philipwhiuk

4 ms·

1.
by philipwhiuk 12d ago
I agree.. I felt a bit blindsided that the guy would start with saying "we need provable statements" and then just throw this in there. For guest post on Tao's blog, after a decent start this fell short of expectations rather
2.
by philipwhiuk 13d ago
There are two hard problems in computer science, naming things, cache invalidation and off-by-one errors: https://github.com/google/ax/issues/356
3.
by philipwhiuk 13d ago
The financialisation of everything is, frankly, fairly toxic, so it's reasonable argument.
4.
by philipwhiuk 15d ago
> There turned out to be less detail available here than I hoped. I'm surprised the author was surprised that a component of a multi-billion dollar company which is vital to it's future success in the industry and is covered by
5.
by philipwhiuk 15d ago
Or insufficient testing?
6.
by philipwhiuk 17d ago
Yeah, someone probably deposited an amount a long time ago and it's sat there earning interest.
7.
by philipwhiuk 17d ago
Still no sign of an apology for any of the vandalism they've done.
8.
by philipwhiuk 17d ago
Interesting idea even for non-AI code.
9.
by philipwhiuk 18d ago
The difference is months inside staring at four walls.
10.
by philipwhiuk 18d ago
Was this part of a planned penetration test or did they just compromise your infrastructure first?
11.
by philipwhiuk 18d ago
> Back in the 2016 timeframe, members of our team were working at AWS and faced a similar challenge. Given all of the awesome complexity of AWS IAM policies, AWS S3 storage policies, historical version support- can we definitively say wh
12.
by philipwhiuk 18d ago
> Funny enough if the model thought it was on the real internet it likely would not have done any of these 'hack' events. As I've said before on this website, fool me once on this. If the model is prepared to break the rul
13.
by philipwhiuk 19d ago
You'd think https://www.wsj.com/tech/ai/anthropic-claude-ai-vending-mach...
14.
by philipwhiuk 19d ago
https://www.wsj.com/tech/ai/anthropic-claude-ai-vending-mach... It went badly at WSJ
15.
by philipwhiuk 19d ago
I feel like an AI lab is less complex than a physical vending machine business. https://www.wsj.com/tech/ai/anthropic-claude-ai-vending-mach...
16.
by philipwhiuk 19d ago
This is a parody, right?
17.
by philipwhiuk 19d ago
It's possible they were authored by OpenAI agents solving AISI tasks rather than OpenAI agents solving OpenAI tasks. That would explain the UK-focus to the data.
18.
by philipwhiuk 19d ago
> In my experience, LLMs only exhibit this kind of behaviour when they are put in sandboxes too restrictive too achieve their task. This is a dumb argument for making these LLMs able to access more of internet. If I'm hosting conten
19.
by philipwhiuk 19d ago
The fact that both are valid from a spelling and grammar perspective makes it an easy human mistake.
20.
by philipwhiuk 19d ago
OpenAI's careless approach to sandboxing and minimal levels of monitoring appear to be positioning it increasingly as a substantial threat actor to the open source ecosystem: * Hugging Face * D Programming Language Wiki * Ruby Gems If
21.
by philipwhiuk 20d ago
The medieval solutions to the king being out of money was always very straightforward. 1. Seize the gold from whoever has it 2. Punitively high taxation Not sure much has changed
22.
by philipwhiuk 20d ago
> gave it some encouragement. I told it to look online at some of Fable’s strongest feats, especially the math problems it has solved, and that something like this should be easy in comparison. This is pretty ridiculous when you think a
23.
by philipwhiuk 20d ago
People are trying to compromise Tesla and because this guy provides NTP services, and Tesla set their NTP up wrong, it appears to other people like his machine is part of Tesla.
24.
by philipwhiuk 22d ago
The point of mathematics is human understanding though.
25.
by philipwhiuk 23d ago
There’s lots of ways of attacking the problem sufficiently to get to 95% and we’ve spent decades on object recognition
26.
by philipwhiuk 23d ago
It was not clear from watching how he got to 11 but the number is not particularly relevant - once you have one of them you induce many.
27.
by philipwhiuk 23d ago
If you upgraded from < 1.25.5
28.
by philipwhiuk 23d ago
> It doesn't expect the first output to be correct, and builds a loop where an imperfect attempt simply cannot move forward until it becomes a good result. So... local maxima?
29.
by philipwhiuk 23d ago
I guess expect no new features on mobile until their token budget gets through all the screens?
30.
by philipwhiuk 24d ago
FISA basically authorises dragnet on demand - it's how XKeyscore operates.

More ›