This is such a pertinent question. I actually think the right combination is a Mac Studio + UPS with ~256GB (to run your local local models with uptime) and then either a MBA or MBP with perhaps 24-32 GB (depends on whether you use Chrome 😛) as the daily driver.
Is there really a practical use case for a laptop with a ~200B+ parameter model?
MacBook power users:
What workloads actually justify more than 64GB RAM?
And when is an M5 Max worth choosing over an M5 Pro?
(Excluding running local AI models)
51
Ninad retweeted
An uncomfortable truth
AI will diffuse through society as a GPU per person
thesis.truffle.net/
Marketplaces provide trust. That’s the key piece of missing infrastructure if this agentic vision is to be realized.
marketplaces will collapse
first - marketplaces that are just aggregators like airbnb in the next 365 days
second - then marketplaces with some ops like delivery, stores like amazon, zomato in the next 1000 days
from first principles, marketplaces aggregated demand for their users, but if an agent like muse, @iradotso or @instinct can do the same without charging you 17% , why would that survive?
@amazon @Airbnb and likes are on a downward slope and it's a steep slope
51
Ninad retweeted
I am begging you, export your ChatGPT history this week.
You've spent years telling it about your projects, your health and how you write. All of that lives on OpenAI's servers.
If OpenAI locks, bans or loses your account, you lose those chats with it. You also can't move them to Claude, Gemini or a model on your own laptop.
In ChatGPT, open Settings → Data controls → Export data, and OpenAI emails you a zip of every chat. In Claude, you'll find the same button under Settings → Privacy.
Save the zip on your own drive. Then copy the parts that matter into one plain notes file, so any AI can read it, including the open source model you downloaded last weekend.
The export takes one click and a minute of your time.
Keep your AI memory on a drive you own.
Apple is not fumbling AI and will end up a big winner.
Although OpenAI/Anthropic and Apple have similar user numbers, their customer bases are orthogonal.
Apple has to ship products that “just work” and do so reliably. AI as a category isn’t there yet; it works in enterprise only after significant data and process engineering — how do make this work across a billion people that you can’t data/process engineer around?
Just ask Meta and Instinct how their products are going after the initial launch buzz.
Apple fumbling AI is crazy because they could have been the OS-level personal agent
Relying on browser use is limiting and (IMO) is why Uber, DoorDash etc. (app-first) are seeing basically no agent traffic
Probably gets fixed with connectors… and opens up opps for startups!
1
1
127
Apple is tantalizingly close to having all the ingredients to dominate consumer AI, but getting there will take time.
* Edge hardware — they’ve been seeding the market for years, and newer iPhones can run ~3B models, but you probably need another order of magnitude capability to support truly edge inference
* iCloud — if not at the edge, a Muse-like architecture with user-level private inference might work, but it is still cost-prohibitive and not business-model compliant
* Harnesses (OS, app stores) — the core infrastructure for context is in place through the platform, this is the least risky dimension
* Product — Apple has built ML into its products very thoughtfully. They clearly understand how AI must be built for large scale consumer products
I think we are about 3-5 years away from edge capability and cloud-inference economics converging in Apple’s favor.
Apple has another major AI gap to close: Computer Use. OpenAI and Anthropic already have agents that can operate software, click buttons and complete tasks on your behalf. Apple controls the Mac, iPhone and iPad, but still lacks a comparable capability. Why that needs to change-
43
If there’s an entire industry that exists just to make your product usable, then it’s probably not a good fit for vibe-coded apps.
Why is Vercel everyone's default choice for deployment these days?
What happened to AWS, Netlify, and traditional hosting?
Is Vercel actually that good, or are developers just choosing convenience?
24
This is important in spirit but absolutely the wrong approach in practice. Here’s why:
1. That 20GB model file means that practically speaking you’re looking at 48GB+ of RAM locally; fixed cost becomes a huge factor
2. That local inference isn’t cheap either; it costs money in the form of utility bills, the need for a UPS, and such
3. 20GB means sub-100B params (best case, really stretching it on luck) if you really pull out all the stops across quantization and any other techniques; raw capability will be significantly poorer, bordering on uselessness in many cases
4. You will have to build all the harnesses, or use an open-source harness; meaning you’re now on the hook for dev ops
If your use case is so generic that you can do it locally, then you should be able to just switch to another paid provider.
Now, the underlying point Alex is making is very valid — AI has the most centralizing tendency of any technology ever. And you must invest in it to truly benefit from it.
But a local model is not the thing to worry about or focus on.
Context is.
If you care about being sovereign, as an individual or as a business, then you should focus on controlling your own context. That’s what gives the chat apps power over you.
The model itself ultimately a commodity. Switch out Sol for Sonnet or GLM or whatever. The chat apps have become addictive because they know more about you with every use; this is Google Search’s flywheel on steroids.
Ultimately the ideal solution here is some sort of AI chat as a service; it will be more expensive to use, but the data will be yours and the model will be just something you switch.
I don’t know much about Truffle but I imagine this is sort of what @iamgingertrash is building.
I am begging you, download an open source AI model this weekend.
It's one file, 5 to 20GB, and it runs on the laptop you already own. You won't need an account, a subscription or an internet connection once it's on your drive.
Your paid AI can change overnight, cost more next year or lock you out. A model file you keep gives the same answers in 2036 that it gives today.
Grab Qwen or Gemma, install a free open-source app to run it, and ask it the questions you'd never type into a website.
Setup takes about 20 minutes.
You'll be glad it's done the day your account gets locked.
55
