9,073 karma · joined May 10, 2020
But the Black Death mixed high contagion and high mortality as an actual example that shows they aren’t mutually exclusive.
It isn't about model capability - it's about inference hardware. Same smarts, faster.
And humans will always deduce that you should switch doors if you are in a hypothetical gameshow and they show you a horse behind one of the doors.
(All I mean is - an example of an LLM answering illogically is not proof that LLMs can't really think logically, as you can equally find examples of humans answering illogically and also find examples of novel questions that LLMs get right)
https://fs.blog/einstein-wertheimer-car-problem/
And that many mathematicians got monty-hall wrong, despite it being intuitive for many kids.
And being at the top of your field (regardless of the PHD) does not make you immune to falling for YES / EYES.
> humans without credentials are bad at basic algebra in a word problem, ergo the large language model must be substantially equivalent to a human without a credential
I'm not saying this - i'm saying the claim that 'AI's get this question wrong ergo they cannot be a senior software engineer' is wrong when senior software engineers will get analogous questions wrong. If you apply the same bar to software engineers, you get 'senior software engineers get this question wrong so they can't be senior software engineers' which is obviously wrong.
I've seen senior software engineers get tricked with the 'if YES spells yes, what does EYES spell?', or 'Say silk three times, what do cows drink?', or 'What do you put in a toaster?'.
Even if not a trick - lots of people get the 'bat and a ball cost £1.10 in total. The bat costs £1 more than the ball. How much does the ball cost?' question wrong, or '5 machines take 5 minutes to make 5 widgets. How long do 100 machines take to make 100 widgets?' etc. There are obviously more complex variants of all these that have even lower success rates for humans.
In addition, being PHD-Level in maths as a human doesn't make you immune to the 'toaster/toast' question (assuming you haven't heard it before).
So if we assume humans are generally intelligent and can be a senior software engineer, getting this sort of question confidently wrong isn't incompatible with being a competent senior software engineer.
If you and me both buy the same car, but i'm better at negotiating than you and get a lower price, I'm not free riding because you 'funded the design of the car with the extra money you paid', you are just bad at negotiation.
The rest of the world isn't free riding - the USA has just setup a market where there is very little bargaining power for consumers because of how the US medical market and insurance works.
Novo and Eli are still making plenty of money in Europe where these drugs cost a fraction of the price, and where there aren't other significant suppliers for GLP-1's like is being implied.
Well I think the article would say that you can diff the documentation, and it's the documentation that is feeding the AI in this new paradigm (which isn't direct prompting).
If the definition of programming is "a process to create sets of instructions that tell a computer how to perform specific tasks" there is nothing in there that requires it to be deterministic at the definition level.
It comes back to fundamental programming guidelines like DRY (Don't Repeat Yourself) - if you have three separate implementations in different languages for everything, changes will be come harder and you will move slower. These golden guidelines still stand in a vibe-code world.
IMO here in the UK we are good at starting tech startups, we are just bad at not selling them to overseas investors early in their life, or having a tax framework that is advantageous to them growing in the UK.
In the UK see Google Deepmind, ARM, Deliveroo... Elevenlabs being incorporated in the USA, Dyson moving to singapore etc - Even outside of the tech space, Cadburys, Sainsbury's, Jaguar Land Rover... If the UK kept hold of everything that the UK created, we would be great!
Even our infrastructure we sell to the French, Chinese, Germans etc just for short-term gain, despite that we are cutting our nose off if we look forward 10 years.
> Probably gonna share this story online more. I mean I didn't expect the situation to be this extremely bad.
Yes it's bad, although I want to emphasise that I have changed the details significantly here - it was something like this, but this was not the infraction because I don't want the company to be identifiable by anyone. I’m also not sure if my example is an actual policy violation - but it was in a similar vein.
To be honest, the surprising thing wasn’t the ban on a policy violation (because they had violated a policy), it was how automated the internals of Amazon were and how slow it was to overturn.
We are talking a very minor infraction - It was something like one of their marketing copywriters putting 'refills can be purchased on our website' on one of their thousands of listings, and Amazon delisting their entire account on the basis that this was moving customers off their platform. No warning - permenant ban that took over a week to remove - c$2m revenue loss (I've changed the details here significantly to avoid disclosing the company - this was not the exact scenario so please treat with a pinch of salt)
They had an account manager, but Amazon is so automated and huge that even at that scale it was a nightmare to resolve. It seemed like account managers couldn't automatically reactivate accounts or anything, they can just fill in forms internally but it seemed like they were getting automated responses back, or it was going to faceless teams etc.
Some customers will churn, some will stay, Bending Spoons are the masters of this model so will have made an assumption on how revenue will change across the next 5-10 years+, but I would assume that they aren't forecasting extreme growth, and instead are calculating that net profit can be changed from c$30m to c$139m within existing revenue, so if they can keep revenue at/near current levels without growth, they can end up with a much more profitable business.
Bear in mind that same revenue doesn't necessarily mean the same number of customers - it can also mean raising prices and having less customers. Bending Spoons might estimate that if they double prices, half their customers might leave - this would still be BRILLIANT for profit, as while revenue would stay the same, some costs would half, and thus profit might jump from c$140m to c$250m based on some napkin math!
> [banning smoking would] afford a direct encouragement to children to smoke. Most boys of a tender age who might be seen smoking in public places did so, not because of any attachment to tobacco, but because they considered it a practice in advance of their years, and something moreover which their elders told them not to do, affording them, therefore, the added pleasure of disobedience which was so dear to boys of their age.
https://hansard.parliament.uk/commons/1908-10-13/debates/6aa...
Perhaps it does make it cooler - but undoubtedly the restrictions reduced availability and reduced the number of children being addicted.
As for the actual age cut-off, it's always going to be fairly arbitrary, or a 'balanced judgement'.
They can function kind-of-the-same in the sense that they can both change things written in a higher level language into a lower level language.
100% different in every other way, but for coding in some circumstances if we treat it as a black box, LLMs can turn higher level pseudocode into lower level code (inaccurately), or even transpile.
Kind of like how email and the postal service can be kind of the same if you look at it from a certain angle.
1) Making the application small enough, and breaking it apart if needed (e.g. I've refactored my old 'big' app into 10 micro-apps).
2) Selecting an architecture that will work, looking after the data modelling and architecture yourself rather than delegating this to the LLM (it can implement it - but you need to design it).
3) Trusting that the LLM is capable enough to implement new requirements or fixes as required.
If requirements change so substantially that it's not possible, you can write new software as requirements change - as per point 1, you will have made your application modular enough that this isn't a significant concern.
I've written my own agent for a specialised problem which does work well, although it just burns tokens compared to Cursor!
The other advantage that Claude Code has is that the model itself can be finetuned for tool calling rather than just relying on prompt engineering, but even getting the prompts right must take huge engineering effort and experimentation.
> But here is what we are not born with: information, data, rules, software, knowledge, lexicons, representations, algorithms, programs, models, memories, images, processors, subroutines, encoders, decoders, symbols, or buffers – design elements that allow digital computers to behave somewhat intelligently. Not only are we not born with such things, we also don’t develop them – ever.
Really? Humans don't ever develop memories? Humans don't gain information?
I think the 'unique' part of fingerprinting here isn't working unfortunately.
Code quality is an issue you need to ignore with vibe coding - if code quality is important to your project or you then it’s not an issue. But if you abandon this concept and build things small enough or modular enough then speed gains await!
IMO codebases can be architected for LLMs to work better in them, but this is harder in brownfield apps.
I’m unconvinced that code is the best way to describe business logic, so here we are! :)
(Code is probably the best way to encode business logic, but most users can’t read it or edit it, which makes it bad)
The idea is that you can write the 'nodes' in plain english rather than pre-written blocks, and then the arrows indicate the flow but don't need to absolutely encode everything (closer to a flow chart). The process of writing the flow chart helps define and document the business logic, and the flows are totally clear because everything is encoded into a state machine.
If you want to be handling peoples secrets, you have to make sure you know what you are doing and build something bombproof (bombproof from a mathematical perspective, rather than relying on your server being secure)
* Your deletion policy says you delete images instantly and via the UI in settings, but I've checked and they are retained in the object store. You need to update these policies to be honest and say that the images aren't deleted, and that you currently retain them and just delete the reference to them.
* Your privacy policy says you can't see user content, but you clearly can (as you have both the data and the encryption keys). You might not have developed the functionality to read it yet - but it is trivial to do. Just be honest and say 'your data can technically be accessed by us, but we promise not to look at it'.
* Your privacy policy only limits your access to 'what is allowed by law' - which is clearly the absolute minimum!
I think your policies currently say how you would like it to be, rather than how it is. You need to be honest with users about how their data is actually processed.
> So while the image URLs aren't encrypted, they're still secure. Only you can access your images.
This isn't true though - and presumably you know it isn't true?
You would be able to access and download all the images if you wanted to.
> But we can't read the actual content of your encrypted items without your encryption key, and we don't have a reason to try.
This is also misleading - because you do have the encryption key, so you can read the content if you want to. "We won't read the content even though we could, because we don't currently have a reason" is the actual state of affairs.