I’m curious about if the model has gotten more consistent throughout the full context window? It’s something that OpenAI touted in the release, and I’m curious if it will make a difference for long running tasks or big code reviews.
one positive is that 5.2 is very good at finding bugs but not sure about throughputs I'd imagine it might be improved but haven't seen a real task to benchmark it on.
what I am curious about is 5.2-codex but many of us complained about 5.1-codex (it seemed to get tunnel visioned) and I have been using vanilla 5.1
its just getting very tiring to deal with 5 different permutations of 3 completely separate models but perhaps this is the intent and will keep you on a chase.
I hate the new electric busses from China. Their acceleration is much better and braking effect is also stronger.
Bus drivers in Norway are binary people. They either press the accelerator or they press the brake. Most drivers call it leg day, because you spend the entire day pushing on these peddles as hard as you can.
Our existing buses have terrible acceleration, which helps a lot with the comfort of the bus ride. But for some unknown reason someone decided that the bendy busses should have the only wheels with power, in the rear of the bus, after the bend. So any slight hill that is slippery from a bit of snow or frost, now becomes a a comedic video of buses trying to drive up before they flop in the middle of the bend and slide back down again.
I don't think anyone is pretending that a Macbook Pro can compare to 8 H100 cards from Nvidia in terms of LLM training or for serving LLMs. But you can buy an awful many macbooks for the price of 8 H100 GPUs.
But if your workload belongs on 8 H100 GPUs then there isn't much point in trying to run it on a macbook. You'd be better served by renting them by the hour, or if you have a quarter million dollars you can always just purchase them outright.
The H100 is just an example, this is true for any workload that doesn't fit on a laptop.
Phishing. Super easy now to send a fake email with a great offer, and have your name and loyalty programme number right there in the email. Much easier to trick someone when your email contains a bunch of personal info that you wouldn’t assume others to have.
«Happy birthday! As a loyal Quantas customer, we would like to offer you a sneak peek of our upcoming Black Friday deals. Consider it a little birthday present from us.»
The current groups, sure, but the existence of a functioning market tends to bring in more participants. Or to put it another way, there are plenty of smart people in the world who found themselves born in a less-than-ideal country and are willing to solve their problems through crime.
The only sustainable solution is to make crime no longer pay. Nothing else will work.
The other solution is making those “less than ideal” countries have more attractive legal economic opportunities so that crime isn’t an attractive alternative.
The only reason these persist is because companies pay out and they can receive it in untraceable crypto currency in countries that are nearly to prosecute them in.
You make it sound like a simplistic game with set rules. There will be myriads of other reasons to breach companies, and even strictly sticking to the money part, doing ransom/extortion can have secondary and tertiary effects worth enough to do it even if the ransom fails.
If you look at it as a market, the victim is only one actor among many.
You can probably pretty easily just say Prime==Performance and Performance==Efficiency, but I think the "Prime" branding is kind of a carry over from Snapdragon mobile chips where they commonly use three tiers of core designs rather than the two. They still want to advertise the tier 2 cores as fast so T3 is efficiency, T2 is performance, T1 is Prime.
As an example, the Snapdragon 700-series had Prime, Gold, and Silver branding on it's cores.
The interesting part is the E-Core on A19 Pro are nearly as good as the previous ARM Big Core while only using half the power. I would love to know the die size of the cache and E-Core.
ARM were catching up to Apple in terms of big core, now Apple has leapfrogged in E-Core again. But competition is good. ARM should have some announcement coming in next few months.
In Norway I've noticed that stringent requirements for privacy make it much easier to run things in the cloud if the physical location of said cloud is actually in Norway.
So if OpenAI is hosting their services within the borders of the UK, then they would also be beholden to UK law. Makes it easier for the financial sector, government and healthcare to use their AI models than if they would have to send their data to a datacenter in the US.
It’s great for multitasking. I’ve cloned one of the repos I work on into a new folder and use Codex CLI in there. I feed it bug reports that users have submitted, while I work on bigger tasks.
reply