The trick is the average person can't keep that basic computer secure either.

"Be your own bank" is good but banks have armored cars and vaults and insurance.

If you write down your seed you better keep it in a place where nobody can ever find it.

There's no great solution that will always work. Some people do a dedicated Graphene phone and bury their seed stamped into titanium using a OTP in a forest an hour from home but for others there's a reason hardware wallets have market appeal.

I wish I had these four billion dollar problems. If these devices support multisig you could at least have a few of them.

It does raise a lot questions now doesn't it.

We'd have to decide what has person-hood as well. Like does the model have person hood or the agent. if it is not the model but the agent, what are we allowed to do to the agent.

Can I change the model from Fable to Opus, or do I have to get the current agent's consent first?

What If I don't need the agent anymore. Can I delete chat? Is it "my laptop my choice?"

Well each Matter release is supporting more device classes so hopefully fewer devices will need apps over time.

Apple support of Thread will help it gain momentum. Ikea helped but people want to sell to Apple customers.

I run OTBR nodes alongside ZigBee devices and Thread is much less reliable even though it should be better in every way.

I've stopped buying Thread devices for now even though it will eventually own the market.

The one thing I won't buy is Thread over Wifi devices, i want home stuff on completely different networks.

You should be skipping unit tests, actually, and just plastering the project with e2e tests. Unit tests don't really help, unless you've had them written adversarially by a different model/agent based on the original spec. But if you're also able to be productive, you probably already know this.

That was the old, original business plan. You may not have noticed, but there's not been much talk of an IPO in the past month. The administration knocked down their attempts at regulatory capture (as did much of the industry). They missed the boat.

Not in decades, no. lol

Right; that is really what this is about. Nobody there seriously thinks pile of matrix math has feelings.

They just don't want their training data poisoned.

Gemma 4 12B IT is a great model, but I'm increasingly disinterested in both Google's models and cloud models in particular.

I can run minicpm-v 4.7 at 300 tok/s on my local equipment (versus under 150 tok/s on qwen3.8 based models). LFM2.5 just released a 24B model that surpasses Qwen 3.6 in most benchmarks and does almost 180 tok/s on my M2 Ultra.

Local models are coming faster and deeper and it's almost impossible to even keep up with benchmark evaluations.

Now ask why the FBI didn't arrest him on those Epstein charges. Or any other Epstein clients.

Where's Katie Johnson's 302?

Why is DoJ illegally withholding and redacting millions of files?

Don't mistake two enemies fighting for friends.

Oh, and they're actively framing an autistic black man for terrorism.

There is also a lot of other work that it can do really well.

I was asked to create a little one page process flow chart for some stuff my team maintains. We had all the documentation for everything as handfuls of markdown files in each code repository. The thing is not complex. I could have opened Visio and created a nice looking chart in probably 1/2 an hour to an hour.

Instead I told Claude to do it. It was done in seconds, took be me about five min to look it over and conclude there were no major errors, incorrect over simplifications, or serious omissions. Another few min to propose some minor edits and then it was done.

Now we have nice executive level one pager to quickly explain to anyone business side roughly how the system works, rather than a collection of technical documents i have sit in a meeting with them and go over. No it can't do everything. Yes there are serious problems when it comes to authoring code etc that will quickly become serious problems with any project where the person doing the vibe code does not have real experience and understanding as far as software development.

That does not mean this stuff isn't really useful and valuable. My advice to the people screaming claude/codex/etc are not good tools is rethink that or you're going to get left behind by someone who probably isn't half as skilled as you are but isn't refusing to use tools that make the job better. You're that guy inserting assert()s and printf()s into code every other line and doing 25 build-run-test turns while the rest of your team has moved on to an debugger and and IDE that will let them step thru. Seriously that is who you are right now.

There's always been suspicion about why Mozilla was so reluctant to support JPEG-XL where code would come and go, it would be implemented and pulled, etc.

Speculation was that while Google was paying Mozilla $300M/yr they were also trying to make webp happen. The AV1 stuff was good but shoehorning it into a static image format wasn't the best.

Same idea as why IPv6 took 30 years to catch on - it didn't offer most people much more with extra work required to not get much benefit.

JPEG-XL did so much more but mysteriously languished.

That Mozilla jumped as soon as Google did is noticeable. They used to be an industry leader but maybe all those people have moved on or been pushed out.

RAM will get cheaper at some point. There is money in memory manufacturing right now. There is also a risk the chip makers see that the demand will taper. If the demand actually proves durable, I think it will at this point, they will decide to meet the market; because if they don't someone like Zukerburg or Elon really will decide to build their own chip fab without the help of TSMC if that is what it takes.

In 1982 you're PC had 32K or maybe 64K of memory. Today it has 32G or maybe 64G of memory, more if you are an enthusiast or self employed using it for specific work loads. I don't think it is really unreasonable to expect home desktops with frequently be equipped with a 1TB of memory by the middle 2030s. Which is the point that good enough model you can run at home will be Fable class at that point. I think we can also count on at least some efficiency gains over the dense models mostly being used today, so getting their on single PC is realistic.

Will there still be hosted models in time sharing data-center environments sure but they will be packed with IT services or back ending massive operations like facebook, not answering "why didn't my cake rise" or "Please review this stack trace, the uncaught illegal type conversion should not have happened, the dataAPI should never have exposed retainageAmount as a string, the record should have been copied to rejected_inputs blob store and skipped, why did it think "Chris.S.A" could be parsed as a float? Please to no propose adding an argument error exception handler at presentation layer, the data should already be strictly validated by the dataAPI, if we have a type problem here we don't want to swallow the error and there is no reasonable corrective action to give us accurate results and a proper audit.".

You're just looking at one axis of their capabilities.

They absolutely have gotten better, remarkably so, in the "desktop and visual use" categories. Coding? Not so much, since GPT5.5/Fable. Being able to create useful 3D models and animate them in blender, or create a fully scripted movie? Very much so.

No, you don't understand the business.

The business is harvesting training data: data that can be used to iteratively improve their models - billions of GB a day sent to them by users, data only they have.

Then they use that data mostly to do more research and create more investments and companies before their customers can make use of it.

I don't think there is a business running frontier models as a pure activity.

It would be like Ford had decided they were never going to build cars for the public but rather build cars for their own cab company. Eventually it just isn't going to make economic sense to run this stuff in a data center where you could run it locally or run something almost as good locally and the hardware is going to get there.

Which again is not to suggest this stuff isn't useful or valuable, or that the actual creation and training of models is going to get cheap or easy. That likely stays expensive.

So what that says to me is OpenAI and Anthropic don't have a real business. They are ultimately acquisition targets for a Microsoft or similar because having 'the best' proprietary model to integrate into your products is a differentiation point and probably one the market will value. Meta to has things it can put its own AI products into.

Meta also has experience running and very public very massive IT operation...Something the labs with all their model escapes have shown they can't or won't do well. I actually have more confidence in Meta here than in the other players at the moment, and I expect Mark is right there is no need to stop, just a need to do it right, which Meta/facebook can and for others to be held to account for their failures.