Interestingly what I’ve said here ☝️ ️comes with an interesting implication. There used to be a type of software “engineer” that couldn’t actually think through a problem and implement a good software solution. They were valuable because they could refactor, write code, but not think through a problem. These are the people who have the most to lose at the moment, and also the ones that find the least value on AI, because AI only does really well the single thing they could do.

I appreciate the intent behind that sentence, and the problem is that is written (or said) by people who don’t understand coding.

Syntax was never the hard part. Code syntax is a set of rules that you can teach a machine and it’s “easy” enough for an LLM to grasp it. The problem here is that this phrase does a “code = syntax” but ignoring all the difficult things that actually are in the code. Architecture, design, tradeoffs, translation of vague requirements into concrete software components…

All of those things are terribly hard, and the reason why many LLM-first coding projects get abandoned at 70% and then companies need to hire a Software Development consultancy to mop up their mess.

Are these any good? I mean, within their size class, obviously - not expecting them to compare to Kimi K3!

Models the size of that 135M one open up some interesting use cases for edge devices.

What they forgot to add: in the same time, they’ve also managed to create more bugs than they’ve fixed, thanks to AI.

Well, MacBook Pro or also Mac Studio. Pre shortage, you could spec out the Mac Studio with 512 GB of RAM, if you had stupidly deep pockets.

Unfortunately AI doesn’t work like that. Any way to explain it would be an oversimplification but I can try.

The training data (songs) are used to create the weights. This is a bunch of numbers that are on their own meaningless - they don’t map to specific songs, but to attributes such as “tone” “rhythm”… And like that but many (millions of) abstract attributes that don’t make sense as people, but make sense to computers.

So if the thing makes a song that is very “rhythmic” but also “tonal”, there’s no specific training song that contributed to that - all did, and it’s a mess to decypher how much each contributed to each attribute. Except the resulting song doesn’t use two parameters, uses many millions, so it’s essentially impossible to know.

It was at the beginning, but lately it’s been hot garbage. I think they’ve nerfed it on purpose so that the migration to gemini is an easier sell.

For me, I’m happy it now understands “hey Google, how’s my timer?” As equivalent of “how long is left on my timer” and no longer answers that question with “it’s seven fifty two pm” 😡

Yes, but at the same time can we stop marketing as “self driving cars” normal cars with a somewhat sophisticated cruise control, like Teslas, and stop pretending their “super full self driving unsupervised for realsies plus plus” is a “self drive mode”?

Well, there’s nothing yet that even resembles a comparable replacement with a different architecture (RISC-V?). So even if they were angling for that, it would have to be at least 3 years away, plus if Intel and Power PC are anything to go by, there’s another 5 years until they drop support. So at a minimum, if someone buys a M-series laptop today, they can expect support for 8 years.

Not terrible, given how Microsoft left 3-year-old computers unsupported by surprise with the TPM requirement in Windows 11.

But this is only because of execs’ stupidity.

For a simple task that AI can actually do, say a boring text processing task, the API calls don’t just cost less than my salary, they typically cost less than keeping the monitor on during the time it would take me to do it.

However when companies are stupid and decide to do things such as “tokenmaxxing” or leaderboards for who can waste more AI compute, you end up with things like calling multi-trillion models to do number calculations, or passing a 300k token context into every turn of the chat and giving it your entire codebase to change two lines of code.

This one I blame squarely on stupid CEOs and execs. The smaller version of Gemma 4 is low-powered enough that can run on phones, and can output usable results for many use cases.

Isn’t this from building data centres and renting them out to other companies (e.g. Anthropic) that have AI tools/models that have at least some utility?

Depends on what the targets are for the C-Suite, but in 2026 one could assume yes.

Like Gmail? Google drive? Slack?

I’m not defending AI, but I can come up with >10 products that would absolutely cripple the company I work at if the provider suddenly says “Soz, terms of service violation”.

Vendor reliance is dangerous. That doesn’t just apply to AI. If the company in OP’s message had both Claude and Gemini they’d been okay, so the problem isn’t with AI explicitly - the problem is with reliance on services that are critical for workflows, and providers being able to change their mind at a moment’s notice.

In any case, leaving aside where the problem is, the idea that 60 employees can’t use Natural Intelligence to do their jobs means there’s something really wrong with that company…

First of all, I’m going to say that I don’t think this comparison actually makes sense and I was just entertaining the question of the message I was replying to - humans are machines are way too different to reduce the comparison to merely “which is more energy efficient”.

But second, I compared to the same level - I stopped at infrastructure. I didn’t consider the costs (energy or otherwise) of building a solar panel or power plants in the same way I didn’t consider the costs of a frying pan, a hob, or farms. Because if we do that, then any point we make about this needs to be a 500 page dissertation, not a Lemmy message.

The good news is that data for how much material/energy is required for a solar panel is freely available, and also that a solar panel can be used for energy generation many more times than a cow.

Well, in pure energy usage, no; however if you take into account the energy usage of the whole chain, they’re orders of magnitude better.

After all, they can even be hooked up to a solar panel directly. For us to get 2000 kJ of energy, we need to water plants for a year, transport them, spend more than 2000 kJ of electricity cooking, and that’s not even considering raising an animal for x months or years which needs >5000 kJ a day to just exist. Our sun->movement energy efficiency rate is pretty appalling and orders of magnitude worse than a robot’s - even if the robot is just hooked onto the regular grid.

At the risk of being called a corporate bootlicker, it sounds like it isn’t their own, it was their employer’s.

If it’s your own, absolutely, fair play. If it’s your employer’s… Then it feels murkier. I wouldn’t blame their IT department for being quite cross if/when time comes to upgrade and return the laptop and give it a second life. For example, I would be quite pissed if IT gave me a defaced laptop like this as a loaner while mine is getting repaired.

I have never seen “immense” spelled that way!

That’s a big difference but not all. The sub-$1000 ultrabook sector has SO MUCH garbage, like Intel Celerons that stutter when you scroll down a web page designed in 2022+. Manufacturers are happy because they can sell rubbish and uncle John with no idea about computers will say “I want a laptop with 1 TB so it’s faster, and it must have free office 365 and an antivirus”…

So when someone puts a phone processor in a laptop and builds a chassis that isn’t a $5 extruded plastic shell, they panic because it still manages to be better in both benchmarks and real world use despite the paltry amount of RAM.

Performance wise it’s an interesting one. I think from a price and energy standpoint it sits squarely against windows ultrabooks with a Snapdragon X, for example, a Galaxy Book 4 Edge.

Based purely on benchmarks, the A18 Pro is weaker than that, plus you have only 8GB of RAM.

However - I have a Surface Pro X with the original SQ1, with roughly 40% of the performance of these… And even at that level, the problem is Windows on ARM, not the performance. It only lets you down for things it’s clearly not meant to do, like video editing.

Another alternative I see for that price is a windows laptop with an i5-1334U, which theoretically gives you a raw performance within 2% of the A18 Pro.

Given that at this price Linux compatibility is an absolute lottery, would I sacrifice half the RAM for having an OS that isn’t Windows? Yeah there’s not much to think. W11 will probably eat half the RAM on telemetry alone, and Apple’s BS is easier to put up with than MicroSlop’s…

The LLM is whatever you want it to be. Self hosted or from any provider with a compatible endpoint. It’s likely a proprietary one… Because the cost of training LLMs means most are proprietary ones.