eugenekolo
eugenekolo on Hacker News
That's a good way to look at it
Just like Graduate Students are good at Gradient Descent, they can see patterns and fine tune parameters 100000 times until the result looks good. So too can LLMs with their ability to examine patterns and brute force things.
Wouldn't hurt to link Looker in the first sentence.
No mention of the RAM size.
Thinking if the RAM is good, this is a great jailbreak target with an A18 chip for only $299.
Somebody needed a promotion project huh?
I believe your use is what they're trying to discourage.
example.com is meant to be legally allowed to be used in text such as "You visit a website (example.com) to browse the internet"
It is not meant to be queried by automated tests or used as a service.
"I work at Anthropic and used idle GPUs running known algorithm to factor a number. I used Claude btw"
Hobby groups projects like this are less fun for a lot of people who used to enjoy interacting with smart people. It's definitely become a game of just spam claude for answers with zero understanding or care for how anything actually works. That's fine to get things done and fine for a lot of side projects, but it definitely ruins the joy that people have in understanding systems and working with intellectuals.
It really is an absolute massive jump. Have no clue what's going on in the back to warrant a 3x increase... 25-50%, sure.. but 3x is wild.
WOuldn't proof of that be some grounds for breach of fiduciary duty?
Without SWE-Bench though, how will AI models properly game their results to show ~5-10% gain each iteration?
Once a benchmark is known and there's billion of dollars on the line, obviously every company will game them.
Generally I agree with the sentiment, for one major reason: debugability/dev loops.
Inevitably you're going to end up having to debug that each of those steps is correct, for that you'll find it a lot easier to break it out.. and the next person who has to do it will as well.
I do think the example is somewhat loaded though, rename "result" to "top5ActiveUserNames" would do a lot there.
I'm not sure Bitcoin won.. it just continues being a ponzi scheme that you can make money in.
You can also accept that certain things and be happy in life either way. Don't need to chase get rich schemes. Some are more privileged than others in being able to do this.
No, because Lockdown Mode disabled JIT which is a part of this exploit chain.
Unrelated bug as far as I can tell.
I suspect you'll see one with this or Coruna soon enough.
Donated. Thank you and everyone else on the PyPy team.
I use PyPy regularly on an app of mine, and very often when I need to do some compute heavy load. Typically over 5x faster than CPython. It makes some stuff that takes impossibly long with CPython (nobody wants to wait 5 minutes...), to returning a response in a few seconds.
But is it running ShellExecute on URIs?
I disagree you'd find "obvious" non-generic names easily. After all, "naming" is one of the hardest things in computer science.
I think paying for search is becoming more acceptable when you look at the amount of people paying $20/mo for their AI subs, which to many people are just search engines.
There will always be users who refuse (not going to convince my parents ever), but for many power users, or semi-power users, it's becoming more acceptable to just pay the $20/mo and get a better product.
So you can detect if it's AI slop quickly.. but they might be catching on to this.
It's possible they changed it from "thisisunsafe" to the b64 version to avoid automatic scanners finding "unsafe" keyword usage.
Confused by his portrayal of Massena, NY. I don't live there and have never driven through, but looking on Google Maps it doesn't seem that bad or depressing as the author (and I guess commentators) make it out to be.
It has a Walmart, Home Depot, BJs (similar to Costco), a main street with several businesses. A walkable grid with sidewalks in that main town area....
Feels like reaching that this place is so desolate and depressing.
It's definitely an attempt to compare models, and Gemini clearly won in the tests. But, I don't think the tests are particularly good or showcasing. It's generally an easy problem to ask AI to give you greenfields JS code for common tasks, and Leetcode's been done 1000 times on Github and stackoverflow, so the solutions are all right there.
I'd like to see tests that are more complicated for AI things like refactoring an existing codebase, writing a program to auto play God of War for you, improving the response time of a keyboard driver and so on.
They heavily compete on "privacy" and "security", so I wouldn't expect them to. Additionally, once you start rolling with one government, every one wants you to do something for them while offering you no additional money for the work and weakening of your project.