unshavedyak

I have a slightly different approach. I've done this as both a profession and a hobby for around 18 years now, and even these days i spend quite a bit of time "making" things i take pride in. Be it decisions (decision fatigue is a battle), code written, architectural improvements, etc. I don't generally feel lacking in this area.

Where i've felt lacking for the last ~5 years though is output. Specifically blocked by having the energy to create all these damn ideas.

I find LLMs neat, albeit their own type of exhausting, as they open up a lot of possibilities for things i want. Ie software i've wanted but have never cared to find the time for or didn't have the time for.

LLMs feel like a software equivalent to a 3D Printer. I don't have to carve it out of wood now. But also like a 3D printer there are some tasks that LLMs are just horrible for. So it's a niche, a skill even, to find software you want that also are a good for for unsupervised development. It's also of course massively more risky than a 3D printed doohickey depending on its internet access/etc.

With that said i still don't really take pride in something vibe coded. I just take enjoyment out of using a thing that i've wanted but never had time for.

Stop, what, exactly? This is simply a new tool, not an AI girlfriend. We've had ML based autocomplete for 10 years now, why is it suddenly a new chicken-little for ya'll?

Yup, or we'll find out there's nothing to fear for folks who are using this responsibly.

It seems a huge "the sky is falling" to think all LLM use is bad for fear of some "addiction" to me. Even many skeptics (eg Hashimoto) have come around to the idea of using LLMs.

One thing is clear in my mind. VCs are burning cash, and so if you have an LLM flow you find useful, take some free cash. The sky is not falling in this respect imo.

Which is not a defense of AI, to be clear. AI may very well be a big societal problem, but in this context i don't see Hashimoto/etc becoming heroin addicts like ya'll are so concerned about. It's just a fancy autocomplete. The fear seems a bit over hyped.

Which is funny because I’ll take the free crack in this case lol. Am I addicted to this new workflow? No idea but there are many providers. I feel like I’m being given free VC money so yea, I’ll use it.

My expectation is that this cash handout is going to stop soon, so take while the giving is good right?

I should note that for my workflow /loop is eating the credits, so it’s basically no skin off my back. I just queue up some more work and let it go.

Huh, don't think i had any of those. Though arguably i had "Lack of focus", difficult to say how much is due to the lack of caffeine or due to undiagnosed ADHD though.

Generally i felt fine. I'll keep it in mind, thanks

Huh, i should look at this. I've been an aggressive drinker for most of my adult life (2 pots a day at my height), but for kicks i decided to cut all caffeine for about 9 months. No real issues aside from very short term headaches, though even those i mitigated by gradually moving down in quantity.

Aside from the headaches what addictive effects are you referencing?

For those of you who get value out of Blender please consider donating: https://fund.blender.org/ :)

It's a wonderful project. Even small amounts help. Thanks!

I'm tempted to try it out. I'm not keep to move but Fable rejected some work i was working on recently and frankly it's infuriating lol. I've been on Claude x20 for like 8 months now and now i'm tempted to switch out of spite.

It is surprisingly offensive.

The only friction for me is the general expensiveness of trying out top tier models, eg OpenAI's Fable equivalent (Sol?) to run for a trial period. I'd like to see a like-like comparison, eg buy x20 on OpenAI and see how much Sol i can use, how well it works, etc.

edit: Though surprisingly Claude seems to think Codex doesn't have hooks? That'll be tough, i use Claude hooks quite a bit.

It's a lossy conversion though. "Mistake" is relative to the stated goals and specifications which are often heavily lacking. So unless you write with a high degree of architectural and implementation specificity then it might make very high quality code that is still not what you wanted.

So i agree with you, but there's no SOTA model that i don't have to babysit. I'm not going to just throw a large pile of code in there unreviewed, and so what i want is faster iteration on code in logical, reviewable chunks. Ie just like i'd normally write myself; small, logical commits.

Faster iteration means i mentally checkout less and am more involved with the code being created.

My hope is that in the far far future, we can get LLMs so fast that i can work in my IDE like normal and the LLM will just be an extension of autocomplete. I can state a goal, rough out functions, code, etc, and it'll just work around me like a very fast pair programmer / autocomplete.

The chat interface is an intermediate step that frankly i hate. The faster it is the less i wait.

Now for vibe-slop i'm making on the side, yea i don't care about speed. But that's not something i'm employed to do or anything i truly care about. It's a different workflow entirely.

I mean for many of us this doesn’t even increase the cynicism. Ie it can’t get worse. These AI labs are already believed to be shifting behavior of these models at runtime so you never really know what you’re getting. Stability is not a thing with any provider, if you need that you have to run it locally.

I’m not offended by that because I’m already maximally offended lol.

Stronger, yes, but not necessarily strong. 2 instances of hearing that the sky is green might be stronger than 1 sky is blue, but that might not fit an expected threshold of say 100 instances of learning for a "fact" to be actually strong. At least in my example.

My point was that i feel like humans have these two aspects, the ability to have a fact, and the ability to have a signal to the facts strength. I propose that as an explanation why we can internally analyze our understanding and come to a conclusion that yes, we do "know" it.

We also at times can't figure out how we "know" it, either because we've made up a detail (filling in blanks, assumptions, etc), or because we forgot where we embedded this detail. The lower the signal strength of this validation the more we can be unsure about a "fact".

I feel like LLMs are one half, but not the other. They effectively need a RAG for all of the knowledge they have, and if the counter for a given fact/idea/etc is low enough, then it's an low signal.

The question i have is how to do this efficiently. Of course i'm just speculating too, i don't know any of this.

Yea, i'm on x20 and while it has been up and down in terms of token-usage-UX, i feel like its the best its ever been. Context: I entirely use Opus 4.8 fwiw.

Now is that because 4.8 is nerfed compared to 4.6 and thus more token efficient? No idea. I just know on x20 with a pretty plain workflow i struggle to use my tokens every week.

It's honestly kinda interesting. Now we're at a point where SOTA model companies aren't the ones who release the best tech, but who release the best and actually usable tech.

A worse product could win right now if it simply does as its asked.

I wouldn’t let anyone else use it in general because it’s an asset i don’t want to lose. As it is I now have to be careful what I use it for as I don’t want to trip any flags. Eg today I was curious how some LLM benchmarks worked and I wanted to talk through how I’d develop some, running some models locally, etc. however I don’t want to be flagged as a potential competitor and have my account revoked from Fable/etc.

It’s feeling quite similar to why I distributed out from google all those years ago. I didn’t want a hugely important centralized google account to be banned and cause friction to various aspects of my life.

The ability for LLMs to more easily catalogue user behavior and intent is going to get more interesting. Weird days. Feels like anyone could become a Facebook level metadata hoarder.

Yea but i don't think the LLM can inspect that. Eg it's similar to a code function , the code executes but the code doesn't have the ability to inquire about the code itself.

In that silly example/thought, the LLM would effectively need the ability to query the strength of a fact/spatial region/etc.

Right now i believe the LLM is more just the output of those weights. It has no way to inspect the strength of the signal. Eg it doesn't know if blue in "The sky is <blue>" is a strong or weak signal, it just predicted that next token.

If we could somehow encode strengths along with every token, eg "The sky is <blue:1.0>" or something we could perhaps give it a sense of [un]certainty. Though i imagine it would look differently than that since we'd want to encode this information in some sort of multi-dimensional space, rather than purely by token - since tokens aren't that valuable. Eg the "knowledge" in an LLM goes beyond tokens, and so too should signal strength.

I'm of course speculating on all of this and i have no clue on anything.

I often wonder how humans "know" things. I suspect (ignorant armchair) we have some ability to signal strength of those facts, via repetition. Without this layer of introspection i imagine LLMs can never avoid hallucination.

It obviously breaks down with humans too, given we so easily hallucinate and confuse things we "know". However i still suspect we're more reliable at probing information we've experienced vs not. Even if the case of poisoned knowledge, eg a crime scene accidentally implying information to a witness that the witness doesn't actually know, we still "know" that poisoned information via incorrect inference. Ie we "experienced" it.

Wonder what architecture would allow for this style of information/weight probing for an LLM.

Yea it's definitely portable, it's just not a friendly formfactor for where my compute sits, where my keyboard sits, etc. If i'm in a chair at the part i'd need a literal lap-top, three components (keyboard, compute, monitor) without a frame connecting them would make that difficult.

I'd buy it but i want it in a laptop form or maybe tablet, or something. Being a monitor means the usefulness for me, ie being able to program outside, is kinda moot.

There is no trusted middleman who does the verification and takes the burden of accountability.

It's a great example of a project that needs accountability, but there's also thousands of other projects that need absolutely no accountability.

It's all relative to the stakes. The lower the stakes the less informed verification is needed.

The real trick is being informed enough on the boundaries to know where it matters. In the construction analogy, you need to be informed enough to know that a house is a bad idea since there are safety concerns. However building something small and non-load-bearing is probably fine to "vibe".

Eg there's no expertise needed to judge a garden trellis or 2 foot picket fence. It either works or it doesn't, and if it fails down the road there's no harm.

This boundary knowledge is the important bit imo.

Depends on the scope. I’m slopping some projects (graphics related) that I’d never have learned before. Mostly because of the huge lift, huge time investment, etc. is it quality code? I’m sure not. But verification of the application is quite easy, even if the underlying graphics and gpu shaders could be done better.

What I’ve learned is quite minimal due to the slop nature. However I’d have never even started this project in the past due to time, so it’s still fun being able to learn small bits of projects I’d otherwise not even touch.

Would be nice if they’d add in a simple back off retry mechanism.

which is hilarious because i was excited when i heard Digg was coming back. Many platforms are having a difficult time with bots, mass thread manipulation, etc. I'd be interested in a platform which attempted to fix that problem. I thought that was "so obvious" that i figured it was going to exactly be Digg's play. .. nope, just another AI play, as if we are missing those these days.

No idea if zero AI/bots is even possible, but at least an attempt would have me interested. A platform like Reddit/Digg of old, offering human connection, features aimed at less toxicity, etc.

Instead they give us this AI crap :s

While Americans Oppose AI Data Centers

I know it's more mixed and complex than this, but i think a big opposition is not to the data centers themselves but to their locations. Too often it feels like the centers are exploiting local resources and community infrastructure rather than paying their share or locating themselves in places that are less likely to cause problems to home owners.

The whole process feels indifferent or even adversarial at times.

This is very close to what i've been wanting. I'm slopping' an IDE to eventually explore this space as well. My general thought is AI should be monitoring what i write for code, and not only make suggestions, but allow using the human written code as a template for the final shape. Either rigidly matching it or treating it as a guide.

I feel like chat agents are missing the ideal UX, especially with the shortfall of their competence. Having to explain, rather than show, the agent feels like the wrong default UX.

Glad to see someone else exploring this space too!

And if you used it to poke around, you would also agree.

Would you elaborate? Not sure what you're describing

Nope, i'm on x20 and almost exclusively use Claude Code. I have a pretty bare bone setup with some custom hooks, skills, etc. I try to keep context lean so i don't like to add much stuff.

It's funny, i'm getting close to not caring anymore how much better a model is. I want it to be about as good as 4.8, but most importantly to be very good at following directions, style, etc. I really like Claude for that in general, but i've not measured in months so i'm not a good judge there.

I don't think i'll want to "hand off" code for several years, and so reviewing and iterating is becoming my #1 interest. A model that's as capable as 4.8 but 10x faster would be amazing for me.

Normally i'm first in line to try new models with Anthropic since i've clearly favored Claude in my personal tests, but this time i just don't think i care. 4.8 is capable, and even if the new one is more capable i don't want it to be slower (assuming it is). Note that i also (almost) use exclusively 4.8 on Max effort, so that also affects my speed comments.

Man did I despise that keyboard. I went from hating that to adoring my M3. Which feels good because I loved my MacBook before those butterfly switches, and again I love my MacBook now.

I almost left apple entirely over those stupid switches lol.

not open the music player?

I'd be fine with it doing that if it actually opened what i listen with. The OS can clearly see i spend 100% of my time in another music player (Spotify), opening Apple Music is at best a poorly designed UX.