Crafty's Illustrated

Like a normal blog, only with pictures. Insights on technology, fractional CTOing, and life in general.

John Gruber’s First Impressions of Vision Pro and VisionOS Gruber conveys the essence of the experience better than other writeups I’ve read so far. Unlike a traditional media outlet with their usual needs to drive readership, avoid pissing off Apple, appear balanced, etc., Gruber pretty much straight says what he thinks, which is useful. June 8, 2023 #

Is Runway Gen-2 Midjourney for Video?

We Tested Gen-2 and Share the Resulting Videos

Orion Browser by Kagi I’m late to the party, probably, but the Orion browser from Kagi (makers of the excellent paid search engine) is really interesting. WebKit-based, super lightweight, zero telemetry / auto ad blocking. And it supports both Firefox and Chrome browser extensions (or at least quite a few of them). Still considered beta but caught my attention for sure. The benchmarks are pretty wild.June 7, 2023 #

Faster sorting algorithms discovered using deep reinforcement learning So how about an AI story that isn’t generative AI. To quote the Nature abstract:

Whereas remarkable progress has been achieved in the past, making further improvements on the efficiency of these routines has proved challenging for both human scientists and computational approaches. Here we show how artificial intelligence can go beyond the current state of the art by discovering hitherto unknown routines. To realize this, we formulated the task of finding a better sorting routine as a single-player game. We then trained a new deep reinforcement learning agent, AlphaDev, to play this game. AlphaDev discovered small sorting algorithms from scratch that outperformed previously known human benchmarks.

Not earth-shattering, not an all-new fundamental sorting algorithm, but achieving a 1% to 5% improvement on code that runs a trillion times a day is meaningful. Already merged with LLVM so it will make its way into real code.June 7, 2023 #

Apple Vision Stratechery’s Ben Thompson with an excellent take on Vision Pro. He was able to get a few minutes hands-on with the device yesterday. His take aligns very much with my own. Pre-event, I kept thinking, “put AR/VR aside, this thing could be a really great replacement for external monitors” and Ben goes there and beyond as he looks at likely evolving uses of Vision products over time.

Well written and worth a read.June 6, 2023 #

Sloppy Apple Keynote Observations

Part 1: Everything Except

Keynote video: Apple blew out the audio levels on interstitial music. Very off brand.
New Mac Pro: I was very amused when, as Apple talked about this beast having 7 PCIe slots and mentioned what you’d use them for, they skipped one little item … GPUs. “Let’s ignore generative AI and it’ll go away.”"
iOS 17
- Stickers is a terrible name, but I think this will be a great feature–like iOS’s “photo subject select / cut” on steroids.
- StandBy: Like it, but for nighttime, does an all-red color scheme make you sleepy?
- Autocorrect is being upgraded to “2019 state of the art” level
iPadOS 17
- Health app on iPad. I’ve been wanting that.
- Notes app massive upgrades, lots of Notability-type upgrades but just like I’ll never buy another Ford, I’ll never trust my notes to the Notes app. Been burned by both Fords and Notes.
macOS Sonoma
- tvOS like screen savers look nice
- Not a gamer but what they’re doing with Game Mode resonates
- Presenter overlay features are interesting and looks like they’re not FaceTime only thank god.
- Loved the car control demo featuring a Lucid Air and not a Tesla. Screw you Elmo.
- Safari profiles and webapps look intriguing. Skeptical that they’ll actually get profiles right though.
AirPods
- Adaptive audio, a combination of transparency and noise reduction.
- Conversation awareness, detecting and augmenting speech. Getting into hearing aid territory.
- Better autoswitching. Currently that’s unusably bad; we’ll see.
Apple Watch
- Road biking wonderfulness, if it actually works. Bluetooth sensor support, FTP calculations, full-screen bike computer modes. I use Cyclemeter but if this works …
- Looks like great upgrades for hiking also.

There you go. Sloppy enough for you? June 5, 2023 #

Sloppy Apple Keynote Observations

Part 2: Vision Pro

Pre-keynote expectations are to Vision Pro reality as “front screen will show googly eyes” is to the actual EyeSight feature
This product shows the degree to which Apple is the anti-Meta
The heavyweight first-gen feel of this reminds me of the first iPad. A super solid chunk.
The theater experience–want.
23M pixels per eye in a postage-stamp-size display.
So so so many things they had to do to get this right. E.g., Zeiss lens options for those needing corrected vision. Reminds me of Apple Watch, but probably 20X longer list.
On-secure-enclave OpticID. Holy hell.
So so so many cameras and sensors.
Required an entire new OS flavor to make this work.
Required an entire new chip family (in addition to an M2) to make this work–12ms latency sensor to display. That’s insane but necessary. Meta will never, ever catch up with this.
5,000 patents. Seems about right.

I predict an Apple Watch like trajectory of success but way way bigger revenues. June 5, 2023 #

It’s infuriatingly hard to understand how closed models train on their input Simon Willison digs into the question of whether the big closed LLMs are training their models based on users’ input. As Willison says:

I’ve been wanting to write something reassuring about this issue for a while now. The problem is… I can’t do it. I don’t have the information I need to credibly declare these concerns unfounded, and the more I look into this the murkier it seems to get … The fundamental issue here is one of transparency. The builders of the big closed models—GPT-3, GPT-4, Google’s PaLM and PaLM 2, Anthropic’s Claude—refuse to tell us what’s in their training data.

Ambiguous language from these big players is the norm; unambiguous statements, such as OpenAI’s (which unfortunately applies only to paid API users and only from March 1st onward) is a rarity:

OpenAI does not use data submitted by customers via our API to train OpenAI models or improve OpenAI’s service offering.

This ambiguity around closed models’ training data makes open, self-hostable models–where there is zero ambiguity about, and total control of, training data–increasingly attractive to developers and companies building LLM-based products.

Definitely worth a read.
June 4, 2023 #

ChatGPT Code Review

I’m working with a fractional CTO client to launch a significant update of their web + mobile platform. The mobile app needs to support offline mapping, and our dev team has fighting through some problems there. My understanding of how map rendering and offline worked in the app was fairly high level, and given the current problems, I decided I needed to deepen my knowledge a bit. So I invoked my trusty intern ChatGPT (GPT-4 model) to help me walk through the app and understand the libraries being used, sources of map tiles, online and offline flows, etc. In about an hour, my understanding progressed from superficial to fairly deep, and potential causes of the problem we’re fighting much more clear. The conversation spanned over 46 pages and 10,000 words. This wasn’t a “saved time” situation–rather it was, “never would have done it alone.” A big win. June 3, 2023 #

A Day Late

Soon after I announced that I’m writing every day and why, I realized, uh oh … what happens if I miss a day? Well, that day arrived yesterday (June 2) when I had a series of unfortunate events trying to streamline my content publishing by doing Hugo site builds in via Cloudflare Pages rather than locally. I spent too many hours streamlining (or trying to!) and zero hours writing. Well, it’s a bit of a cheat, but I’m doing yesterday’s writing today, and I justify that because the subject is yesterday 😜. I’ve reverted to my previous local build workflow for now, and will continue to plug away on the Cloudflare issue while keeping daily writing as a higher priority. June 2, 2023 #