> Funny how these things when done by a human is a positive and when done by an LLM is a negative.
> Humans generate perfect code on the first pass every time and it's only LLMs that introduce bad implementations.
That's not what the "anti-llm experts" are saying at all. If you think of LLMs as "bad first draft" machines, then you'll likely be successful in finding ways to use LLMs.
But that's not what is being sold. Atman and Amodei are not selling "this tool will make bad implementations that you can improve on". They are selling "this tool will replace your IT department". Calling out that the tool isn't capable of doing that is not pretending that humans are perfect by comparison.
> QA has to some how think of all the inane ways that a user will actually try using the thing knowing that not all users are technically savvy at all.
The classical joke is: (this variant from Brenan Keller[0])
A QA engineer walks into a bar.
- Orders a beer.
- Orders 0 beers.
- Orders 99999999999 beers.
- Orders a lizard.
- Orders -1 beers.
- Orders a ueicbksjdhd.
First real customer walks in and asks where the bathroom is.
How does PBKDF2 prevent an offline decryption attack with unlimited attempts?
All it does is slow down the attempts, but for the average person's easy-to-remember password, it's probably increasing the effort from milliseconds to a few days.
The solution in Australia is simpler - you don't submit the vote that you took a photo of. You can get a ballot, fill it out the "right" way, take a photo, erase the markings, write on your preferred vote, and submit that.
Even if you ignore the pencil they give you and use a pen, you can simply tear or damage the paper, take it back to the elections officer, ask for a new ballot, and fill that out instead. We make it as hard as possible to coerce a vote while maintaining secret voting (noting that it is definitely still possible, just hard).
Are there no polling stations where you can submit the ballot in a private location, like a drop box inside a booth or whatever? In the US I've only voted electronically, and it's done in a private booth with a curtain preventing external visibility, so somebody can easily video record the entire process with no realistic way of altering their vote.
No, the booths where you fill out your ballot are on one side of the room, and the ballot boxes are on the other, supervised by an election officer, and near the exit. You:
- collect blank ballots (usually 2 pieces of paper, one for the House, one for the Senate) at the entry,
- walk over to the booth,
- fill out your ballot in secret at the booth (taking as long as you like),
- fold the ballots,
- walk over to the ballot boxes,
- drop the folded ballots into the corresponding box (House ballot in the House box and Senate ballot in the Senate box),
- then leave.
As no-one sees what you write at the booth, you can vote legally, draw pictures on your ballot, write obscenities, write nothing, or a combination.
> It's easy to assume making a word disappear is always the right choice, but you forget it changes the word it leaves behind as well. Very clever.
It's also easy to make the opposite assumption, that the goal is to change the other word. I initially felt weird about changing from a letter at position 3 to the same letter at position 1, but eventually realised that the goal is just to slide the word around, not necessarily to make a new word.
> Humans generate perfect code on the first pass every time and it's only LLMs that introduce bad implementations.
That's not what the "anti-llm experts" are saying at all. If you think of LLMs as "bad first draft" machines, then you'll likely be successful in finding ways to use LLMs.
But that's not what is being sold. Atman and Amodei are not selling "this tool will make bad implementations that you can improve on". They are selling "this tool will replace your IT department". Calling out that the tool isn't capable of doing that is not pretending that humans are perfect by comparison.
reply