Congrats Tom. I wonder if a particular model could be used as a baseline for these values or if they are already doing that to check first level prior to a human in the loop? I myself have been using AI for this purpose and have found it getting pretty good. I know its not a replacement for thoughtful moderation however, a tailored model for HN would also promote the tradition of HN in terms of having it not just be about who is there and have it more trained on its best practices to promote consistency, possibly as an aid.
What struck me was the interruptions to the AI speaking which seemed commonplace by the team members in the demo. We will quickly get used to doing this to AIs and we will probably be talking to AIs a lot throughout the day as time progresses I would imagine. We will be trained by AIs to be rude and impatient I think.
Currently, what I get out from it is a good quick overview with some hallucinations. You have to actually know what you're doing to check the code. However, this is a fast moving target and will in no time be doing that part as well. I think stepping back and thinking maybe this thing is just giving us more and more agency and what can we do with that? We need to adapt to not constrain ourselves to just being programmers. We are humans with agency and if we can adapt to this, we can be more and more powerful having our technical insight that we've gained over the years to do some really cool things. I have a startup and with ChatGPT I've managed to do all parts of the stack with confidence and used it for all sorts of business related things outside of coding that have really helped move the business forward quickly.
IMO Web Components generally suck in developer experience but are good for bridging different js projects or web frameworks as the web components are very compatible between js frameworks. You can literally make a thin wrapper to expose your js flavored framework-based component such as react for example and expose it as a web component to another system such as angular 1 for example and it will work pretty well. You could potentially use this strategy to have parallel tracks of development where you can create an entirely new framework from scratch and start to share that functionality in the legacy project while you're in the process of upgrading which could take a long long time in some cases. At least you won't be continually working on the old one and adding more tech debt... although one could argue this wrapper approach is tech debt.
I'm currently using Deno deploy and found it to be fantastically performant and dumb simple for my lone wolf project. I'm experienced in AWS development in larger teams and it is nice to see a move away from complexity for a change where you can easily set something up without having to think about setting it up at all. The DNS stuff was just dead simple as well and automatic ssl certs was super nice. I have 0 complaints for what this is trying to be and am excited for the road map.
You can also continue to chat with it in each of its answers and ask it why it said this or that and see how it works out corrections and clarifications which to me is so much more significant as that's how people converse. One-shot answers are for google.
Asking it why it said something isn't a great pattern, because it can't answer that truthfully: each interaction with the bot resets its "memory" entirely.
You're effectively asking it to invent a rationale for what you are telling it was the thing it told you last time round.
Asking it to "think out loud" during its initial answer is a better way to get insight into why it answered in a particular way (and also often causes it to provide better answers).
Asking it follow-up questions is almost always useful, it's just the "why did you say X?" pattern that I'm recommending against.
>each interaction with the bot resets its "memory" entirely.
Completely separate interaction maybe. You can certainly ask a question then ask it to rephrase the answer, etc. Like "say five words", followed by "reverse the answer"...produces the same five words in reverse.
Asking a question then following with "why did you say that?" usually cites the previous context reasonably.
Sure, but the thing you have to understand is that there's no hidden state from the previous answer that can be revealed through extra questions. It's lost all of that state the moment it spits out the reply tokens.
> You're effectively asking it to invent a rationale for what you are telling it was the thing it told you last time round.
There’s lots of evidence that this is what people do, too. Asked about their reasoning, people will do things like include information they didn’t have at the time. The part of the mind that moves the fastest doesn’t encode the process to memory[0] but we’re really uncomfortable saying “I don’t know” about our own decisions. So, we rationalize.
If you want to be more accurate, you need to do the same thing of “thinking out loud”, either subvocalization, voice or paper. Even then it’s not perfect. Easy to make up your mind and justify it after.
[0] Which, of course it wouldn’t. You don’t log every line of your fastest piece of code. Then it’s not fast.
Yes. I strongly feel that reinforcement learning should be applied to punish the LLMs for speculating about their past behavior. They should respond along the lines of “I’m sorry, I don’t know why I said 3 + 5 is 9, but I will try to answer again.”
I think didn't do a good comment here because I don't mean that exactly. I don't necessarily care so much about it telling me how it got there. Rather I like it when I ask someone about something then I can ask them... "Tell me more about this or that or explain that in more detail" which I can never do in google.
I can think of it like the wheels are still there, the frame, engine, body etc but the design and materials have been completely refined and overhauled to be modern future tech today. The car even has new stuff like electronics and wifi. Meaning there are still base conceptual things there but they are not the same...knowing that reasoning by analogy is flawed but still satisfying to me to feel like I understand it enough.
There's quite an interesting analogy there. I guess you could have a four wheeled wagon structure similar to wooden wagons of old, then an engine added and now electronics added, as an analogy to the triune brain. It's not that modern cars are built on wagon chassis but that the basic body with four wheels structure is similar for functional reasons. Probably similarly with the 'reptilian brain.' Our bits are maybe similar because they do the same job rather than because they evolved from that.
Personally this is my experience. I've had a couple of very distressful/low times in my life and I've always instinctually exercised rigorously in terms of weights, running, swimming, etc during those times to take my mind off of things outside of my control and have found it very therapeutic. The adrenaline and pain felt from pushing yourself through a hard workout takes your mind off of things and also makes you feel more alive and energetic. It is very difficult to feel down when you are experiencing these feeling from working out IMO. From my perspective, I like to describe it as I could not escape my own mind of a stream of negative and impending doom of thoughts but when working out and giving it your all, it allows you to feel and focus on the environment around you and get outside your head or negative thoughts, giving you a break mentally from that constant stream of bad mojo.
I would have to guess working out like that puts you in flight or fight mode. Which I think is part of our evolution when in fight or flight mode to have a heightened sense of awareness of the environment around you which due to our limited locus of attention, shifts attention or focus from inward negativity to the "immediacy of being". This also seems to coincide with my normally perceived fast tempo of music that seems to slow down when in a hard workout.
I imagine an analogy of summer camp as that is my only way to relate as I'm in a rich country. I'm going to go to a summer camp not of my choosing and participating in activities not of my choosing and being bunked with people not of my choosing for the rest of my vacations for life and then when i'm about 80 years old then claim there is negligible delta from not having to do that. That last bit, I can't imagine.
It's going to get more interesting or disgusting depending on your compass (no judgement either way here) when robots are connected to this as that reminds me of times Elon was snickering when talking about Optimus and "other uses".
There's an interesting set of ethical questions surrounding the ways that people with socially unaccepted preferences/fetishes use technology to satisfy forbidden urges. If we accept that individuals aren't able to (and generally shouldn't have to) control what they find sexually attractive, then it seems admirable to be able to use tech in ways that don't harm any actual people. If LLM erotic roleplay begins to replace the highly exploitative realworld porn industry (which of course uses real humans, many of them psychologically, socially or financially vulnerable), that seems like progress.
I definitely see the logic in your statement (and it seems like tech/logic-oriented people tend to agree with you and me), but I lament that most of society operates on a much more "feelings-based" basis, and thus they generally see something like "$Person engaged in a detailed $deepTaboo simulation with computer using the $SoftwareProduct" and start looking around for someone to crucify (usually both $Person and the owner of $SoftwareProduct) for 'promoting' $deepTaboo -- without asking "who was harmed?"
Even if the constitution is actually consulted and the parties are spared pearl-clutching prosecution about it, the court of public opinion will punish the company just as hard, pressuring gatekeepers like Apple to ban it, etc.
And thanks to the deep, disturbing taboos that can be involved, it's super hard to find someone who will stick their neck out to challenge all this. I know I wouldn't, because I know I would be immediately branded as "Joe Bloggs, Fan and Advocate of (Insert disgusting perversion)"
I wonder if this is something more advanced societies would come to terms with. I know ours is not equipped to handle it in a smart way.
