mianos

I noticed this. What it seems to be is a very aggressive pruning of the context that seems to be done the more you use it. It's quite possibly something they are doing as the model gets used by more people to manage shared resources.

It gets so stupid it's frustratingly useless when it does not follow instructions give just a few prompts back.

I am not sure why I am still arguing, but maybe if I explain it another way, you will understand why what you are saying isn't true.

When you build an LLM, you base it on an immense knowledge base that you didn't create yourself.

It processes that information and, in simplistic terms, constructs its own underlying logic based on what it read. It's once removed from your direct instructions. If you think a human can manually edit or curate all that input content, go for it.

The developer who built the application isn't the author of the output, it is authored by the billions of data points it was trained on. Its choices and biases stem from that training data. No human is writing those billions of logic statements. The LLM is generating its own responses from the input knowledge, not the developer.

This is not how it works. LLMs at large scale have emergent properties. They are not just a box of 'if-then-else' statements. That was the major revelation when they scaled them up. They are doing beyond what that naive view of 'if-then-else' expects.

As dense as MacBooks are, it is nearly trivial to fit a spying device inside them, especially compared to small hardware wallets like Ledger Nanos. As secure as macOS is, once a device is inside the chassis, there are numerous ways to read data. USB serial is easily decodable, as are the many SPI, I2C, and other buses inside the machine.

Keyboard traffic is plain. Bluetooth access depends on where you are positioned. DVI is not hard to decode for capturing the entire screen. KVMs are so small now that one could easily be snuck in. Even though the motherboard is very compact, there is space above it and near the batteries.

There is not a lot of floating point in an operating system.

The issue with floating point and aliasing preventing vectorisation was from the late 1980s when early C compilers lacked sophisticated alias analysis and standards were loose. is not a really a thing anymore. C can go as fast, specially compiled with strict aliasing. Maybe more work in compilation.

> He's pretty damn smart that guy. This is probably the understatement of the year. I am literally ROFLing.

I think Ken Shirriff, CuriousMarc and friends need a new project to step it up now they have their Apollo systems all running like new. :)

As an Australian, I'd rather them spend the time and money fixing their insecure web sites than establishing a 'task force'. Because I can't see much useful stuff coming from herding a bunch of monkeys into a room and letting them just screech at each other, as monkeys do.

The writes where to get past cloudflare.

It's full RRP for everything less than 1000 units. If you want a 100 unit refresh you are better off going to your local corner shop for price and service.

But you can bend the pins of the power cables and make them fit. :)

Having spent a lot of time in the UK, EU and US, I feel Sydney is way more like California than say London or European cities, in terms of living, shopping and culture. Even things from the US are coming here, like the growth of Halloween.

If you walk into a shop, the US is much more like Australia in terms of service and attitude (even though retail staff are paid vastly differently).

I always thought this was due to TV and Australia having remote branches of many US retail chains.

Outside of California, it's similar but diverges a lot in areas like socialism. In the recent 20 years Australia has pretty much become more like the EU in terms of socialism.

I agree. Lot's of people really like it. For me it often just forgets all context and starts showing random slop. It's super clear as, when I ask it what happened to some element earlier in the conversation it tells me it does not have that. It might be good if it told me, but randomly lose the plot is quite frustrating.

Claude does it occasionally but it's a more a soft landing earlier context seems to be compacted, not completely lose the plot.

I just cancelled my pro subscription. I really wanted it to be good but not yet.

It's not 'whataboutism'. It's an identical situation, and the statistics relate to the same events.

But they don't. Thousands of people are being killed at intersections, yet unless there is some proven intent, it's extremely rare to go to jail.

Maybe this might be valid if this was a realistic thing and every bad driver who pulls out of a side street and kills someone suffers serious consequences, but they don't and it's not.

I have a microscope, but once you get used to using hot air, you don't need it for soldering, just inspection.

I have a pro subscription, I think they have just given up. Likely because when they test their new models against the other frontier models they are so bad, they just pull it back. This leads them to try and innovate in other areas where there is currently less competition so they can compete. Not a bad play.

There is no right to refusal at border control. It's outside the 4th amendment so it's not even relevant.

They do reach out pretty quick when your usage takes a huge dive after you move to one of the single provider plans that offer hard to believe possibly unsustainable value.

The flaw in 'calling attention to their methods' is that you rarely have a meaningful opportunity to argue against a bureaucratic system that operates this way.

Whether individuals inside the system are true believers or cynical operators, the response to calling out their methods will be identical:

If they genuinely believe the rhetoric: They will simply double down and say, 'Yes, and protecting the children makes this absolutely necessary.'

If they know it is a manipulation tactic: They will still reply with, 'It is for the children,' because they know it is an impenetrable PR shield that gets them their way.

In either scenario, the bureaucracy is entirely immune to the criticism.

Exposing the game might help persuade third-party observers, but it does absolutely nothing to stop the bureaucratic machine from using that exact same line to steamroll the agenda forward.

> Anyone who brings up kids is trying to manipulate you into giving up your freedom for security.

The immediate reply is the accusational: "why are you trying to trying to prevent safety for kids".

It's a simple next step to accuse you of hiding your own child abuse ring.

This sounds like hype, but it's the normal path here in Australian.

I just asked it to do a summary or a status page for my staff so I can share the high level details. Somehow it retrieved a version of the page from two weeks ago and summarised that. When I said, this is old, it happily said, yep, I'll do a summary of the current page. I kinda feel there may be people in between the box and some lame LLM.

With the explosion in token usage, why not write it in z80 assembly and run it in a z80 emulator written in webasm?

There is still a place for 'but rust’, in between too much and too little LLM usage.

There are also a lot of mcp agents for Joplin: https://discourse.joplinapp.org/t/ai-agents-and-joplin/46168 Dare I say, another, more popular, note taking app for you and your agents.

I really like it as I run a webdav service and can read/write and sync notes between multiple Mac desktops and Android. Encrypted while I am at it.

Markdown native.

Also, highly related, in the previous paragraph, they say: "We know a lot more about the biology of disease - although God knows, not nearly enough".

Anyone smarter than a 10 year old would not we actually know almost everything about car design. How it works is not at all hidden. We have iterated on pretty much the same thing for 100 years. The current state is about efficiency, materials and manufacturing.

When you have deep knowledge and experience in a field you will get near 100%.

Once I had more than a few esp32 sensors about the place, with my own firmware, I got serious about getting the OTA updates working and a good c++ wrapper around the nvram management. I don't like the vibe of multiple sensors not having the fixed firmware, just OCD. Github mianos.

Same, I often wondered why the openai logo looks like an interpretation of goatse.

I run zfs as the storage pool for my incus (next lxd) services. It is the ideal fit. Here is a list:

- Instant, zero-copy container cloning from images via Copy-on-Write. If you boot a new image like the existing ones it's seconds.

- Atomic, millisecond-level instance snapshots regardless of storage size

- Block-level container migration using native 'zfs send' and 'zfs receive', very short command lines and seems to work perfectly.

- Granular dataset nesting (every instance, image, and custom volume gets its own ZFS dataset). You can see every filesystem even on the host.

- Transparent, inline data compression (LZ4/ZSTD) enabled automatically per dataset. For services that don't change much, you might as well use a compressed image to make them even smaller.

- Mirroring / Raid

- Sub-volume sharing and direct management via native ZFS administration tools. If my home directory has a build area and a million files, I can just save time and put my home, pre cooked into a new machine and not copy or even rebuild on my new machine.

- Dedup keeps blocks with the same data as a reference. This costs a lot of memory and has not saved much for me as a lot of my images are similar and already shared I think, but it's cool.

I have a 15 year old Mercedes. I got the old integrated CD player and radio pulled out and Alpine makes a nice dash surround and canbus support module for the steering wheel controls. It has a 7" display where the old stuff was and wireless Android Auto.

It's pretty much like any new car as far as entertainment and maps. Plus, it's got physical buttons for everything else. I think recent cars may well be coming back to what I have on my old car.

ps. not mentioned, Anrdoid auto and Carplay are H264, later H265, video players with touch control and buttons in a return channel. The phones render the display internally and project it as video to the display. That's all. This is why much older generation displays with at least reasonable SOCs can run them near perfectly. They are not running Anrdoid themselves. In fact some of them are operating system-less SOCs that have nearly no firmware but a video player and the support for the touchscreen and buttons. Much like that crappy video player built into old TVs ten years ago.

I have them on my motorcycles too. It's way better than a phone, in that the user interface is designed to be less interactive and more focusses with larger buttons.