Last Week Off My Mac: A1 and AI

So much water has gone rushing under the bridge this week that I have barely registered that WWDC is already over for another year. The huge political events which have been in progress through the week have all but obscured the excellent news that Apple is at last shipping new Mac models, with even better coming in the pipeline. The iMac and promised iMac Pro at last rectify an extraordinary oversight: Macs do have a solid future.

Apple has also been very wise in its approach to macOS 10.13 High Sierra. The big change is the introduction of its new file system, APFS, and there are valuable advances in support for virtual and augmented reality (VR, AR), and for artificial intelligence (AI), which I will return to later.

Many of our early concerns about APFS have been allayed, and its implementation in iOS 11 is moving closer to HFS+ in its approach to handling filenames with normalisation. This is not to say that upgrading to High Sierra is going to be trouble-free for everyone, but Apple is continuing to work to minimise its impact without imposing penalties on its features and performance.

High Sierra’s log

Unlike WWDC 2016, this year Apple’s engineers did not single out the new unified log system for special coverage. The implication is that High Sierra won’t bring any particular changes in that, nor any wonderful new tools to work with it. For some, that will be a big disappointment; for me it is a challenge to keep developing Consolation and other log-based tools to fill the gap.

It appears that Apple considers the log is the preserve of its engineers, with their in-house tools, and developers alone. One strong signal of that was the unannounced removal of log access when in normal user mode. Most peculiar about this is the way that it has been implemented: for example, try using the log show command in normal user mode, and you are not warned that it requires elevated privileges, nor does it need to be run as root, the usual reason for invoking sudo. Instead, it just returns empty results. It doesn’t even have the decency to behave as a normal Unix tool should.

There’s nothing wrong with that behaviour if normal users shouldn’t be peeking in their log. Apple’s engineers and other developers know better, and can get what they want.

None of this would be important if macOS ran faultlessly, and any glitches or problems were reported fully to ordinary users. We know from long experience that this simply isn’t the case. There are lots of vital features in macOS – keychains, Time Machine, updates, and more – which continue to cause problems. In many cases, the errors and clues to solutions are to be found in the log. Making it impractical to access the log for such information prevents the user from self-help, and makes life much harder for system administrators, forensic analysts, and many more.


The inclusion of support for AI in High Sierra is very progressive, but courts problems.

Doing AI properly is not a simple task. Developing AI solutions classically consists of three main phases, once the problem has been framed and a promising methodology has been selected:

  1. The chosen neural network or other solution is trained using example datasets, until a predetermined level of fit or accuracy is obtained.
  2. The trained network or solution is then tested using completely separate datasets to determine its efficacy; if that attains an acceptable predetermined level, then the system is accepted into production. If it does not attain an acceptable level, then the methodology is reconsidered and the whole process starts again.
  3. Once accepted into production, monitoring checks that efficacy remains within acceptable levels (quality control).

A lot of current ‘consumer’ AI systems do not pass through this full sequence. They perform well during stage 1 (training), but are never properly subjected to stage 2, and released without further monitoring in stage 3. We then end up with the sort of experience that many of us have with current ‘consumer’ machine translation systems: they perform fairly well on easy tasks, but once the grammar or vocabulary starts to get more tricky, they quickly produce results which are comical and of little practical use.

AI has a great deal to offer, but half-baked AI does everyone a disservice, and risks bringing the whole discipline into disrepute. The danger with providing easy access to just stage 1, as macOS High Sierra does, is that the other stages will be skipped, and that inappropriate models and neural networks will be used.

This also has more general, even political, ramifications. Computer-illiterate politicians have come to think that, just because there have been some impressive achievements using AI (many of which don’t actually use AI at all!), AI can solve anything – including detecting ‘hate speech’ in social media, ‘adult content’ in movies, and more.

Providing access to a limited range of AI methods in High Sierra is therefore a bit of a mixed blessing. May they always be used wisely.