Given this trend, GPT 5 or 6 will be trained on a majority of content from its previous versions, modeling them instead of the expect full range of language. Researchers have already tested the outcome of a model-in-loop with pictures, it was not pretty.

Forewarning : ops here, I’m one of the few the bosses come to when the “quick code” in production goes sideways and the associated service goes down.

soapbox mode on

Pardon my french but that’s a connerie.

Poorly written code, however fast it has been delivered, will translate ultimately into a range of problems going from customer insatisfaction to complete service outage, a spectrum of issues far more damageable than a late arrival on the market. I’d add that “quick and dirty code” is never “quick and dirty code with relevant, automated, test coverage”, increasing the likelihood off aforementioned failures, the breadth of their impact and the difficulty to fix them.

Coincidentally , any news about yet another code-pissing LLM bothers me a tad, given that code-monkeys using such atrocities wouldn’t know poorly written code from a shopping list to begin with, thus will never be able to maintain the produced gibberish.

Subjective take: there’s worse than FreeCAD - sure it’s a bit “old school” but it’s bearable. O. The other hand, the solver has crashed on me so many times… The workbench way of doing things requires some time to get usdmed to, sure, but a crashing solver is far worse.

Secrets don’t belong anywhere inside an application code. They’re related to the runtime environnement - 'cause you don’t use the same password for production and integration, right? - and should come from an external configuration source. That might be as simple as environment variables.

Application deployment should never require modification of a file that resides inside the application itself. PHP and other interpreted languages has a tendancy to promote laziness when it comes to proper release management.

And don’t start with “but it makes development complicated”: fix your onboarding and then tooling instead of putting the security of your users and customers at risk.