Bugging me to try Apple Intelligence

[heretiq]

I feel a little out of place, chucking in my 2c worth into this conversation. I'm just a draughtsman (and am in the process of becoming an energy efficiency assessor), and a nerd, but surely at some point the AI bubble will burst, like most tech hype cycles, and LLMs and diffusion models will no longer have their next big thing hype, and become just another set of technologies for use in making better tools.

As an assistive technology, spoken chatbots have great potential, but as an everyday interface, I don't want to be walking around like I'm in Star Trek, having to talk to a small screen to make it do what I want! Standing in the queue in the supermarket: Computer! Read that message from my boss!

Well said. I think we can largely trust professional app designers and developers to use conversational UIs where it makes sense. But definitely expect to see some cases of overuse to leverage AI hype factor. Getting the hybrid UX worked out is not a trivial exercise but I’m optimistic that progress will follow similar maturity curves as precursor UI/UX paradigms and become largely invisible at some point in this decade.

 

[heretiq]

There are whole areas of just finance which have bigger than Apple sized chunks of money floating around with absolutely no use case or interest in LLMs at all. In fact they see them as a risk due to the unconstrained risks.

Finance is not monolithic. In 3Q of 2024 we finished an underwriting department process improvement consulting project with a leading US commercial insurance provider.

The only way to achieve their required productivity improvements was by improving the effectiveness of their underwriters because they had largely exhausted underwriting process automation opportunities. The target opportunity space encompassed manual fact-checking, application data validation, web-based research and myriad due diligence activities conducted by underwriting associates and underwriters.

We built a Proof of Concept that demonstrated that these functions could be aided by an LLM-based Agent application that performed these functions, summarized findings and generated preliminary underwriting recommendations based on company guidelines complete with decision support citing the data and applicable business rules. The preliminary recommendations were provided to underwriters for review and final decision making and produced the desired productivity gains. That solution is now in production development — for a function that is literally charged with financial risk management — so risk may be viewed differently in the are of finance you participate in.

This is just one use case that represents (a) LLM-based automation that is being implemented in insurance companies across the globe, and (b) a fundamental repeatable pattern that is applicable to many similar use cases across the finance industry.

So, again, while this doesn’t provide definitive evidence of unfailing success, it does support the contention that pronouncements of certain failure are premature.

 

[cjsuk]

Finance is not monolithic. In 3Q of 2024 we finished an underwriting department process improvement consulting project with a leading US commercial insurance provider.

The only way to achieve their required productivity improvements was by improving the effectiveness of their underwriters because they had largely exhausted underwriting process automation opportunities. The target opportunity space encompassed manual fact-checking, application data validation, web-based research and myriad due diligence activities conducted by underwriting associates and underwriters.

We built a Proof of Concept that demonstrated that these functions could be aided by an LLM-based Agent application that performed these functions, summarized findings and generated preliminary underwriting recommendations based on company guidelines complete with decision support citing the data and applicable business rules. The preliminary recommendations were provided to underwriters for review and final decision making and produced the desired productivity gains. That solution is now in production development — for a function that is literally charged with financial risk management — so risk may be viewed differently in the are of finance you participate in.

This is just one use case that represents (a) LLM-based automation that is being implemented in insurance companies across the globe, and (b) a fundamental repeatable pattern that is applicable to many similar use cases across the finance industry.

So, again, while this doesn’t provide definitive evidence of unfailing success, it does support the contention that pronouncements of certain failure are premature.

You're at proof of concept still. It's not the only way to improve productivity in insurers I guarantee that. They are some of the most inefficient **** shows on the planet.

I've literally worked in the risk space on ML projects before LLMs came along. Prior to that I designed fact find and analytic modelling tools right down to rule engine DSLs and recommendation engines.

The original ML approach, which was functionally unrelated but had higher determinism than an LLM system would, was discarded because the aggregate risk of a poor decision that impacts the organisation and leads to regulatory capture is higher than using simple heuristic analysis. On top of that because decisions need to be reproducible and auditable you need to capture the entire state of the model that was used at the time. In the ML space that meant capturing the entire model state / weights which was non trivial to store. That's even more fun when you have to do something computationally complex to test the models under extreme conditions via monte carlo method.

Also you can't really go to your compliance and audit team and say "well our CI is 95% on this making 90% of decisions within the existing heuristic model test cases but that other 10% we insure a Russian oil tanker made of rusty holes". They go "pfft. project canned".

There's too much faith and not a lot of science going on here.

 

[heretiq]

You're at proof of concept still. It's not the only way to improve productivity in insurers I guarantee that. They are some of the most inefficient **** shows on the planet.

I've literally worked in the risk space on ML projects before LLMs came along. Prior to that I designed fact find and analytic modelling tools right down to rule engine DSLs and recommendation engines.

The original ML approach, which was functionally unrelated but had higher determinism than an LLM system would, was discarded because the aggregate risk of a poor decision that impacts the organisation and leads to regulatory capture is higher than using simple heuristic analysis. On top of that because decisions need to be reproducible and auditable you need to capture the entire state of the model that was used at the time. In the ML space that meant capturing the entire model state / weights which was non trivial to store. That's even more fun when you have to do something computationally complex to test the models under extreme conditions via monte carlo method.

Also you can't really go to your compliance and audit team and say "well our CI is 95% on this making 90% of decisions within the existing heuristic model test cases but that other 10% we insure a Russian oil tanker made of rusty holes". They go "pfft. project canned".

There's too much faith and not a lot of science going on here.

Your points are historically accurate and still valid in many cases. My point is that the issues you point out are not conclusively terminal. They are taken seriously and have considerable effort directed at working them out.

Advances in modular AI application architecture, knowledge graphs, agents, and rigorous testing have leapfrogged historical AI/ML systems and enabled apps that are more efficient and capable of directing reasoning to deterministic subsystems where required, and to non-deterministic subsystems for processes that benefit from LLM capabilities.

And this is not just POCs — there are many production systems that have been approved by BODs, validated by auditors for green-lit use cases, and deployed to production in insurance, manufacturing and other industries.

Is there still much experimentation? Yes. Are the success stories definitive proof that LLM technology has no risk of failure in the long term? No. Are the failures conclusive proof that LLM technology is destined to fail? No. Net-net: The LLM experiment continues because it has not conclusively failed.

P.S. Not being argumentative. I respect your knowledge, perspective, rationale, and appreciate the opportunity to discuss this topic in a non-superficial way. Also not saying or suggesting that you’re wrong — only that I think it’s too early to call this technology a dead-end.

P.P.S. If there’s any doubt that LLM-based Agentic AI is the future of software development, look no further than Microsoft’s announcement yesterday: https://arstechnica.com/gadgets/202...rosoft-reorganizes-entire-dev-team-around-ai/

 

[I7guy]

Agreed, but a reported 73% of Apple users chose to not install Apple Intelligence when the honking great window was part of the install at the last OS update. Very little has come out since in the Media (tech or otherwise) to suggest that it's an essential bit of kit, while the option to enable it afterwards does exist. Like lots of hidden features / controls that exist on even the iPhone, but we don't get popup nags about them. The web can always direct even the non-tech nerds to them, if need be. And if an app / feature doesn't gain traction, then yeah, maybe it isn't worth the trouble.

A reported 60% of eligible iPhone users installed apple intelligence.

 

[Tfones]

This better not be a thing. Both my phone and mac today...

View attachment 2472237

If it comes back again then this is Windows 11 level of nag.

He wants you bro

 

[adrianlondon]

I installed MacOS 15.3 beta 3 yesterday and on completion it asked for a WiFi network (by complete coincidence I was rebooting my router) so skipped that after which it told me that Apple Inteligence was being activated! No option to say "no" or cancel.

So clicked "next" and luckily within a few seconds I was back at the desktop with System Settings open, so immediately disabled it.

I seem to have escaped its clutches for another day.

 

[navaira]

I seem to have escaped its clutches for another day.

It already took your storage space, though :/

P.P.S. If there’s any doubt that LLM-based Agentic AI is the future of software development, look no further than Microsoft’s announcement yesterday

This feels a lot like “if there’s any doubt that Arm is never going anywhere near real computers, look no further than Intel’s announcements” did before M1.

 

[heretiq]

This feels a lot like “if there’s any doubt that Arm is never going anywhere near real computers, look no further than Intel’s announcements” did before M1.

Not sure what the point of that comment is but Microsoft is unifying their entire development divisions behind AI. In Microsoft’s own words:

“This new division will bring together Dev Div, AI Platform, and some key teams from the Office of the CTO (AI Supercomputer, AI Agentic Runtimes, and Engineering Thrive), with the mission to build the end-to-end Copilot & AI stack for both our first-party and third-party customers to build and run AI apps and agents. This group will also build out GitHub Copilot, thus having a tight feedback loop between the leading AI-first product and the AI platform to motivate the stack and its roadmap.”

Microsoft is not asking for permission to embed AI tooling into its development platforms and runtimes that are included on every windows PC, they’re doing it. And developers will use these tools to deliver apps based on LLM-based Agentic development paradigm. LLMs are a fundamental evolution of computing architecture is here to stay whether consumers ask for it or not.

In my opinion, this is not comparable to Intel doubling down on x86 vs embracing the competing ARM architecture — a decision which was likely seen by Intel as an admission of failure. This is Microsoft doing what Intel should have done: Accept the fact that the world has changed, embrace the new, potentially unsettling but disruptive reality and use it to your advantage.