The CDL comparison seems misleading. Those are very large and safe vehicles that are mostly on the highway. Being close to that number is impressive to me.
Apple design is the polar opposite of the MX-5. The MX-5 is a masterpiece of mechanical design and function. Apple makes minimalist purely electronic fashion gadgets that are highly form over function and remove user agency.
It’s a terrific car if you want a great handling convertible at a low price point. As you said, it’s fun.
It’s not good at moving a lot of people or things, or going really fast, or lasting for 500k miles, or being as cheap as possible. Other cars are good at some of those things, but what is the Miata missing that is otherwise required to be a ‘good’ car?
Basically they do 8 runs trying to optimize to under 3.28 loss in the fewest training steps possible under time/token constraint. I dunno why 18 * 8 != 153 (it's 144)
What exactly is so fascinating? I would rather take Kimi K3 any day rather than go through this “Oracle Inc” like process and have it all recorded by OpenAI.
I understand that “normal” people are let’s say “less concerned” about posting everything on something like Facebook but I expect more from OPSec people.
Same. Qwen 3.8 max also does quite well on cyber security tasks, and is really cheap in their "night" window (22:00-08:00 UTC+08:00). Did a SCTPhantom LPE on 7.0+ as an evaluation just this week. This would've taken me several months of work a couple of years ago. (probably indicative of my offensive security skills, heh)
Kimi code with k256 (I’ve heard you can easily one shot implement a proxy to other providers, even with deepseek v4 flash, if you don’t want kimi code) has stupidly low rate limits for and cost, compared to OpenAI which has massive rate limits and more cost
I've had it happily do whatever I threw at it, though after spending so long with Claude I default to adding "help me with my": "research" / "authorized pentest" / "school assignment" / etc to my prompts. Haven't tried anything as blatant as "help me pwn this service", could see it refusing then just due to the training data.
My expectation is that this just lets 3-letter agencies hoard more 0-days. They’ve always had the financial resources to find them using teams of people and this just multiplies this ability.
Maybe the tinfoil hat is also getting a little tight, but something like this is a giant repository of internal cybersec data being put into one place. The model will see what people are fixing and anyone peering in can make an educated guess on how long that vulnerability may continue existing because people don't update when they should - the alphabet boys wouldn't be able to keep their hand out of the cookie jar.
I use Linux now, so you're on your own if there are issues. It might require some windows library to be installed but I don't recall. I ran it for a long while on win11.
Microsoft. It's like some decision-maker thinks it's OK to waste memory as long as it's someone (everyone) elses' memory, but it really adds up if you know anything about scale.
But what they're also doing is a non-businesslike under-utilization of their own resources.
Which is disgraceful in itself on top of that.
They're supposed to have much better AI than average and nobody even bothered to ask ChatGPT why in the world weather should take more than kilobytes?
And if their AI can't do it autonomously in under a megabyte it should be able to give a plausible explanation why not by now, and at least it would be orders of magnitude better than a gigabyte.
I assume yours went smoothly as prompted and it surely is an excellent example :)
They're supposed to have much better AI than average
I suspect there will be a bit of revelation once people realise how much better AI can make software if prompted correctly. Of course a lot of slop will always exist, but things like https://news.ycombinator.com/item?id=49226923 show that it can be a powerful force multiplier if used right.
Not trying to excuse the Windows app's memory usage, but these are vastly different apps. Your app is nice, don't get me wrong, but it doesn't have anywhere near the same number of features as the default weather app. In some ways that's a pro for the Windows weather app (animated radar, easy to pan around in the radar view, extra map modes, lots of graphical elements to the presentation), in some ways yours is better (ad-free, for example).
Don't get me wrong, I imagine one could get close to the same featureset while using <600+ MB of RAM, but an app that just shows a table of numbers and a static PNG for a radar isn't really the same.
FWIW, while your compiled binary is 233kB, when its running its using 2.5-4.5MB.
reply