It's like Intel i3/i5/i7/i9, or nvidia 50/60/70/80.
It is meant to give a rating inside a generation. Different generations are not necessarily comparable. an i7 is better than an i5, but not if you compare an i7 7xxx with an i5 14xxx.
A nvidia xx80 card is better than a xx60 card, but not if you compare a 1080 to a 5060.
The only difference for OpenAI is that not every generation has every model.
> While employed, you’re using corporate accounts and can’t be banned.
Google/openai/etc detect that you are using your corporate account from the same home where you use your banned account, corporate account get banned for trying to circumvent the ban.
You'd have a point if those countries had followed Europe in replacing coal with imported gas, but few countries outside of Europe have done that. Without imported gas, electricity is generally cheap.
China burns coal and has still had time-of-use pricing since 1999.
The price of electricity is a function of demand with respect to supply. Even if the fuel is cheap, heavy demand will cost more to supply than low demand.
The UK largely did - there was a "dash for gas" throughout the 1990s, a response to the decline of UK-mined coal and the availability of North Sea gas.
Coal is no longer used, and the remaining mothballed coal power stations were demolished a couple of years ago. North Sea gas is now in decline, and is rapidly being replaced by renewables for electricity generation but we'll still rely on (expensive!) LNG imports to make up the difference until the end of the decade.
It did not. Germany's percentage of gas use for electricity stayed mostly flat. In absolutely numbers it went down because of less electricity use. Energy Charts from Fraunhofer publishes this data [1].
It already has made them poorer and miserable. The Draghi report was all about how they were poorer due to these things. Of course they decided not to follow the recommendations.
I like Bun too, I think its ergonomics are much better. If you are writing a front end you can produce the final output without an external bundled. Similarly if you are working on a backend TS project (less of a problem these days but transpiling TS used to be full of gotchas depending on your dependencies) . You can even just build a single binary to deploy in your container. Useful for command line tools to (they come out a bit big, but for personal use is not an issue)
It's not been the case for probably more than 20 years, but I remember that in the past the official website of the Italian Postal Service [1] used to be closed out of office hours.
Japanese websites still do this. If you buy shinkansen tickets between 23:00-05:00 you get:
* JR central - you can buy, but not reserve seats, as the system actually registers your ticket at 5am
* Jr east - you cannot buy.
Some banks/government sites also simply don't want to work at night.
Very frustrating. To me it spells incompetence, but maybe there's some non-technical stakeholder who made a stupid rule once. For JR its probably some nightly accounting system - not 100% sure but I assume all the modern systems are just plugged on top of the ancient ones just like the airline industry does it.
> scaling problems due to the influx of LLM assisted coding
You mean people pushing LLM assisted coding, or github's own coders pushing LLM assisted code to production ? :-).
I am more inclined to think that it is the second, but if the reason is the first, what is particular hard to scale ? It's not like everybody is vide-coding into a single repository, so I would expect that most of the load would naturally be scalable by increasing the resources. If for some reason that is not possible, I would expect a corporation as big as Github to keep the paying customers working by gracefully denying service to the free tier, not to collapse like an house of cards.
Last year I started self hosting my email. I found that once you setup the “usual suspects” (DMARC, DKIM, SPF…) outgoing mail works without issues. In fact it works now better than the (small) provider I was using before.
And setting them up properly is not even that difficult nowadays
Whenever HN discusses self-hosting email, half the people with self-hosting experience say they never had any issues and the other half says they had so many issues they had to give it up.
I self-hosted for years and I did all the standard things you're supposed to do, but I had tremendous problems with Gmail and Outlook. Other providers usually went through without issue.
I guess there is a certain amount of luck in this, probably related to the provider, or the range that your IP reside into. (I guess google and microsoft keep their own blacklists, regardless of what public blacklists say). I wonder if the best course of action in that case is to just spin up a new server on a different provider and try again.
FWIW, in my case, I am running the server on Hetzner, even though is supposed to have a bad reputation, but I have been using its IP for more than 10 years for other things (web server, personal services etc) so I guess that helps
Hard to tell. My server's IP was not on any public blacklists (confirmed by https://multirbl.valli.org/) and it was whitelisted on DNSWL.org at medium trust level.
Not all non-trivial workloads are web-scale. There are plenty of on-premise applications out there that have at most hundreds or thousands of concurrent users, and the connections come from a bunch of fat spring boot servers that handle most of the pooling by themselves.
This is all quantifiable. I regularly have my model run benchmarks against all the config permutations and then choose the best based on my criteria, which typically boil down to trading prefill and decode times
I'm running on Strix Halo so memory bandwidth is my constraint. In that example I'm describing the choice between using ROCm or Vulkan. I have a llama-swap config that can call different instances of llama-server running a toolbox with either runtime.
Memory bandwidth is my constraint too (M4 Max) but prefill and single-token decode don't run at the same time. It's best to use batched prefill so you can benefit from processing multiple tokens with a single pass through the model weights.
the slogan came out when the alternative was Perl, and in Perl you could do the same thing in a million ways, and each of them was equally "idiomatic".
> There should be one-- and preferably only one --obvious way to do it. [1]
Note the `should` and the `obvious`. Is it not a strict rule about having a single way to do things. It is about the aspiration that, if you do something, there is single obvious way to do it, much better than the others.
(I agree though that not even this is true anymore, see how many different ways you have to interpolate strings).
It is meant to give a rating inside a generation. Different generations are not necessarily comparable. an i7 is better than an i5, but not if you compare an i7 7xxx with an i5 14xxx.
A nvidia xx80 card is better than a xx60 card, but not if you compare a 1080 to a 5060.
The only difference for OpenAI is that not every generation has every model.
reply