antonis-gr

As much as Russell's teapot. I mean the teapot itself, not the analogy. Non zero for philosophers, priests or cognitive scientists, practically zero for the general public while conversations with AI always start with a human asking the AI and never the other way around.

20/month buy enough intelligence to build comprehensive plans for dirt cheap models to execute. My observation is that unambiguous specs create high quality code regardless of model, and cheap models are more obedient than SOTA.

I may be wrong, I just describe what works for me. Exhaustive documentation and frequent checkpoints with SOTA on subscription, and execution with cheap Chinese models on openrouter. It is not easy to overspend.

Considering a human body consists of 10^28 atoms, it is 15 orders of magnitude different. The emerging properties of 10^23 parameters may be much better defined without needing to know the properties of each parameter.The theology of such a big llm would be interesting. Current ones, not so much.

Once I asked ChatGPT "it takes 9 months for a woman to make one baby. How long does it take 9 women to make one baby?". The response was "it takes 1 month".

I guess it gives the correct answer now. I also guess that these silly mistakes are patched and these patches compensate for the lack of a comprehensive world model.

These "trap" questions dont prove that the model is silly. They only prove that the user is a smartass. I asked the question about pregnancy only to to show a friend that his opinion that LLMs have phd level intelligence is naive and anthropomorphic. LLMs are great tools regardless of their ability to understand the physical reality. I don't expect my wrenches to solve puzzles or show emotions.

42 is an extremely non-special number. Does anyone know if it appeared in the CS field before Douglas Addams "invented" it?

No provider is better than two providers.

Hosting on second- or even third-tier providers allows you to overprovision and have much better redundancy, provided your solution is architected from the ground up in a vendor agnostic way. Hetzner is dirt cheap, and there are countless cheap and reliable providers spread around the globe (Europe in my case) to host a fleet of stateless containers that never fail simultaneously.

Stateful services are much more difficult, but replication and failover is not rocket science. 30 minutes of downtime or 30 seconds of data loss rarely kill businesses. On the contrary, unrealistic RTOs and RPOs are, in my experience, more dangerous, either as increased complexity or as vendor lock-in.

Customers don't expect 100% availability and noone offers such SLAs. But for most businesses, 99.95% is perfecty acceplable, and it is not difficult to have less than 4h/year of downtime.