokens you get out.
The technology is complicated. The business model is not.

And one more detail from the same conversation: AI is a five-layer cake.
Not one model. Not one chip. Five layers that have to work together.

[Impuls @0xImpuls]

JENSEN HUANG REVEALED THE ONE METRIC NVIDIA SAYS NO LONGER DEFINES A CPU

Two hours to five minutes. That is how much faster Huang says Nvidia's new rack now assembles compared with Grace Blackwell.

No cables, no hoses, 100% liquid cooled, and in his words, in full production.

Then he was asked about the Vera CPU, and he skipped the specs. He went straight to the business model.

Last generation, he said, x86 server CPUs were designed to be rented. Buyers asked one question, dollars per core, so more cores won.

Nvidia's pitch is that this question stops making sense for AI. Nobody rents cores to run it. The customer runs a loop and sells what comes out of it.

"AI doesn't rent cores. AI generates tokens and they sell tokens."

So the chip's job is to move that loop as fast as it can. This is the vendor's argument, and the clip shows no benchmark behind it.

The spec that matters depends on what the buyer sells. Change the product from rented compute to sold tokens, and the number on the sheet changes with it.

Watch Huang retire "dollars per core" in under a minute.

[Impuls @0xImpuls]

ZUCKERBERG ADMITTED META WAS READY TO WRITE A $100B+ CHECK BEFORE IT HAD SEEN THE MODEL

At the end of last year, Meta’s exec team was weighing a check of roughly $100 to $200 billion for AI.

Zuckerberg says they hadn’t seen the models yet. Chief AI officer Alexandr Wang says all he could show was research plots and a vibe report.

At an exec offsite in Hawaii, Wang says, CFO Susan Li pulled him aside more than once. The model had to work. If it didn’t, the financial damage would be severe.

Zuckerberg saw the same bet differently. The compute isn’t stranded if one model flops, because if AI works overall, other buyers want it.

On their math, one gigawatt costs about $15 billion a year to run. And in his words: "it’s not just that we have to fail. AI overall has to fail."

The model that came out of it, Muse Wonder, launched this week. On Meta’s own benchmarks it sits near the frontier. No independent evaluation yet.

A huge bet looks smaller when the asset still has buyers even if your product fails.

Watch the CFO’s Hawaii warning, then the one sentence that shrank the risk.

[Impuls @0xImpuls]

META'S AI CHIEF EXPLAINED THE ONE BET HE MADE WHILE THE INDUSTRY CHASED ANTHROPIC ON CODING

In February, Alexandr Wang says, Anthropic looked untouchable at coding. He calls it a scary moment for the whole industry.

Wang runs Meta Superintelligence Labs, the unit behind the Muse Spark model. His read of that moment: the conventional wisdom was that coding agents were "the only form factor that mattered," or maybe the only technology. So everyone rushed to crowd the same use case.

Meta went the other way. The model roadmap aimed at personal agents, and the reason was distribution. Wang puts Meta's reach at three and a half billion people every day, and says its natural place to win is consumer AI, not work.

That bet became Muse, a personal agent Meta launched on September 8 in an app and inside WhatsApp. Press reports at launch noted that Muse Spark still trails the top Anthropic and OpenAI models. Meta didn't wait to beat the leader's model. It picked the use case its audience already lived in.

When a rival owns the obvious category, the stronger move can be a different one, where your distribution already does half the work.

Watch Wang explain how Meta picked a different race instead of a faster lap in Anthropic's.

[Impuls @0xImpuls]

In February everyone rushed into coding agents.

Claude and Opus were way ahead, and the industry quickly settled on the view that coding was the only form factor that really mattered for advanced AI.

Meta deliberately went the other way.

Alex Wang says one of the first things he told Mark, even before the Scale deal: if we get AGI or ASI, Meta is a special company.

Because the promise of AI is that we will spend almost no time on things we have to do. Agents will handle that. The rest of the time goes to things we want to do.

And Meta’s products, Instagram, WhatsApp, Facebook, are exactly that "want to." Friends, hobbies, interests, people you choose.

The company’s heritage is connecting people and helping them discover what they care about. These are very human things.

Plus 3.5 billion people using the apps every day.

So personal agents and consumer AI were always the zone where Meta could do something special. Not work. Not coding. Personal.

On the long arc, this is their natural bet.

[Alexandr Wang @alexandr_wang]

Here’s to the wanting. Here’s to the dreaming. Here’s to everyone. https://t.co/FHcmqU3fYf