jiggawatts

Is anyone else annoyed by articles that simply repeat themselves instead of providing more detail? You know what I mean! The first paragraph is basically just the title, but with more words. Then the second and third paragraphs rephrase the same. Weirdly, next two paragraphs just does do the same! You expect some detail, some explanation, something in-depth, but no, there’s the same sentence again! And again! /s

In all seriousness, I prefer LLM writing now over some of these “weirdly abbreviated and then padded out” articles because AI output these days is dense and doesn’t repeat itself to get to a target word count.

'bs58' jumped out at me as a random looking package name.

It implements the base58 encoding, which is used primarily by... Bitcoin.

I'd love to see the explanation of why "a TUI file manager" needs a crate closely associated with crypto coins and not much else.

It is precisely this kind of thing that makes Rust a no-go for most enterprises.

"I just want a CLI tool."

"Congratulations, your servers are mining crypto!"

PS: bs58 is maintained by one anonymous person and their package is used in just about every Rust crypto library and hence application. Compromise his account in the same manner as the 'xz tools' attack and you could still billions in crypto!

YAML seems reasonable because it allows the sequence of steps to be treated as data, which then enables GUI visualisation, introspection, etc... without actually having to run anything.

That's critical for a platform like GitHub and for devops pipelines in general.

The failure is that "data" ends up being a "terrible custom DSL" that is bad at everything: Not good at data, not a good DSL, and not even a proper programming language.

The best approaches I have seen to this kind of thing are:

- Pulumi: You get to run custom code, but it outputs data. In other words, your "build automation script" must be a pure function taking data in and returning data out. The resulting data is then treated as the "thing" that the pipeline executes, which means that all decisions (parameters, inputs, etc...) have to be "baked in", before the pipeline starts executing.

- Google CUE (Configure Unify Execute): lets you build up JSON using a strongly typed constraint language. Great for huge, complex configuration.

It doesn't have to be "externally hosted, proprietary AI model"! The argument is against "self-assembled small AI pieces" versus frontier monolithic models.

A) You can always self-host something like Kimi, DeepSeek, or GLM.

B) Just because you use a specific proprietary AI for programming doesn't actually bind you to that provider in any meaningful way. The authored code remains even if you stop paying them!

Of course, if you use AI as an active component in some sort of service, then the EULA, rug-pulls, etc... suddenly start to matter. That's a different story.

See: https://en.wikipedia.org/wiki/Bitter_lesson

Everyone assumes that carefully crafting a specific AI architecture with bits and pieces bolted together based on their human intuition is necessarily superior to simply using a bigger monolithic AI model. It turns out that the opposite is true, and has been demonstrated over and over again.

The bitter lesson is this: You can simply ask a frontier model to do the things you suggested, in a few terse lines of English. Dump a few lines in AGENTS.md and you are good to go.

Your approach is to "fiddle with inadequate tools" for weeks or months until you can finally attain a pale imitation of what the frontier models can do effortlessly.

It's the classic "But I can customise EMACS endlessly, why would I use an actual IDE?" argument all over.

I get it. You don't feel ownership over someone else's AI. You don't feel involved, you don't feel like you have agency.

It's like LEGO or IKEA furniture: study after study has shown that people enjoy things more if they "put it together themselves", even if fundamentally the thing is worse and/or still essentially nothing more than plastic made in a factory.

People keep forgetting that programming is not just about knowing the target programming language, but also an enormous volume of tacit knowledge:

    - Understanding of protocols like HTTP.
    - HTML, JS, CSS, SVG, and everything "web".
    - Understanding of databases, SQL, etc.
    - Abstract code architecture patterns.
    - Understanding the users' requests in English.
    - Responding in English.
    - Command line tool usage (agents/harnesses)
    - Industry-specific knowledge that can be applied.
    - Frameworks, SDKs, applicable libraries.
    - Relevant legal requirements.
    - Etc...

I.e.: If I tell a frontier AI that this project is for a "local council in XYZ location" it can immediately figure out that a scalable, globally distributed architecture is not required. It can also figure out that using local time instead of UTC is not only "fine", but even desired. Or that globalization/localization is not required... or.... required if the council is in some place like Belgium or Canada where multiple languages are officially recognised and supported by the government.

Valuations are not value.

There’s a point where if you ask to be paid a billion dollars for something that is worthless, people just laugh at you.

Until the next transaction you can keep pretending the old inflated valuation is real, but you can raise capital, not without popping the bubble.

At some point people will stop giving Tesla and NVIDIA more money.

Seeing what AI has been doing to maths conjectures recently, I decided to give theoretical physics a go.

I threw some random "crank" ideas into GPT 5.6 Ultra and let it consume most of my weekly Pro subscription quota.

I'm actually pretty impressed with what can be achieved! If your simply ask it to use adversarial agents and include honest self-ratings, it can produce output competitive with a smart but not exceptional PhD. Not imaginative, but the effort that it sinks into even the smallest matters is just amazing to watch.

Did I get anywhere by applying AI to my toy models? Probably not! I ran out of quota before I could tackle anything really interesting.

But... GPT did seem to discover something genuinely new by essentially brute-force effort of combing through hundreds of papers on ArXiV: a new "constant" extracted from a formula over quark masses that seems to evaluate to exactly 3/4 at the one-loop level and then with various corrections at two-loop and higher. Link: https://chatgpt.com/s/t_6a80f2c794388191971350733a6f4378

This is "novel" in the sense that it found a vaguely similar formula in a paper where the authors simply missed the opportunity to extract a simple ratio from quark masses, combined with updated constants the AI found in in later papers that allowed it to guess that this might be exactly 3/4.

This is the super-power of mechanised attention!

We've produced a truly ludicrous amount of scientific and mathematical output, far past the ability of any single human mind to be even vaguely aware of all of it.

Now we can scrape together all of the jigsaw pieces we have made and see what fits together.

I wouldn't be surprised if someone, perhaps even a hobbyist comes up with a viable theory-of-everything in the next year or two simply by feeding in some clever starting point and then turning the handle on the machine until a complete and consistent theory pops out the other side.

Exciting times!

Similarly, the UI must not interrupt me when I’m typing into a text field!

I’ve lost count of the number of modal dialogs I’ve “accepted” because they popped up out of nowhere milliseconds before I pressed enter coincidentally.

> exactly how security and compliance

Not everywhere. I go out of my way to assist teams to achieve a secure outcome with less effort.

Things like: “instead of admin access to the production servers the devs can have fully automated deployment pipelines combined with OpenTelemetry for observability so they don’t have to spend half the day scrolling through gigabytes of logs.”

That’s more secure and and more better.

Nobody had to be told “no”.

Similarly, I replace key store access with secret-less managed identity, etc.

Cash doesn't buy time machines.

They want to build the data centres right now, but the power companies and other energy suppliers are booked solid for 5+ years out.

Money or not, you can't just click your fingers and instantly get regulatory approval, or spin up a facility to build parts that only a handful of places have ever figured out how to mass-manufacture.

Gas turbine blades are especially difficult, but batteries and panels are not far behind. There are a surprisingly small number of shops capable of making power switching gear and transformers too.

Solar works as well as it does only because we’ve spread it out and connected it to a grid. When it is overcast in one spot it is sunny elsewhere.

That “elsewhere” has to be far enough away to get different weather.

This means a grid capable of transferring 100% of the needed power.

So even if you plopped a new DC in the middle of a field of panels you’d still need a ~1 GA grid connection… or a battery array many times the size as you’d otherwise need just to cover night time usage.

The reasons are:

- Running billions of dollars worth of GPUs only 30-50% of the time insanely wasteful. Everyone else is running them at 99% duty cycle or close to it.

- Solar is big and generally has to be built far from where the data centres are. Transmission is already a bottleneck in many states. Adding tens of gigawatts requires non-trivial large scale infrastructure investment by… the utility companies, not SpaceX. Not to mention regulatory hurdles, etc.

- Battery production is already mostly maxed out, so obtaining tens of gigawatt hours of capacity on short notice is a decent percentage of global production.

Gas and other fossil fuel power plants can be deployed adjacent to the data centres for effectively zero transmission loss and no long powerlines that upset the NIMBY types.

They can work 24/7 and are largely available right now, but as others have pointed out the turbine blades are a limiting factor with these too.

Batch mode execution has been in Microsoft SQL Server for a while and just recently gained AVX-512 support.

I’ve done some experiments replacing spatial SQL queries with custom vectorised batch mode code in C# and the speed up was astonishing.

The people dubious about these claims have no idea what their computers are really capable of.

GPT 5.6 has similar language quirks that makes its comments nearly unusable.

I wonder if this is a side effect of MoE models — they can write excellent prose, but not simultaneously with writing code.

I had a similar conversation with a global warming sceptic friend.

It turns out that the "more CO2 = more warmer" model is hilariously simplified, and the real modelling gets into the weeds to put it mildly! There's a NASA database of the super-high-resolution absorption and emission spectra of every isotopic combination of every common molecule and ions, excited states, and more! It turns out that most of the forcing is determined by the behaviour of the upper atmosphere at high latitudes where the air is so thin that exotic excited states can persist for appreciable durations, and are made in large amounts by absorption of UV light. The "glancing angle" of the sunlight near the poles also means that even minor constituents participate in the exchange of IR radiation. Then, then, the simulation has to be run in many thin slices because air is so opaque to IR radiation that it bounces many times on the way up and down, and of course, the isotopic mixes (and excited fractions) are inconsistent between layers.

An insanely complex supercomputer model is required to come up with even a rough estimate of the actual warming.

> reduced to rubble

By whom?

China would be mad to destroy the most valuable building on that island.

I doubt Taiwain would bomb themselves.

I can't see anyone in the US acting like this either.

Both NVIDIA and Apple derive much of their wealth from off-shored manufacturing to… drumroll… China! (PRC+ROC)

If China decides to take over Taiwan and nationalises TMSC, redirecting their top tier silicon to 100% Chinese companies then what are the American firms worth?

"X is less than Y".

"X is literally more than Y".

"You are exaggerating how much X is!"

"It's still a lot more than Y."

PS: This whole thread reminded me of several managers I've worked with who were pathologically unable to estimate... anything, be it driving time or development effort.

They always focused on the "minimal aspect", ignoring everything before and after. Walking to the car park. Standing in line at the machine. Paying at the machine. Getting out of the car park in the car, surprisingly long during busy times. Driving through traffic. Any delays that could -- and regularly do -- occur. Finding parking. Actually parking. Walking from the car park. Etc.

"It's just a 5 minute drive!"

No, it isn't, not door to door.

They might be head hunting other Apple staff and wish to control the narrative to avoid scaring them off.

“If you have the prerequisite hardware, then… know which model you want out of thousands of a variants… and your drivers are up to date, then it is fast!”

Three hours is a lot longer than one minute.

You two have just listed several tools needs to clean up after another tool: not a strong argument for “simpler”.

Neural nets can extrapolate past their training data, and there is no reason to think LLMs don’t inherit this capability.

The extent to which they are able to do this is the more interesting question!

This is the diametric opposite of the rent-vs-buy scenario that this entails.

Local: You need to invest $thousands into GPU and/or very-high-end CPU+Memory hardware.

Vendor: You can use any existing device, even a phone or tablet. A very low-end laptop is fine.

> takes literal minutes to get started

Local: Typical scenario is hours just to download the software, the model weights, and then faffing around with CUDA and matching your GPU drivers.

Vendor: Free-tier available instantly on a web URL. Even local agents have free tiers from multiple vendors. Install is a single command and/or download and "next,next,next,finish" wizard that takes ~1 minute.

> you can just `rm -fr` it and forget the whole thing existed.

I'm still cleaning up multi-GB model weights floating around in hidden subdirectories under my user profile from months ago when I was experimenting with local models!

Meanwhile I simply... stopped using Gemini. That was the entire process: I no longer actively use it. They stopped billing me for my token usage, because it is now zero. That's... it.

You have it totally backwards.

A big part of the metaverse was "buying land" in exclusive locations, a similar concept to NFTs or crypto in general.

Tulips, NFTs, and the Metaverse all revolved around exclusivity and collectibles. Art with no real utilitarian value.

I pay OpenAI $100/mo, twice what I pay for my mobile phone service.

It’s a useful service that I use every day.

I have never even heard of anyone in my social circle expressing even a mild interest in tulips, NFTs, or the metaverse.

This bubble has much more in common with the rail road boom and the 2000 fibre build out bubble.

That's not my meaning. There is a larger game being played. Sometimes one side is the aggressor in one particular local skirmish, sometimes the other.

Over the last few decades, the United States has regularly been the "unprovoked aggressor".

But again, the numbers aren't the point. The individual skirmishes aren't the point.

Neither side is "good" or "evil" like some child's cartoon. These are merely Great Powers poking and prodding each other, looking for advantage.

This is what it feels like to look in at Linux from the outside as a Windows admin used to corporate-enforced consistency.

Sure, Microsoft has stumbled a bit recently, but around the 2010s it was a glorious era of PowerShell and Windows Server releases that were still improving.