Dave Plummer worked as a programmer at Microsoft in the 1990s. He wrote the Smart Drive CD-ROM cache and DISKCOPY. He also invented and wrote the first version of Task Manager. He just posted a video about his latest project:

Are you sick of windows getting just a little fatter every single year? Tired of apps that need an account, a cloud sync, and 30 background services just to open a damn text file? Well grab a seat, because today I’m actually doing something about it.

He built a modern version of notepad.exe with the same features as were present in the Windows XP-era version, and called it TinyRetroPad.

The final executable size is only 2.5 kb. That’s smaller than the default size for a single disk cluster in NTFS (i.e. the smallest unit of disk space that
can be allocated). It’s the same amount of disk space as the audio that goes along with a single frame of his video.

About a year ago, Andreessen Horowitz (the largest venture capital firm in the world), published a blog post showing that the cost of LLM inference is dropping 10x per year.

Andrej Karpathy, a member of the founding team at OpenAI and subsequently Tesla’s Sr. Director of AI, says that in 2019 it cost $43k to train GPT-2. Today he can train a model that performs similarly for $73. That’s 600x cheaper in 7 years. Not the same metric, but maybe more important.

In the last few months, a bunch of new AI browsers have launched, including Comet, Atlas, and Dia. I have no interest in these applications, they seem like a privacy nightmare, and I’m perfectly happy to use Claude and ChatGPT in a browser tab or in their native Mac apps.

Six years ago, the team at Kagi decided to build a new browser called Orion using the WebKit engine that powers Safari. They just released Orion 1.0 for Mac. It’s already available for iOS and iPadOS.

I’ve been an occasional Orion user for several years. It’s faster than Safari every time I measure it. Orion loads and runs Safari, Chrome, and Firefox extensions, including must haves like uBlock Origin and 1Password. Unlike Chrome, Kagi offers full support for extensions using both manifest v2 and v3. Orion has a robust AppleScript dictionary, and the user interface is excellent. You can tell a lot about how much a Mac dev team cares about their app by looking at the settings. Orion’s settings page is beautiful. This is a Mac-assed Mac app.

Orion has no telemetry and no AI built in, and that receives top marks from me. It’s not open source, but the development is done in the open with an active user community. Kagi is working to bring Orion to both Linux and Windows.

I’m going to give Orion a try as my daily driver, and I hope you do too.

In the olden days, every computer keyboard had all the alphanumeric keys, a row of function keys, an inverted T arrangement of arrow keys, edit keys (page up, home, end, etc), and a number pad. Today there are many different keyboard sizes, with various combinations of all of these types of keys. Keyboard nerds know what a 60% keyboard is. The rest of us call it a laptop keyboard. I use the editing keys a lot when typing at my desk with a full size keyboard, and I really miss them when using my laptop. Inspired by Brett Terpstra’s Home Row Arrow Cluster, I decided to see if I could figure out a solution.

I use a Macbook Air, so my approach is only useful for Mac users. Like Brett, I decided to use Karabiner-Elements, an outstanding low-level keyboard customizer. In Brett’s implementation, you hold down the ; with your pinky, and then use I J K L for the arrow keys.

That feels uncomfortable to me, and I wanted to be able to use P and ; for Page Up and Page Down. I wanted something where I could enter an edit key mode, then have the keys work as arrows and edit keys, then leave the mode and they go back to normal. Kind of like how the Num Lock key works. I can’t use Num Lock, because Apple laptops and keyboards don’t have it. In it’s place is the Clear key, which nobody uses because nobody knows what it does.

While experimenting with several keystroke options I discovered I was already using them for something else. Funny how your fingers know to type a thing, but your brain can’t remember all the things you know how to type. I finally settled on using ⌃+Delete to activate and de-activate edit mode.

Once in edit mode, I had to decide which keys to use for arrow keys and edit keys. I also experimented with this a lot. On my full size keyboard, I move my right hand over to use the arrow keys, so I wanted to use keys on the right hand side of the keyboard. I use my left hand for the modifier keys, so when I want to select backwards by word I use my left hand to press and hold ⌥+⇧ and then use my right hand to repeatedly press Left Arrow. I want to be able to do a similar thing on my laptop.

I ended up with the following keys:

I configured I J K L for the arrow keys, in a similar configuration to the gaming W A S D but on the right hand. Because I exclusively used Unix and Linux before ever using Windows or MacOS, I have Home and End move to the beginning and end of a line. U and O are logically positioned for those actions. I originally had H for Delete, but changed it to N so I can have Y and N for Page Up and Page Down.

To implement this in Karabiner, you need to create a new complex rule, and paste in this JSON:

{
    "description": "Control-Delete toggles 60% edit mode (I/J/K/L → ↑/←/↓/→, U/O → Home/End, N → Delete, Y/H → PageUp/PageDown)",
    "manipulators": [
        {
            "conditions": [
                {
                    "name": "sixty_edit_mode",
                    "type": "variable_if",
                    "value": 0
                }
            ],
            "from": {
                "key_code": "delete_or_backspace",
                "modifiers": { "mandatory": ["control"] }
            },
            "to": [
                {
                    "set_variable": {
                        "name": "sixty_edit_mode",
                        "value": 1
                    }
                }
            ],
            "type": "basic"
        },
        {
            "conditions": [
                {
                    "name": "sixty_edit_mode",
                    "type": "variable_if",
                    "value": 1
                }
            ],
            "from": {
                "key_code": "delete_or_backspace",
                "modifiers": { "mandatory": ["control"] }
            },
            "to": [
                {
                    "set_variable": {
                        "name": "sixty_edit_mode",
                        "value": 0
                    }
                }
            ],
            "type": "basic"
        },
        {
            "conditions": [
                {
                    "name": "sixty_edit_mode",
                    "type": "variable_if",
                    "value": 1
                }
            ],
            "from": {
                "key_code": "i",
                "modifiers": { "optional": ["any"] }
            },
            "to": [{ "key_code": "up_arrow" }],
            "type": "basic"
        },
        {
            "conditions": [
                {
                    "name": "sixty_edit_mode",
                    "type": "variable_if",
                    "value": 1
                }
            ],
            "from": {
                "key_code": "k",
                "modifiers": { "optional": ["any"] }
            },
            "to": [{ "key_code": "down_arrow" }],
            "type": "basic"
        },
        {
            "conditions": [
                {
                    "name": "sixty_edit_mode",
                    "type": "variable_if",
                    "value": 1
                }
            ],
            "from": {
                "key_code": "j",
                "modifiers": { "optional": ["any"] }
            },
            "to": [{ "key_code": "left_arrow" }],
            "type": "basic"
        },
        {
            "conditions": [
                {
                    "name": "sixty_edit_mode",
                    "type": "variable_if",
                    "value": 1
                }
            ],
            "from": {
                "key_code": "l",
                "modifiers": { "optional": ["any"] }
            },
            "to": [{ "key_code": "right_arrow" }],
            "type": "basic"
        },
        {
            "conditions": [
                {
                    "name": "sixty_edit_mode",
                    "type": "variable_if",
                    "value": 1
                }
            ],
            "from": {
                "key_code": "u",
                "modifiers": { "optional": ["any"] }
            },
            "to": [{ "key_code": "home" }],
            "type": "basic"
        },
        {
            "conditions": [
                {
                    "name": "sixty_edit_mode",
                    "type": "variable_if",
                    "value": 1
                }
            ],
            "from": {
                "key_code": "o",
                "modifiers": { "optional": ["any"] }
            },
            "to": [{ "key_code": "end" }],
            "type": "basic"
        },
        {
            "conditions": [
                {
                    "name": "sixty_edit_mode",
                    "type": "variable_if",
                    "value": 1
                }
            ],
            "from": {
                "key_code": "n",
                "modifiers": { "optional": ["any"] }
            },
            "to": [{ "key_code": "delete_forward" }],
            "type": "basic"
        },
        {
            "conditions": [
                {
                    "name": "sixty_edit_mode",
                    "type": "variable_if",
                    "value": 1
                }
            ],
            "from": {
                "key_code": "y",
                "modifiers": { "optional": ["any"] }
            },
            "to": [{ "key_code": "page_up" }],
            "type": "basic"
        },
        {
            "conditions": [
                {
                    "name": "sixty_edit_mode",
                    "type": "variable_if",
                    "value": 1
                }
            ],
            "from": {
                "key_code": "h",
                "modifiers": { "optional": ["any"] }
            },
            "to": [{ "key_code": "page_down" }],
            "type": "basic"
        }
    ]
}

If these keys don’t work for you, and you are handy with JSON, you can probably figure out how to edit this.

If you aren’t handy with JSON or aren’t a Karabiner expert, paste the code above into your favorite chatbot, tell it this is a Karabiner-Elements complex rule, and use plain english to describe what you want to change. It will do a pretty good job of making the changes for you.

Anthropic bought paper copies of books, scanned them, and used the resulting data to train their large langugage models. Authors sued, claiming this was not fair use under copyright law, and got their case certified as a class action.

The case was assigned to District Judge William Alsup, a seasoned US district court judge with deep experience in technology law, including two jury trials in the Google v Oracle saga over the Java APIs in Android. Last month Judge Alsup ruled that it’s fair use for legally obtained books to be used to train large language models. That’s a big win for Anthropic and the other frontier model companies.

However, Anthropic is also accused of pirating digital copies of more than 7 million books, and these illegally acquired works are not part of the fair use ruling. Judge Alsup also ruled that Anthropic must face a trial over the author’s piracy complaints.

US Copyright law includes statutory damages of up to $150,000 for willful copyright infringement of a single work. With 7 million works in question, Anthropic faces potential damages of up to a trillion dollars.

In their first major filing since Judge Alsup’s ruling, Anthropic is appealing the certification of the class. It’s their best tactic to reduce their financial exposure. Ashley Belanger summarizes Anthropic’s appeal for Ars Technica:

Confronted with such extreme potential damages, Anthropic may lose its rights to raise valid defenses of its AI training, deciding it would be more prudent to settle, the company argued. And that could set an alarming precedent, considering all the other lawsuits generative AI (GenAI) companies face over training on copyrighted materials, Anthropic argued.

No, Anthropic won’t lose any rights if they decide to settle. They made their choice when they willfully decided to pirate copyrighted works and hoped nobody would notice. Now that everyone knows, they may choose to settle and avoid a trial because if they lose at trial it’s a company ending event. Yes, the precendent would be ground-breaking, but it’s only alarming if you are infringing author’s rights. If other frontier model companies have done the same thing, they should have to face lawsuits as well.

Copyright class action lawsuits are incredibly complex, the case against Anthropic will take a long time, and the outcome is far from certain. Authors Guild v. Google took 10 years for the appeals court to finally rule that Google’s Books project of scanning copyrighted works and publishing them on the internet “provides a public service without violating intellectual property law”. It was appealed to the Supreme Court, which declined to hear the case.

I’ll be following Authors Guild v. Anthropic with great interest, even if the certification of the class is overturned. If the individual authors win against Anthropic, the floodgates will be open for claimants to sue every frontier model company. We already know that Meta pirated 7.5 books and 81 million research papers. If there is a precedent set in Authors Guild v. Anthropic, I’ll bet it will only take one phone call to find a lawyer who will help you sue Meta, OpenAI, and Google on the same grounds. A trillion dollars of damages from Anthropic, another trillion from Meta, and pretty soon you are talking real money.