"Training frontier models is a bit like alchemy:"
It's a load of old bollocks, and it doesn't work.
1041 publicly visible posts • joined 12 Jun 2009
"Turner said the reasons tend to fall into one of two categories. First, there are the mission-critical applications that can never be offline for any reason. Upgrading is, therefore, highly problematic and likely to be a business risk."
The third reason: because the new version (e.g., maybe, version 8.0 does not deliver like-for-like performance compared to version 5.7) is worse may be more common than both of them but together.
"When your application database version is multiple updates behind the rest of your systems, this adds to the workload and increases the challenge."
That's not true, if you can upgrade straight from version 5 to 8 then doing it in one go rather doing upgrades from 5 to 6, 6 to 7 and 7 to 8 a year apart reduces workload.
If you can't go straight from versions 5 to 8 then you are better off if the team doing the 6->7 and 7->8 upgrades have recent experience of doing the 5->6 upgrade
"added a click box to our AI platform"
This is not a sufficient response to misleading the court. If the person who wrote this was a lawyer, then they should no longer be.
"With a repentant heart, I sincerely apologize to this court, to my firm, and colleagues representing defendants for this mistake"
This is more the sort of thing I would hope for
This seems to be an example of Betteridge's law of headlines - Any headline that ends in a question mark can be answered by the word no.
"These are big claims. What lies behind them?"
"behind them?" could be replaced be a full stop.
"domain experts arguing about what the right answer should be"
The main issues aren't disagreements between experts, it's AI spurting out absolute horseshit.
"While putting control over receiving the summaries into the hands of users is helpful, it would be better for Apple to make this an opt in feature until the issues have been ironed out."
The summaries are so egregiously wrong it would be better for Apple to completely disable the feature until it can guarantee a reasonable level of accuracy.
I suspect the time that they can make that guarantee is 'never'.
" free/cheap models have no control over "temperature", and it's never zero. I have no idea why"
To encourage confirmation answers. If you run a generative AI 10 times, then its more likely one is accidentally right, and that is the one that lodges in the brain.
I recently had a case where the AI summary gave (what turned out to be) the right answer when there was a typo in my query, which it corrected to the wrong answer when I fixed the typo.
Unfortunately, we're no closer to rightly apprehending Babbage's confusion of ideas, because the AI is a black box.
"If insurers refused to cover not just business losses from wonky security, but also didn't extend cover at all if standards could not be shown to be in place."
If a couple of the larger insurers did this, then companies would insure themselves with the others.
Offering cheaper insurance to companies with standards in place could work. Or offer cheap insurance, with the option to add an expensive schedule for covering terrible security practices.