Given your experience with data centers, what do you think of this quote from the article?
> In any case, even the noise worry is overstated: many viral videos of supposed AI data centre noise are really Bitcoin mines, not state-of-the-art hyperscaler facilities.
Are there really major differences between a hyperscaler, neocloud, and crypto mining data center in terms of noise pollution or other negative effects?
No, there really isn't. It's more a question of how the data center is cooled. There are quieter more expensive options, which is why they often aren't used.
Interesting. What ways? Suppose I wanted to call the governor and say “I have an answer to how to regulate these that solves the noise problem” what would we be asking for?
Regulations about 95th percentile decible levels at certain distances, and regulations about max decibel levels at certain distances, for all buildings.
Sure. But what’s the technical way or meet that requirement? Is there a specific method or is it just that they can but they don’t care because there’s no law?
Thanks, that makes sense. I wasn't sure why The Economist would make this distinction, though I can see why certain companies (e.g., 'poorer' non-hyperscalers) might opt for cheap/loud options.
I'm a founder who went through YC (S12), raised VC funding, etc. Now working on other businesses. I also know a few other founders who raised tons of VC funding on their first go, some who have succeeded and some who shut down.
Anecdotally, I feel like the second-time founders see VC as a very specific tool and many prefer not to raise... They'd rather bootstrap and scale more slowly, retaining control, freedom, and flexibility. Granted, a lot of founders also have bad experiences with VCs.
I think your framing VC as a tool is the right one. Not sure if first-time founders are still like this, but ~15 years ago, it seemed like raising VC funding was validation in and of itself and you were perceived as successful due to the raise. The second-time folks know this is neither validation of your product nor is it any sort of guarantee of success... If you need to raise lots of money to capture a speculative market or build something that won't see revenue for a while, VC is great. If you want to build a decent business that grows 20% to 30% per year and compounds over decades, then maybe it's not for you.
Finally, VC is a broad category. Raising money from former founders or from strategic VCs who can legitimately mentor you is very different than getting a check from a fund that makes 100s of investments per year and simply manages you as an arms length portfolio entry.
I recently read a few articles about how the harness is a bigger factor to successful LLM usage and wish they discussed this here.
I use GLM-5.3, Qwen3.8, Claude (all of 'em), GPT Sol/Luna/Terra across direct API calls + local models where I can (128GB Macbook Pro)... The harness and whether the model or underlying system prompts know how to make the best use of iterative LLM calls makes such a big difference...
For example: one-off articles on a news topic (e.g., "Update me on the US-Canada relations") yields very similar results across all models... But "run a web search, write a draft perspective from three points of view, and structure data around it" will make everything but Claude + GPT struggle.
Funny enough, I have a simple test that I have been running on each new model that catches my eye on OpenRouter. It is just a short prompt that asks the model to research yesterdays news and summarize it in a specific format along with a critique of the article or a highlight of any bias: https://gist.github.com/james2doyle/6afb04ea6b18e1a36bc45258...
I have been doing this test for about a year now. I run it on different harnesses and apps as well just to give me an idea of what they differences between them might be. Since I have been running it for a while now, I have a good sense of the correlation between this output and how the model will be on the rest of the things I want it to do.
I would say that almost every model (just tested Mercury 2.5 Preview and IBM Granite 4.2-8b) are pretty much the same on this task. Some are more diligent with how many sources they go out and get, but for the most part they are all very close in quality and will follow the instructions very well.
So saying "everything struggles" with that has just simply not been my experience.
Anecdotally it's a mix. Claude is so good in part because the models are clearly trained to use the harness, and the harness (despite questionable UX) is really best-in-class when it comes to its functionality.
> run a web search, write a draft perspective from three points of view, and structure data around it
Ironically that's not harness-heavy at all, is it? Apart from sterring via system prompt, that's largely relying on the model itself to reason through the task (what to search for, which links to follow) and then synthesize the information and present it in a way that meets the user's request. Seems like a good test of pure LLM capability to me.
I find it hard to believe that if GLM 5.3 struggled with that task in, say, Pi, it would do any better in OpenCode. Unless you're talking about some next-level research stuff to provide strong guidance/steering and context offloading.
I had OMP running GLM-5.3 download multiple CSV datasets for fantasy drafts, classify them according to a strategy I am following, and then design a local site I could use for my draft, that would dynamically account for other picks and give a useful overview of what was available for what I had at the time of pick. 4 inputs from me, two related to how I'd run it on draft day. Worked great!
I usually use pi or minimal harnesses when I am working with Anthropic or OpenAI models. With local llms they seem to work well with maximalist harnesses like omp, Hermes etc.
>But "run a web search, write a draft perspective from three points of view, and structure data around it" will make everything but Claude + GPT struggle.
Did you mean with some specific harness? GLM5.3-Flash handled a very similar test I ran in opencode (with Kagi MCP to let it search) pretty well, and coming from Claude it was incredibly refreshing to not have to decipher its absurd techbro-speak.
> Citadel has sold more than 80 per cent of the portfolio it took on from buying the majority of Situational Awareness’s stock bets last month after the AI-focused hedge fund was almost toppled by the recent tech sell-off.
...
> “Over our nearly thirty-six-year history, we have prided ourselves on being front-footed and proactive during periods of market dislocation,” Griffin wrote. The firm’s flagship fund was up 6 per cent in July, while many of its rivals lost money or were relatively flat.
Author here. Everything is written by me, by hand. What makes you think it's AI slop?
I see you edited your comment. Yes, we use LLMs to monitor news --- otherwise we wouldn't be able to read Japanese news sites or policy documents. However, the articles and analysis are written "manually". I take this very seriously --- the only "AI" is spelling + grammar checking.
For starters, you don't have domain experience in the space. Secondly, you are sourcing from your harness.
> However, the articles and analysis are written "manually". I take this very seriously --- the only "AI" is spelling + grammar checking.
And you do NOT have a background or domain experience working on (eg. In this case) Japanese economic and monetary policy.
Grammar and writing is the least impactful aspect - use AI for that, it doesn't matter. But using it to synthesize and then make judgements isn't valuable.
> But refrain from unfounded allegations of AI slop or incompetence.
What you are doing is BY COLLOQUIAL DEFINTION AI Slop.
I see. Well, feel free to disagree with my insights and ideas; it makes us all better for it.
But refrain from unfounded allegations of AI slop or incompetence. Gatekeeping is even worse.
EDIT: you keep editing your comments. I'm going to stop engaging with you since one can't even have a linear discussion. Good luck with your research and ideas. If you are an expert on Japan, I encourage you to share your views in the general comments here.
lol a website about "Create your own event-powered feeds, newsletters, & risk maps about topics that matter to your business, big or small" that decided to blog about Japan
I'd be more concerned about private credit socializing losses and being bailed out, than bad AI loans leading to private credit problems which then lead to the bailout.
Private credit has a million other problems right now, like bad loans to auto parts dealers[1], iron traders[2], and so many more.
The issue is lack of due process in private credit. THAT is the huge @#$%ing problem.
> I'd be more concerned about private credit socializing losses and being bailed out, than bad AI loans leading to private credit problems which then lead to the bailout.
A distinction without a difference. Insurance is the back door for socializing losses and bailing out private credit. That was the case with AIG, the media paved the way with headlines like "without AIG, the world will have no insurance", and the crooks got paid.
They learned the lesson well, just BS about some vaguely specified "other problems" that take forever to sort out, so we can be bled dry in the meantime - through the back door. Ergo, closing the back door is the first order of business.
I generally agree with you but given your comments, you might enjoy some additional details... Or please challenge me if you think I am wrong.
I've been paying for order-level data feeds on stocks and one thing you'll find is that a lot of the 'sensitive' trades will be anonymized or broken down in different ways to obfuscate who is trading. Citadel would still be able to see there's a surprising level of interest in a certain stock but might not be able to deduce it's one actor. A broker working for SA should know they need to do this, as it helps the broker do better via commissions, etc. too.
My understanding is that Citadel negotiated directly with SA to buy the book, so the final trades were likely taking place outside of the formal market feeds.
Author here. There is absolutely no AI used in any of the writing I do. We use AI to track news and better understand global and economic developments.
> In any case, even the noise worry is overstated: many viral videos of supposed AI data centre noise are really Bitcoin mines, not state-of-the-art hyperscaler facilities.
Are there really major differences between a hyperscaler, neocloud, and crypto mining data center in terms of noise pollution or other negative effects?