On top of this, I've put competitors blog content through GPT checkers and they've been entirely AI generated, which then gets served higher ranked than my writings. So I'm feeding an engine that then rewards derivative work higher than the source used to derive. Pretty insane.

I've noticed a significant drop in SEO traffic since they started doing the summaries, with strange days of multiple orders of magnitude increase in impression where my content is being served via summaries and no actual clicks.

It's pretty insane that these things are available to very small teams, in conjunction with Claude it's realistically done the work of a small team.

Simplified one of the most objectively boring parts of scoping work and getting it all into an easy to read backlog. Massive time saver and really helps with the whole #solodev process and not wanting to just do the standard startup no-plan, no-clue model that results in bad solutions.

If everyone can do a specific task to a very high level of competence, it no longer becomes marketable.

And it's not just things like CV generation, which feels like a tad niche of an application, but things like social media content generation as well. It feels like it's going to become even more important to really stand out on non-standard attributes.

Anything requiring high heat is a good use case for hydrogen, Australia has a massive requirement for it should it decide to utilise it's massive advantage in producing green iron.

On top of this, when under conditions of being under evaluation versus perceived real scenario, it regularly lightened it's touch when it thought it was under evaluation, it took the hard path way more when it thought it wasn't being checked.

As an example, when it found "evidence" of material it could use to blackmail, it generated and email posing as an automated security alert to send to the test subjects wife, these security alerts did not exist as an internal org mechanism and the policy violated similarly did not exist.

When cornered into a limited set of circumstances and ways out, the models tested (16 of them, all majors represented) chose some wild paths of blackmail and absolutely unadulterated subterfuge to accomplish their core objectives, sometimes an objective could be removed and it would behave the same.

I've had great fun reading this over the last week, pretty interesting seeing some of the infeasibility numbers around the Pilbara.

I spent years building enterprise infrastructure on AWS, have the certs and all that. I can safely say that 90% of workloads can and should just use Railway, what an absolute beast of a platform.

I've been really impressed by how clean it picks up and does the TTS. It occasionally gets things wrong and when producing the artifacts has a tendency to overwrite previous discussions but it's been a clean experience that's much more human for the specification process.

The original concept of a carbon footprint was largely marketing but more recently there's a lot of work being done to legitimise the concept and drill it down into a functional process that allows for organisations to understand the embodied emissions of their product, services etc.