Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Yikes!

Definitely avoiding that mess.

If you use google then... no you aren't...


As an aside, I barely use google any more because I get more relevant results out of ChatGPT - it searches the web and filters out the irrelevant crap for me, deep research is even better.

Has been invaluable for 2 major projects I've knocked off this week.
 
....and The Browser Company has already released a beta version of their own "AI browser" named Dia for macOS, remember...

It's insane how much change is happening in the AI field...wasn't it just 2 years ago when we were barely able to generate a realistic video using AI? Now, halfway through the 3rd year with LLMs, we are already seeing competition on who can come out with an AI-first browser.

Just an insane pace of progression of AI...when will it slow down?

I remember thinking this same thought when the iPhone was first introduced, and Apple was showing those graphs at WWDC or at their annual iPhone events every year showing at the rapid performance advancement of the A-series chips and the hockey-puck curve...now, Apple doesn't show that slide anymore because the A-series chips/M-series chips have basically matured and the performance curve has plateaued. Not the case the AI. Who knows what is going to happen to our society once AI has matured and the performance has plateaued and tech has moved on to the next thing...

Unimaginable times we are living in folks...

Just wanted to vent a little and put down my 2-cents.

Thanks for reading...
Separating the wheat from the chaff is extremely difficult because we are in a hype cycle but you are right with the level of progress. Deepseek was known among the research community roughly 9 months ago and that feels like ancient history now.

I think there's a chance that without a major fundamental discovery we will see a cooling-off period within 18-36 months, then an extreme ramp up again afterward toward the end of the decade as the longer-term research starts to pay off. But honestly, a lot of people used Siri regularly even though it didn't really improve – the same thing may happen with current GenAI tech and the demand may outpace supply until the next generational breakthroughs happen.

I saw deterministic code generation including full test cases in one shot about 5 years ago at a research company, this was well before GitHub Copilot and also performed much better in the domain it targeted.After that presentation I kind of stared into the middle distance for a while just absorbing the ramifications of what I saw.

I have posts as recently as ~6 months ago in this forum where I vehemently argue that AI cannot summarize because it misses key points, and there was research to back that up. That is no longer the case although there are still some errors and issues they are drastically reduced.

With the current technology we won't ever perfect things (World Models will likely be necessary, not LLM + tools), but they are now to the point where they can often accelerate rather than detract. What's really crazy to me is if you used the latest models in February of this year and based your judgments on that, you are already woefully out of date. I have never seen anything move at this speed, and while there are scaling limits I wonder how much they can be minimized with tool use and agentic workflows etc. that kind of suppress the drawbacks.

18-24 months ago ChatGPT was widely being torn down as being a stochastic parrot which was true in some sense. Now, with scaling and some additional breakthroughs there are many more tokens that are intermediary and actually do function as a sort of "reasoning" in a high-level metaphorical (not biological!) sense. Emergent behavior is real, but don't let anyone tell you that will lead to AGI or some similar nonsense, it will not.

Unimaginable is a good word for it, and I don't love the implications of any of this GenAI technology on the world. As a larger public, we are not prepared at all, and I don't mean from a safety perspective, I mean from a societal one.

...

On-topic, integrated web search dramatically improved the utility of LLMs since you can point them at documentation or have them search for up to date information which helps get around knowledge cutoff limits in their training data. It's smart for OpenAI to release a browser and if they product manage it correctly I could imagine it competing at a pretty high level, probably surpassing Firefox within a year if they have a fully fledged and well thought-out product.

Many of us switched to Chrome rapidly when it launched, OpenAI is wise to pursue this avenue, particularly as "there is no moat" to some degree. They need propriety offerings to help lock-in customers.
 
Last edited:
I can’t wait for this. ChatGPT has removed the need to use a browser to search for something then scroll through endless links. Ask, get an answer, done. I think this will make traditional browsers redundant.

Yup, a lot of people are poo-pooing AI, but it is moving fast and the paid services are actually good.

Legitimately, this week it has been genuinely helpful for me with a few tasks:
  • 365 tenant migration
  • Hyper-V cluster build and issue diagnostics
It has also helped explain WHY some of the things have been configured the way they are in other places in my environment. You can tell chatgpt things that are relevant and it will adjust the results to work with your specific scenario.

Why do you search the web for things? To get answers. ChatGPT and others can give you answers with citations for them for you to judge the accuracy.

Much faster than sorting through the "sponsored link" garbage and SEO trash that most engines are filled with these days.
 
Will try it out. But really not sure on how privacy will be respected in such a browser. More browsers will definitely try to incorporate AI as much as possible into their products.
 
  • Like
Reactions: mganu
If this has proper ad blocking, I’ll give it a try. I’ve stopped using Chrome because Google nerfed ad blocking extensions. Firefox and Safari are my current browsers.
 
  • Like
Reactions: chibihabibi
That is just the start. The next step is a new type a user interface for a complete operating system. The current point and click GUI is so old school and a shakeup is needed.


A whole OS that just "does stuff for you" and "gives you answers" ..

..and all of it only correct "some of the time" and full of skewed information and pushing narratives that have been steered by the incentives of the company (or even a single person!) making the system.

So cool!
The future!
/s

UGH
 
As a larger public, we are not prepared at all, and I don't mean from a safety perspective, I mean from a societal one.
Indeed. Well said. This is what I am most fearful of as well.

This technology has come at exactly the wrong time - when the younger generation is reading less and less, the attention span and the importance of critical thinking is shrinking at a rapid rate - and we are blindly putting our faith into a tool that we understand at a primitive level.
 
I can’t wait for this. ChatGPT has removed the need to use a browser to search for something then scroll through endless links. Ask, get an answer, done. I think this will make traditional browsers redundant.

The notion that there are more people like you is frightening. ChatGPT gives the wrong answer half of the time. Relying on it is practically dangerous!

If you know what you're looking for, you'll know that it's giving you half wrong answers, and if you already know what you'r looking for, it's pointless. Only if you have no clue do you think ChatGPT is correct most of the time.
 
Yup, a lot of people are poo-pooing AI, but it is moving fast and the paid services are actually good.

Legitimately, this week it has been genuinely helpful for me with a few tasks:
  • 365 tenant migration
  • Hyper-V cluster build and issue diagnostics
It has also helped explain WHY some of the things have been configured the way they are in other places in my environment. You can tell chatgpt things that are relevant and it will adjust the results to work with your specific scenario.

Why do you search the web for things? To get answers. ChatGPT and others can give you answers with citations for them for you to judge the accuracy.

Much faster than sorting through the "sponsored link" garbage and SEO trash that most engines are filled with these days.

I've used various LLMs to aid me in development, and all of them are varying degrees of "telling a rookie how to do something for you", GPT 4.1 in particular just all over the place.

The time-savings aspect is an illusion.
 
The notion that there are more people like you is frightening. ChatGPT gives the wrong answer half of the time. Relying on it is practically dangerous!

If you know what you're looking for, you'll know that it's giving you half wrong answers, and if you already know what you'r looking for, it's pointless. Only if you have no clue do you think ChatGPT is correct most of the time.
In my opinion there is a lot of FUD around use of AI tools, and a lot of opinions seem to have been cemented when it was less reliable than it is now.

Personally, I've used it successfully to save myself and my company hundreds of hours of time and thousands of dollars. In my experience, saying it gives the wrong answer "half the time" is wildly off. Maybe 10% of the time, and that's probably overstating it. Yes, you need to know how to use it correctly, and yes, I proof any output that is mission critical, and occasionally spot check stuff that isn't mission critical, but I do the same thing with my junior employees' work output (In other words, I treat it like a junior employee).

I believe people dismiss it at their peril. Learning how to use it is going to be essential. But obviously, that isn't going to apply to all occupations equally.
 
The notion that there are more people like you is frightening. ChatGPT gives the wrong answer half of the time. Relying on it is practically dangerous!

If you know what you're looking for, you'll know that it's giving you half wrong answers, and if you already know what you'r looking for, it's pointless. Only if you have no clue do you think ChatGPT is correct most of the time.
The same can be said for the entire internet.
 
In my opinion there is a lot of FUD around use of AI tools, and a lot of opinions seem to have been cemented when it was less reliable than it is now.

Personally, I've used it successfully to save myself and my company hundreds of hours of time and thousands of dollars. In my experience, saying it gives the wrong answer "half the time" is wildly off. Maybe 10% of the time, and that's probably overstating it. Yes, you need to know how to use it correctly, and yes, I proof any output that is mission critical, and occasionally spot check stuff that isn't mission critical, but I do the same thing with my junior employees' work output (In other words, I treat it like a junior employee).
In my experience, 10% is understating it. I guess it depends on your subject, and even Perplexity, no matter which LLM you chose, writes up answers that are frequently wrong. Nearly every answer has at least one thing wrong in it. I find LLMs to be like a keen work experience kid, and Perplexity is like a keen work experience kid who loves researching, and is actually good at it.
I believe people dismiss it at their peril. Learning how to use it is going to be essential. But obviously, that isn't going to apply to all occupations equally.
This is the truth. I have been against LLMs for a long time, but I felt the need to learn to use them for myself. Firstly because my arguments can’t be very solid if I don’t have personal experience of what I’m arguing about unless I restrict my arguments to purely philosophical grounds, and secondly, while I expect the bubble to burst at some point, LLMs seem here to stay, and if I’m to help guide the next generation, and assist my peers, I need to know how to use them.

They don’t impress me much, though they can be very useful in specific circumstances.
 
  • Like
Reactions: cjgrif and surferfb
I've used various LLMs to aid me in development, and all of them are varying degrees of "telling a rookie how to do something for you", GPT 4.1 in particular just all over the place.

It really depends on how you are using it and the questions you ask. If you have a solid idea of what you are doing it is an excellent source for indexing and summarising online documentation on the subject.

For app development? I don't use it for that. I can see how you can use it for brief snippets where you're basically giving it a process/algorithm to turn into the required language of choice.


in summary: search, translation summarisation/analysis of documentation are the things LLMs are actually really good at. People avoiding using them for search (when they source the information provided/link directly to it) is a bit odd.... this is probably the biggest thing you can actually use AI for effectively right now.

If you want something to help with code you're better off with Claude.

But seriously, right now in this field, 3-6 months is the equivalent of 3-6 years for the rest of the software industry. It really is moving that fast right now.

Anyone who is basing their assessment on Siri or some experience from last year, etc. should re-assess on a regular basis.

Right now you also need to pay a bit of attention to which model you are using and verify that it's the right tool for the job. They're not all the same and "reasoning" models aren't always "better" for the task at hand.
 
Last edited:
  • Like
Reactions: novagamer
Personally, I've used it successfully to save myself and my company hundreds of hours of time and thousands of dollars. In my experience, saying it gives the wrong answer "half the time" is wildly off. Maybe 10% of the time, and that's probably overstating it.

This. You can also ask it to cite sources and judge for yourself. If you use a paid account and use things like deep research, everything is cited.

Yes, you need to know how to use it correctly, and yes, I proof any output that is mission critical, and occasionally spot check stuff that isn't mission critical, but I do the same thing with my junior employees' work output (In other words, I treat it like a junior employee).

Also this. Treat it like an assistant/junior to do bulk amounts of grunt work that you can validate. It is generally a lot faster to validate output than create it.

its not a replacement for humans, nowhere near. It's simply a tool, like a shovel. You can dig holes without one, but it's a lot quicker using one.
 
Last edited:
This. You can also ask it to cite sources and judge for yourself. If you use a paid account and use things like deep research, everything is cited.
Often cited wrong, in my experience. I am actually finding Perplexity Pro using Claude 4.0 Sonnet useful, but for its citations, not its answers. In the field I'm using it in, it's wrong nearly as often as it's right.

It does find better sources than I can do by myself with Google though.
 
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.