nwienert

Sol does not follow instructions well at all. I've caught it multiple times a day now since release going off in incredibly bone-headed directions. It's so easy for it to over-interpret, make wildly out of scope changes, or just completely mis-understand what you're saying. The code it writes also is quite bloated still, and in my still-forming understanding I feel Fable is much more reliably smart. Sol is just more persistent and fast, so it often gets there quicker or after many tries, where Fable will just get it right from the start albeit more slowly.

Yea, two more years for the last 10.

The feedback I heard was definitely not that. The mistakes it makes are incredibly hard to predict, and they were lucky that no one was on the side of the road as they could've killed someone if a person had been there and they'd been a half second late.

That's the point. It's actually worse than a dumber system, because it's not even past the margin where it can go a month without a mistake, but you don't have to correct most days. The absolute worst possible scenario for safety.

A friend of mine just got one, ex-Chrome core dev so a fairly sharp guy, his one month review was that it was incredibly capable but had already done two maneuvers that would've led to an accident without intervention.

I built an iOS simulator simulator, though only for RN. Runs in browser but covers 100% of the API of RN, iOS UI, and the top 1k native libraries basically now. Been an ongoing agentic experiment of mine that's about ready to release.

Kind of fun, you can develop iOS and Android both without a build step and without a Mac even.

I agree we're at diminishing returns, but when brain scanning gets good enough you have a better dataset than the internet, Facebook is deep into that research.

I did initially through some miracle, as I wasn’t overweight. But after that year was up I now do grey market. Finnrick does testing which seems like a decent way to source if you’re looking that way. Not affiliated though and haven’t done a lot of background research on them.

Btw I have a couple autoimmune issues and found Tirzepatide strongly preferable to Semaglutide.

I've been posting about this including here for years now. I wrote a long post about it a while ago here and on Reddit. At time no one was talking about it, and actually my Reddit post was buried behind tons of others which was frustrating at the time given I had basically shared a partial cure.

Now if you search "reddit eds glp-1" or tirzepatide you'll see tons and tons of long threads of people all saying the same thing - it's the only thing that actually helps.

For me it was something of a miracle, I have two overlapping immune issues and it seems to just turn me into a much more normal, functional person. Including fixing my sleep.

Never was overweight beyond maybe ~15lbs btw when I started or took it, and the effects are 100% not because of just fasting or weight loss. I had tried keto and OMAD before, and been at healthy weight my whole life.

There's a definite auto-immune modulating mechanism and it's so strong it seems better than basically most first-class drugs. Even things like prednisone which are like nuclear weapons don't give me relief like Tirzepatide does.

Btw highly recommend Tirzepatide of the three GLP-1 drugs, for me at least it's by far the most effective and least side effects.

I easily burn through 3 $200 plans in less than a week. I am often using 4-6 sessions at once and do run overnight goals though typically 2 at once. Almost never use fast.

Claude plans are more generous now by about 2-3x but Anthropic slowed their tps a month or so ago so you’re not getting the speed. It’s flip flopped, Codex tightened it significantly recently and used to be more generous.

I do split between work, personal and OSS projects, which is why I have the plans.

It's so disrespectful to not give feedback to people you reject, companies that do it should be shunned.

Have some respect, be human.

I don’t think you’re really reading between the lines.

I’ve hired many asian developers anywhere from 1-4k a month.

I get a lot more out of a 200/mo subscription now in a week than I did from them in a month.

Now obviously in today’s world they’d be using a 200/mo subscription themselves. But it’s not like money is nothing, software development doesn’t scale down below 1k/mo for anyone competent even in the poorest areas.

I somehow take the opposite on almost everything here.

4.8 xhigh or max has a slight edge on 5.5 xhigh, for very complex logic perhaps it loses but it's just better in almost every other way, especially code quality. GPT is a slop machine outputs way too much and over-abstracts, plus its communication is so bad in comparison.

Fable was for sure a step above GPT, I tried them both against a few of the same hard tasks and it was not a small difference.

I actually was a Cursor advocate / CC hater (go back in my comment history), and now I use only TUI coding harnesses.

To start a big part is just the efficacy of them, which comes down to the model and the harness logic itself. CC is good, it's sub-agents, loops, background jobs / agents, skills/hooks/etc have typically been pretty far ahead though others are constantly catching up.

But you're sort of missing something. I use iTerm, so to me it's not the TUI itself, it's iTerm. And while it's imperfect, what I get is this:

I can open and close sessions nearly instantly and tile and window and tab them as flexibly as I want, plus it's a system I'm familiar with in terms of shortcuts etc. Has my configured theme, fonts, etc all set up. Every GUI app is different, every TUI app has half of the UI already incredibly familiar to me, it's not "just text", it's iTerm.

That also means they all are the same - I run Codex and Claude and pi side by side, and i switch between them with no overhead and minimal mental model shift. Sure, different harness does suck but that's the same issue with GUI just with an additional new layer to learn.

Smaller thing is because it's all text, there's no limits on my ability to copy things out. And it's a really fast text renderer that can render tens of thousands of rows efficiently. Many GUIs have various dialogs, unselectable areas, virtualization, or just slow past a point. I trust my terminal scales.

Just a few reasons.

It's the data. To do RL.

You have to have the models the create tools for you to paint.

I have a side project which is an experiment to build an interesting quick UI for local AI. As part of it I want a very very specific, interesting look involving shaders, animations, and so on.

I was trying to just get a prototype in place by prompting and it was going nowhere, just constant yo-yo'ing and never really getting what I wanted. This also was quite de-motivating and I found myself "yelling" at the model.

So I told Codex:

- Make this API first-class in our framework, with easy parameters (it had been sort of a hacked low-level thing)

- Add hot reloading to our system so I can edit it without any state loss or refresh

- Give me more knobs (X, Y, Z) so I can tune everything here as I need

- Add a HUD that lets me also drag sliders to tweak the same things

And I got my desired look within a few seconds.

The principles of good design and products have always been this btw, you need your feedback loop to be as tight as possible. Good design has always come from the ability to iterate incredibly fast, your brush needs to move precisely with your hands, and can't have delay from the time you put it down to the time the stroke shows up.

Apple does rein them in heavily, they push back on specs all the time, somewhat effectively.

If you decry bloated web apps and use Chrome on their Mac... there's Safari. It's far more efficient and has a far snappier UI.

Safari is better than Chrome and FF in enough ways I'd argue it can be considered the best of the three, even to people in tech. The dev tools are just way behind.

Bun has never really been well run. Every feature it had was full of bugs and gaps. And every release fixed a few but broke others.

They released more major features and breaking changes in their last patch release than most software sees in two major versions.

I've been using it just as a script runner and npm package manager basically, and it's incredible the amount of work you have to do to find "good" versions. We've had patch versions suddenly freeze on install more than once, we couldn't upgrade for quite a while due to this. I think they broke postinstall scripts with trustedDependencies entirely two minor versions ago - not a mention in release notes, and somehow no one reporting it in GH issues. In 1.1 or so you could get Bun to do trustedDependency builds in postinstall, and then after that you couldn't. I looked around for release notes and saw nothing mentioned. It's been broken for months.

No counterfactual there either though.

It's always changing, but this is the start of my default prompt:

https://gist.github.com/natew/fce2b38216edfb509f7e2807dec1b6...

I've had 0 issues with Codex once it adopted it. I use it for Claude too, which seems to also improve its continuation.

It was revised for friendliness based on the Anthropic paper recently, I'd have been a lot less flowery otherwise.

With one paragraph in your agents.md it's fixed, just admonish it to be proactive, decisive, and persistent.

The way I’ve come to think of LLM is that what the produce in a single reply even with thinking turned up, is akin to what you’d do in a single short session of work.

And so if you ask it to do something big it will do a very surface level implementation. But if you have it iterate many times, or give it small pieces each time, you’ll end up with something closer to what a human would do.

I imagine the pelican test but done in a harness that has the agents iterate 10+ times would be closer to what you’d expect, especially if a visual model was critiquing each time.

4.5 was clearly better than .6 and .7. Like, clear as day.

.6 is some sort of quantized or distilled .5 with a bit more RL, and the current .5 is that same cost reduced model without the extra RL.

Agree on all counts, 4.5 was a monster, 4.6 a clear regression, and then 4.5 was dumbed down so I moved on.

It's significantly worse on Mac than iOS, which gives you the answer. On iOS it's fine, even good. I prefer it, as a designer. On Mac it's a mess, and obviously spent less time baking.

Minimax is nowhere near Opus in my tests, though for me at least oddly 4.6 felt worse than 4.5. I haven't use Minimax extensively, but I have an API driven test suite for a product and even Sonnet 4.6 outperforms it in my testing unless something changed in the last month.

One example is I have a multi-stage distillation/knowledge extraction script for taking a Discord channel and answering questions. I have a hardcoded 5k message test set where I set up 20 questions myself based on analyzing it.

In my harness Minimax wasn't even getting half of them right, whereas Sonnet was 100%. Granted this isn't code, but my usage on pi felt about the same.

I found it worse, in a very clear way.

4.5 is better than 4.6 though in practice. 4.6 was purely a cost savings change with enough benchmark gamification to look better.