Prompter prompts you say?
What areas, outside of LLMs and AI will this benefit? They mentioned 2x FP16 also.
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.
Apple is falling behind, M4 Max nowhere close to RTX 5090.
- Thread starter Zest28
- Start date
- Sort by reaction score
What areas, outside of LLMs and AI will this benefit? They mentioned 2x FP16 also.
For gaming, it'd be MetalFX. Nvidia's DLSS is done on the Tensor core. MetalFX can use these new cores.
There is some additional interesting tech likely included with the new GPU MXU units, so who knows, maybe the performance uplift will be even better.
The matrix stuff is pretty much limited to AI, but it’s becoming increasingly important in graphics too. Nvidia has been pushing neural shaders quite a lot, and there are nice things like procedural materials etc. that can be done.
That could help some gaming shaders if they are optimized appropriately. For example FP16 is sufficient for a lot of operations related to lighting.
I’d think that ANE is the better fit for MetalFX. It’s more energy efficient and frees up the GPU to do other work.
Will be really interesting if they keep the Neural Accelerator per GPU core for the M series. If so, for certain tasks it will be a monster performance improvement.
Really looking forward to M5 now and I'm glad they delayed it to (hopefully) get these improvements and others in.
You started out more or less promoting LLMs as a magic button you press to get production worthy code, yet now that a few people have posted about their experiences with them making mistakes your tune has changed. Interesting.
If a LLM requires hypervigilance to use, why should I use it? Handing off work to something that's likely to screw up, in unpredictable ways, with greatly varying levels of subtlety from one attempt to the next? That's not my ideal productivity enhancer.
People have compared LLMs to junior devs, but at least a bright junior dev can be expected to (a) understand and reason about the problem statement to some degree and (b) learn from their mistakes, neither of which are properties of LLMs. There's also a long term problem here: if you're using LLMs instead of junior devs, where are you getting the next generation of experienced devs to watch over LLMs?
And if you are an experienced dev, be aware that there are studies showing that programmers who use LLMs are slower and more error prone than programmers who don't. Yes, these studies covered people who love LLMs and swear by them. There's even evidence that heavy reliance on LLMs decreases cognitive ability over time - instead of exercising your own reasoning, you're training yourself to stop thinking and ask the AI to think for you.
AI bubble hype is causing problems even for people who haven't opted in. For example, the curl project is being DDOS'd by fake AI-generated bug reports. Many of these seem to be filed by sincere people who got fooled into believing they could "contribute" to an open source project by just asking a chatbot to analyze it for security bugs. They've tried putting in the rules that you must disclose up front whether you used AI, or face an immediate ban, but that just got many of these AI believers to start hiding what they were doing. After all, they think the AI has told them about a real bug, so they better do what they can to get it in front of people's eyes.
Death by a thousand slops
I have previously blogged about the relatively new trend of AI slop in vulnerability reports submitted to curl and how it hurts and exhausts us. This trend does not seem to slow down. On the contrary, it seems that we have recently not only received more AI slop but also more human slop. The...
daniel.haxx.se
There's also a ton of ethical and environmental issues with so-called "generative AI", but I bet you're one of the people who would just handwave such concerns away.
they delayed it? i thought only the Mbp are delayed and those are rumoured to get the M6 family
M5 probably ipads pro and Mba in the spring
You're completely missing the point. I said it can one shot new apps. It doesn't mean I'm going to let it run wild in my business critical application without review. It writes 90% of the code in my business critical app but all of it is reviewed by me.
The rest of your post is just regurgitating anti-LLM garbage, no offense. It looks like you just searched for why LLMs are bad, and then posted what you found.
Last edited:
A19 Pro Geekbench Metal score
browser.geekbench.com
iPhone18,2 - Geekbench
Benchmark results for an iPhone18,2 with an ARM processor.
browser.geekbench.com
Massive 41% better: https://browser.geekbench.com/v6/compute/compare/4765303?baseline=4764370A19 Pro Geekbench Metal score
View attachment 2545753iPhone18,2 - Geekbench
Benchmark results for an iPhone18,2 with an ARM processor.
I must say I was quite confused by the attitude regarding junior devs in some posts. In my organization people who are incompetent, rude, and unwilling to learn do not keep their role very long. It’s not my job to babysit entitled beginners. Everyone contributes according to their ability and grows over time.
LLMs are a useful tool, as long as one uses them responsibly. And they are getting larger and more accurate at some tasks. But I don’t see a qualitative improvement in the last generation models, at least not on tasks that are relevant to me. I do believe that future models will have better performance, but they will use different architecture than token prediction.
Last edited:
These are pretty much the same cores, so yes, the new GPU cores will certainly be present in M5. Apple is serious about making the MacBook more interesting to ML researchers and developers.
Better cooling is just a small part of it. The GPU execution units got substantially wider. The new GPU can now execute two 16-bit precision operations per cycle, or a certain mix of 32-bit operations. These things add up.
Yeah, I mean we all were at that level at some point, its kind of sad. I used to train programmers in my job, and its something that the more you put in to help the more you'll get out, both personally and professionally
Agree but if the cooling cannot keep up with it, its tech wasted, especially in something so thin like the air
I don't think that's the plan, I meant MBP M5 in early 2026 vs. the yearly cadence, which is the rumor. I'm really looking forward to these new chips and may upgrade M4 Max -> M5 Max if they pan out, if the A19Pro is an indicator of what's coming. Really good stuff.
I get why you say this but their point specifically about LLMs not learning from their mistakes is extremely valid. It's an incredible, probably intractable problem given the nature of the tools.
We do have better versions of MoE and a bunch of stuff happening behind the scenes now where they can review things, especially code, to find errors much of the time which is great and makes them useful for scaffolding and the "thinking" / "reasoning" model paradigm plus web search has opened up a lot more every-day use cases, but without being corrected and internalizing that knowledge there will be a hard wall on how far these tools can advance since this also blocks off self-learning which is going to be an even greater challenge.
Exactly, you need to be able to learn, to have true agency and improve. Right now that's impossible.
I agree with you that a future novel architecture or hybrid will improve this but scaling LLM technology to consume all the power on earth can not solve it, despite the foundation model companies wanting us and the VCs who fund them in particular to think so.
There's a gap between people who write off the technology that has real use cases now and also the people who think we've solved all of the problems or we're on a path to inevitably do so. I think you recognize this, but some others–I'm loathe to use the term–"on both sides" really miss this and it's a large blind spot. AI research talent is being scooped up to try to solve these problems, as well as stifle competition of course per usual Silicon Valley antics.
They will probably have at least some success and it's a pretty cool time to be alive with respect to some of these technologies existing at all, but it's prudent to take a somewhat reasoned approach about how and where to fit them into workflows and daily life. Society will do what it will, and so will corporations, but we must decide for ourselves ultimately when we can and push on the things that aren't good enough despite what we're told.
Last edited:
Same, though as I've mentioned I'm really happy with my M4 Max as it stands. I'm more curious to see if apple changes the architecture or just add improvements to the existing setup.
What would constitute a change of architecture to you? From what we’ve seen yesterday there are major changes in both the CPU and the GPU in this new generation.
The two rumors I saw in various websites seem to fit the bill. I'm not saying these are viable, just stuff you see on the interwebs.
The first, a change the unified memory architecture and focusing dedicated memory for gpu operations, and the second is more dynamic options when buying a Mac, that is, having the ability to choose 30 gpu cores, or 40 w/o needing a new cpu selection. That may not strictly be a architecture change, but its a change in how apple does business.