bobm_kite9

Amazing that they're able to vibecode this kind of stuff together in a couple of days, but how is it possible there's no API for organising my ChatGPT chats at all?

Welcome to a new kind of lock-in I guess

Of course OpenAI used these guys' training data. This is part of their ratchet strategy: if you can monopolise the creation of new knowledge, you win. The moat they are building is user data, which a billion people are now throwing at them every day. They are solving difficult math problems using it. The only way to get the this data and these techniques back out is by distillation - expensive, slow work.

This moat ensures that Anthropic and OpenAI will remain at the frontier - no one else can train a super-intelligent model because they lack the trove of user data.

My main takeaway from reading this is - why do we always fall for this same trap? Every single time we allow a software company to build a monopoly in pretty much the same way. Microsoft, Oracle, AWS - They have a cool tech and we let them run away with it.

this might be burying the lede a bit, but I feel like BDD tools deserve a renaissance given the amount of agentic coding going on. There's an article linked in the readme that explains why I arrived at this conclusion. Would be interested to see what other HN'ers think

I think this is set to happen: https://www.agcc.co.uk/news-article/hydrogen-will-be-pumped-...

Good article but this misses the obvious - but slightly more distant - eventuality: the no-code vendor goes bankrupt meaning no further support or security fixes, or is bought out by some other firm which means exorbitant future licensing costs and no further features.

Again, some people might not care about that but it seems kind of a biggie to me.

My pet theory is that OpenAI releases are on hiatus pending the US elections. It would be a huge disaster to release SORA in the next few months and have it some way sway voting. What do you think?

That seemed easy. What was required to make the batteries dispatch power in this way? I have rooftop solar and batteries in the UK and I would definitely engage in a similar scheme like this except I wonder how much work the power companies would have to do to make it work.

I'd love to see Java adopt Kotlin's more advanced type system: Proper support of nullable/non-nullable types, better generics and no primitives.

Not sure how they could get there but that would be amazing.

Also Kotlin has a pretty good JavaScript and Native support story now.

Yes, you're right. There could be multiple ways to tokenise a sentence. Shouldn't all the valid tokens be included in the vector?

This sounds a lot like the approach taken by GraalVM. Can someone better versed in this area comment?

And hands. Try searching for ‘hand holding apple’ for instance

I’m going to have a stab at actually answering the questions, rather than just taking the opposing point of view.

There is another way: organising regular off-sites for the employees to meet up, get to know each other, build relationships, share culture and attend workshops so that they can learn the common tools and techniques that are valid across an org.

I would argue this is a better way to onboard new staff too, if you can hire them to coincide with these off-sites.

Even if the off-site is 2/3 week-long all-expenses-paid trips pa, this will probably work out cheaper than hiring office space for a year

Those are way too upbeat

https://twitter.com/risk_first/status/1487374583024005123?s=...

I think the best ice-cream is dairy-free.

I've always been a huge ice-cream fan so the fact that it's been retooled in a more addictive form is very concerning.

This seems like a nice thing to have badges for. Couldn’t there be a certification programme for ‘dark-pattern free’ websites?

Although I like the idea of DDG's bang operators, I rarely use them (mainly !g when I'm feeling desperate).

What I would find useful is to be able to whitelist a bunch of sites on DDG, so that it prioritises results from them first, when I search.. basically most of the sites with ! operators I guess.

That way I wouldn't get all the SO clone-sites returning their rubbish.

I touch on both these perspectives in my book ‘Risk-First Software Development’. You can read all it online at https://riskfirst.org if you want.

Hope this is useful to you.

Yes! This is one thing humans are still much better at than AI: taking blame.

Every year I do a photo book. Used to use apple but they discontinued the service so now it’s snapfish. It’s a lot of work to create but worth it

> A Ponzi Scheme is where the sum of all investor's reported account values exceed the total value of assets held by the institution.

Also how fractional reserve banking works

Everyone seems to be joking about this, but doesn’t this result indicate that human neurones have a better learning algorithm than the ones we are using to train AI?

It’s interesting that this is evident in even small clumps of human brain tissue.

Really interested to know how this might work.

Talking about risk as it affects software projects is the main topic of

https://riskfirst.org

if anyone is interested in reading more on this.

(disclosure: I am the author of that site)

The law of large number works well for some things (say, shoe sizes) and fails to work for others (e.g. income distributions, which follow a power-law).

So it's not odd at all that it would fail to adequately account for people. It's just sad we don't have anything better. Can't we find something better?

My takeaway is that danluu is saying that, at the small scale, people understand how individuals contribute to teams. But at the large scale (like the HR department) we don't.

At the small scale, we're able to apply systems thinking - a team can be viewed as a system composed of a set of interacting parts.

However at larger scales, systems thinking becomes unwieldy - the interactions increase quadratically. There, we fall back to the Law of Large Numbers [1].

"Fungibility" and the related failures arise as a result of this.

[1]: https://en.wikipedia.org/wiki/Law_of_large_numbers

I just need to write a GAN to generate plausible looking irises, then - Profit!

Can't someone turn this into a feature of HN?

I guess I'm interested to see if this performs qualitatively better than GPT-3, given how many more parameters it has.

However, I think this is really a dead-end: throwing more hardware at this is just going to generate better-sounding nonsense. Yes, we are learning the "model" of the English language - which words go with which others, but successively larger transformer models don't really expose much more about the nature of intelligent conversation.

I think we need a better algorithm now.

If you’ve ruled out mental causes (like perhaps ADHD, as discussed in other comments) then perhaps it’s something about the work.

I’ve often found in these periods that there’s some underlying part of the work that makes me not want to do it, usually that subconsciously I’ve categorised it as being pointless.

It might be worth brainstorming the nature of the work itself rather than blaming yourself.

> Notably, GPT-3 is licensed by Microsoft, and is therefore closely guarded. You interact with it only via apps which act as oracles. It’s basically a genie stuffed in Microsoft’s basement you can Zoom with.

This brings to mind Eliezer Yudkowsky's AI-box experiment [1]. Although GPT-3 is mindless, how long will future iterations _remain_ in Microsoft's basement?

[1] - https://rationalwiki.org/wiki/AI-box_experiment

I think you're right. This is discussed in the article- GPT-3 does very well at the small-scale, like short poems. The architecture of GPT-3 is based on Word2Vec, and attention - it doesn't really have a good understanding of large-scale semantic structures in writing. (Also mentioned in the article - it can't rhyme: it doesn't understand sub-word semantic structure either).

I think the next big advance will be when we are able to apply attention/word2vec-style algorithms at other semantic levels.