I just got Apple Intelligence on my MacBook, and I’ve spent today playing around with the summarisation, rewriting, and proofreading features.
My hot take: It’s fine.
I’ve had all these features and more running via my launcher, Alfred, for months now, so experiencing them doesn’t feel as magical as when I got LLaMA 3.1 8B running locally on my machine back in July. The model does a decent job—it’s not bleeding-edge for me, but for many people using it for the first time, it absolutely will be.
What really strikes me is how quickly my expectations have shifted. Summarise, proofread, rewrite—these tools just sit alongside spell check and grammar suggestions. It’s 2024: of course my computer can summarise a chunk of text, make a list, or rewrite a paragraph in a different tone.
And every time I use this feature, somewhere behind the scenes a cached 12GB LLM Isekais into existence, gets handed some text, processes it, and then blinks out again. No uncanny chatbot negotiation. It just works.
And that’s today’s takeaway-how mundane it all is already. This brand-new technology is becoming invisible, baked into workflows so seamlessly we stop noticing them. This shift feels inevitable. Apple Intelligence is already the boring AI future.
Going into 2025 AI is advancing in two distinct directions. On one side, frontier models—GPT-4, Claude, and whatever comes next—will continue adding features, growing more capable, and more expensive. We’ll also see more specialist models, smaller systems, being hyper-optimised for specific tasks, rather than the general purpose models we have now. These big and expensive frontier cloud systems will evolve into clusters of intelligences, with different kinds of functions and capabilities based on what you’re doing.
OpenAI’s $200-a-month Pro subscription doesn’t actually seem that unreasonable for the kind of capabilities described above. $200 bucks a month is way cheaper than asking an intern to do something for you (like read a book) a month ago. And these ‘pro’ tools will keep getting better.
But as I wrote back in July: The Only Acceptable Price Point Is Free.
Which is the other direction things are going. Away from the flashy announcements and innovative cloud products like MidJourney’s Patchwork which wrote abut yesterday, local models are quietly making big progress.
For everyone wondering about how much value $20 a month for GPT/Claude can truly provide, someone else is running a good enough model locally. I have Apple intelligence AND a Llama 8B nano variant running on my machine right now.
Adobe is making a big deal about its generative tool Firefly, but Krita’s seamless integration with Stable Diffusion is super fast, super performant, and all running locally.
This is just the beginning. With the imminent release of Mat and Holly’s Public Diffusion model from Spawning, people mad about copyright (or data centre energy consumption) are going to have to find something else to get mad about.
Also, Meta’s new 3.3 70B model has ChatGPT 4 level intelligence means that it’s March 2023 in local model land – which wasn’t that long ago. Things are getting optimised fast. True, it’s a little too big to run on my machine (it needs 38GB of RAM), but it’s plenty small enough to be running on a souped-up Mac Mini somewhere in the corner of an office, shared by everyone, integrated into your OS, for free, forever.
Whilst a lot of focus this week has been on the release of Google’s video model Veo, Gemini Flash 2.0 was also just announced. Variants of its mini model are just months away from running on our phones, handling not just grammar and summarisation, but maybe, if and when APIs are made available at the OS level, even stranger things.
The only thing these smaller nano models require is battery power. They aren’t burning down the rainforest, using the power of a small city, and the water of a small country. They are just lil’ guys in your phone and everyone is going to end up with devices that have ‘Intelligence Inside’.
We are already in the era of intelligence is too cheap to meter. The real question is what are we going to use it for?
Little Computer People, which I’ve written about a lot, Tamagotchi-like agents, onboard inside our devices. I find the idea very compelling.
These tools may no longer feel as magical or news as the once did, but it’s still a weird and exciting time.
Subscribing to SSRZ supports my online work and creative projects.
As a thank you, I send you my zine four times a year, just like it’s 1994.
No spam. No email. Cancel at any time.
Leave a Reply