This won't happen in your lifetime.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
xAI Launches Grok-3 AI Model, Claims Superior Performance Over GPT-4
- Thread starter MacRumors
- Start date
- Sort by reaction score
This won't happen in your lifetime.
Did you check all the links and prices were real which ChatGPT will tell you itself isn’t reliable. There’s literally a small text at the bottom of the window telling your to double check the data.
Yeah don’t reply with ‘YES it wAS aLL rEALL’. We know the results can wildly swing.
HobeSoundDarryl
macrumors G5
His AIs will soon be able to give you the social security, address, email of any US citizen. You may also ask his AIs how to hack into the IRS, medicare and other US agencies and why not, also ask for the admin password of the Pentagon.
Glad to see everyone cheering for King Musk the first…
Glad to see everyone cheering for King Musk the first…
Well put, and I agree. If these AI models were so great, they would seek out the objective truth and offer it. That's not the reality. Howver, this would eliminate the human emotional aspect of certain jobs like doctors (it would also eliminate sleep depravation), psychiatrists, judges, etc. I can't wait for AI to replace some and maybe all of those jobs one day. As long as it is 100% accurate of course.
HobeSoundDarryl
macrumors G5
They kept hard drives in the sky inflated for about 15+ years. This- as next big thing (whether it is or not)- will probably have AT least as long. Look for the prefixes to show up to keep extending the spin: Hybrid A.I., A.I. cloud, A.I Hybrid Cloud, etc. Where's there a dollar to be harvested, there will be fresh tech spin... and sometimes some actual innovation.
Last edited:
After rejecting Elon’s $97.4 billion bid to buy OpenAI, Sam said: “I wish he would just compete by building a better product.”
Well, Elon did just that and smashed it out of the park with Grok 3, which wipes the floor across AI benchmarks. It’s the first model to score over 1400 on Chatbot Arena and has wowed everyone who has tested it. Looks like Sam’s playing catch-up now.
Well, Elon did just that and smashed it out of the park with Grok 3, which wipes the floor across AI benchmarks. It’s the first model to score over 1400 on Chatbot Arena and has wowed everyone who has tested it. Looks like Sam’s playing catch-up now.
Yes of course I checked, it was all accurate and up to date. I find the results to be more and more reliable. In the beginning a lot of answers were just fantasies and made up stuff.
ChatGPT evolved into a faster and more precise Google when searching something complex.
Winning AI benchmarks is like being the skinniest kid at fat camp.
HobeSoundDarryl
macrumors G5
The rejection was only days ago. I'm doubting Elon since "did just that and smashed..." This had to be in development for longer than last few days and thus "out of the park" would have been realized whether Sam rolled over and accepted the offer or not... as some accepted deal would not be close to being complete by today too.
If you're from the US, I have bad news for you...
Model benchmarks are quite silly anyway. Math benches are fine, but most of the other benchmarks are subjective. They aren’t tangible benchmarks like Geekbench measuring CPU performance.
The AI guys even have a name for cheating at benchmarks. If engineers know the tests ahead of time they can tweak the models to perform well on benchmarks. They call this Benchmaxxing.
One super important thing to always remember when it comes to generative AI.
Never trust any demo you see online. They are all carefully cut and edited. Use that stuff yourself and then you’ll understand where they are and what they can and cannot do.
Then you’ll always understand the lies in the media. Journalists are terrible at talking about tech and sometimes that is deliberate.
Last edited:
"maximally truth-seeking AI,"
It feels incredibly snake oily to even suggest something like this
I'm certainly no expert in these (perhaps someone will chime in?), but I thought it was fairly "known" that an issue with LLMs is that they absolutely have no way of knowing on their own or through their own methods of learning, what is factually correct
This part about analyzing "information from X" as a data source for seeking "maximum truth"...
"The new feature is designed to analyze information from the internet and X (Twitter) to provide comprehensive answers to user queries."
Not great
Generally speaking, anyone seeking to suss out factual information usually doesn't speak in terms like "maximal truth".
That verbiage is trying to simplify something that isn't simple or black and white and is far more nuanced and full of subjectivity and required context
It feels incredibly snake oily to even suggest something like this
I'm certainly no expert in these (perhaps someone will chime in?), but I thought it was fairly "known" that an issue with LLMs is that they absolutely have no way of knowing on their own or through their own methods of learning, what is factually correct
This part about analyzing "information from X" as a data source for seeking "maximum truth"...
"The new feature is designed to analyze information from the internet and X (Twitter) to provide comprehensive answers to user queries."
Not great
Generally speaking, anyone seeking to suss out factual information usually doesn't speak in terms like "maximal truth".
That verbiage is trying to simplify something that isn't simple or black and white and is far more nuanced and full of subjectivity and required context
Last edited:
Yeah but they saw that Will Smith movie with the robot so it must be true, especially if the crooks who run CNBC keep saying its coming any day now.
I will not use or buy anything associated with Elon Musk.
Ask any of the top performing models to generate code for an Asteroids clone or Pac Man clone. They came out in 1979 and 1980 respectively. They barely used 50 lines of machine code when they came out.
The best models can barely get Snake right in one shot. They require a lot of hand holding to generate a super basic version of Tetris. They utterly fail at Asteroids or Pac Man.
So despite eating all the data on GitHub, where there are a dozen Asteroids and Pac Man clones, and despite the massive amount of compute power being used, the programming skills of the best models are still utterly sloppy by themselves.
They produce good boiler plate code, because they have been shown tons of boiler plate code, but they cannot program applications by themselves. Even the current wave of ‘reasoning models’ just get stuck in a loop asking dumb questions and failing if they are tasked with creating applications.
I didn’t say whether it was any good!
Open AI has a model that it hopes will provide an objective truth to such extent that it will monetise itself.
If it’s seen to be biased or deliberately misleading it will lose its purpose and go out of business.
This new model isn’t so interested in truth, it’s interested in power. It doesn’t have to make money, it’s buying something, influence.
Its sister social media site didn’t make money, independent sources value it at far below its purchase price. So the question becomes, “What did this otherwise smart man get for his money?”
Power
HobeSoundDarryl
macrumors G5
Ahhhhhhhhhh, but identify the finest, most capable programmer in the world. Let's call him/her Casey. Back up time to when Casey was a toddler and ask them to code Pac Man or Asteroids. I would guess Casey might only try to chew the keyboard. But give Casey time to grow knowledge and eventually Casey becomes smartest/best programmer in the world where such a challenge is trivial.
A.I. Casey can learn faster than H.I. (human intelligence) Casey. Thus A.I. Casey can grow up far faster than H.I. Casey. While current A.I. Casey might also try to eat the keyboard, skilled A.I. Casey is only "how fast can it learn to be a great programmer" away. That timetable will be measured in the speed that computer knowledge & skills can accumulate vs. being measured in number of days and years a human can spend in various tiers of school.
Last edited:
Perhaps there's nine of him?
Well we know there's now 13 and counting offshoots 😂
Deepseek has made all of this irrelevant. Open source in full privacy is where it's at. Mac users can run the 70 billion parameter R1 model at eight bits MLX on a M3 MacBook Pro (128 Gig Ram) entirely off-line. It's the best model I've ever used hands-down. Next generation of the Mac studio ultra and we should be able to run the full R1 at eight bits. Musk who is supposed to be so open source is going to release his outdated version open source at some future date. Hah! Big thanks but no. His commitment to open source is zero. Speak against that guy with your wallet.
For anyone wondering why "they don't just shut AI down" or why AI is so important, perhaps read The Economist's article "How AI will divide the best from the rest" at https://archive.ph/Sl9sA
