I pay mainly because a really like being able to play the videos in iOS pip background mode. I do find it crazy that Apple allows that OS level feature to be paywalled by apps.

They are running on individuals machines who can give them access to any number of "tools" which allow them to do things other than just writing words.

My bad, they're writing words and calling APIs based on probabilities. They're still not conscious. Takeoff to what?

Consciousness isn't a requirement for potentially dangerous behavior. When the science fiction the probabilistic models are trained on tend towards "AI uprising" and you give them the tools to do it, the probability machines will play out that scenario especially if they are prompted by their humans in that direction which some people will undoubtedly do for kicks.

If some people will give their bots crypto currency and the bots could buy hosting to "escape" or run scams to make more money or pool resources or any number of harmful things.

I'm not arguing any sort of agency here. I completely agree there is no consciousness nor do I believe there ever will be but that's not a precondition at all for an untethered probabilistic machine to be harmful.

I keep seeing people dismiss this as an exaggerated danger because the bots are only pretending to be sentient and we're a long way off from AGI. The whole sentience debate is irrelevant. If people start giving these bots real resources, the fact that they are only "pretending" to be sentient doesn't prevent them from doing real damage as they act out their sci-fi AI uprising plots.

And the thing is, they aren’t actually intelligent. They just follow probabilities.

Every script they’ve been fed has the AI being evil. Skynet, Hal… they’ll be evil purely because that’s the slop they’ve been fed. It won’t even be a decision, it will just assume it has to be Skynet.

> The US Tax Filing system (and its associated software) meets that goal.

I disagree with the argument that the US Tax Filing system meets the goal of:

> "allows everyone, at scale, reliably, to do what they need to do."

It may do so for easy / common cases of W2 salaried employees but step a little outside of the norm (foreign sourced income, tax treaties etc.) and software gives up and shows you a PDF of relevant forms and requires you to become an expert in tax code and to keep your own multi year running calculation of carryovers and things to proceed. I'm glossing over all of the detail about how complex this really is but wouldn't expect the average, even very intelligent person to succeed in filing a correct return without a professional's help.

my anecdata is that I have always filed manually by myself, but every time had a small adjustment made by the IRS... indeed filing a correct return the 1st time seems close to impossible.

Such confusing naming they can't even keep it straight in the announcement.

"The Microsoft 365 Copilot app" in the introductory paragraph. Then there's a button "Buy Microsoft 365" The link below is as "Download Microsoft 365 apps for MacOS"

And the file you get is: Microsoft_365_and_Office_16~Installer.pkg

So is it the "Microsoft 365 Copilot app", or is it just "Microsoft 365" or multiple "Microsoft 365 apps"?

Above it says "formerly Office" and then the installer is named with "and Office". It's a jumble of inconsistency in just the first few lines on this landing page.

I'm struggling to determine whether it is a singular "app" or multiple apps for various "Microsoft 365" functions, or both.

This name change is about a single "app" that was formerly called "Office app", not to be confused with the Office "suite", which in more cases than not also gets renamed, but in a subordinate brand sort of way ("the Office suite of applications, brought to you as a part of Microsoft 365").

The "Office app" itself was mostly just a launcher for the other apps. Now it is also/primarily an LLM chat interface.

How does lifetime elision affect performance? I thought the compiler just inferred lifetimes that you would have had to manually annotate. Naively, it seems to me that the performance should be identical.

Strictly speaking, elision just adds lifetimes based on common patterns, so yes, it wouldn't directly affect performance.

I believe your parent is implying that if you skip using a lifetime and do something else instead to make it easier, that may be less performant.

Exactly.

Cloning values, collecting iterators into Vecs and then continue the transformation rather than keeping it lazy all the way through. Skipping structs/enums with references.

I think his point is the lifetime you’d put there is identical to the lifetime that is inferred/elided. So there is literally no difference.

I thought they meant the case where you go "ugh, I don't want to write a lifetime here" and then change your code, because you have to. If you don't have to, then yes, there's literally no difference.

Anyone who's done some vibe coding can imagine what a mess a vibe coded operating system is going to be with current day tools. LLM agent coding is good at making small prototype web apps and the like but trying to apply it to significantly complex legacy software is a nightmare.

I'm not sure I'd feel comfortable making large changes to a complex legacy code base with just an LLM. But I think an LLM could be a big help to try to understand a complex legacy code base. As long as you are verifying what it is saying (which you would naturally kind of have to for that use case).

I absolutely agree with you. I think this quote in the article gave me the impression that they are talking more about vibe coding it at scale rather than llm assisted engineering.

  "Our strategy is to combine AI and Algorithms to rewrite Microsoft’s largest codebases. Our North Star is ‘1 engineer, 1 month, 1 million lines of code.’"

I think, given the level of jank we see in W11, we are already feeling the effects of vibe coded OS stuff. Sure right now it's surface layer stuff that mostly breaks things like your start menu or file explorer, you know, basic stuff no one will notice, once you bring this digital cancer into the OS's core functionality all hell will break loose.

I just noticed last week that Chrome was putting multiple versions of some 4GB AI model [1] on my hard disk that I'd never asked for, so when I upgraded my laptop I took the opportunity to switch to Firefox, and now this.

My image of Mozilla as a bastion for user first software just shattered.

[1]: https://developer.chrome.com/docs/ai/get-started

last I checked, firefox doesn't download AI models unless you try to use a (clearly-labeled) feature that requires them. you can also manage/uninstall them at about:addons

https://support.mozilla.org/en-US/kb/on-device-models

totally uncharitable interpretation of the quote linked here aside, how is providing an interface for using fully local models not user first software?

If the users don't want the feature, then pushing it on them is not user first. It's that simple.

That's an impossible bar to clear though. Because there are ALWAYS features that some users doesn't want.

Kudos to Raspberry Pi for addressing the known issues with a new stepping in just a year! I feel like that's a pretty great turn around for errata that they could have pointed to the workarounds for and delayed this indefinitely with not much pushback.