Astra is barely usable even on the $100 plan. And that is if it doesn't just burn through 80% of your weekly quota in a couple of hours by continually expanding the scope of the task you gave it - while not noticing the failing tests that are right in front of it.
Opus is at least actually usable even on the small plan. The main downside is its insane writing style, but 5.5 seems to address that somewhat. Otherwise, you can just use your $20 OpenAI plan to have Luna de-slop Opus' prose, which seems to work fine.
I have a Claude Code hook that calls codex for a code review on commit time (Codex is set to Astra Medium) and it's been pretty good in general. It sometimes hits the 5hr limit but most of the time it provides really good feedback and because it's a completely different model it's mostly complementary to what Fable/Opus do themselves. IMHO it's been $20 well spent.
...but the few times I've tried to use codex for a moderately difficult task it burned through its limit extremely quickly.
The issue is that in order to have the agent write good code, you need to implement standard SWE best practices. But that also means a lot of manual intervention in terms of writing specs, checking acceptance criteria, and reviewing code. So you end up spending a lot of time on managing your agent, which means you won't get a 1000% productivity gain, you get maybe 50 or 100, possible less in some areas and with some issues.
A 1000% productivity gain is quite possible on solo greenfield projects.
At work, with a team and code reviews, the 50%-100% figure seems much more likely.
This can probably move towards the more spectacular productivity gains as the AI's output becomes more reliable, people realize this, and less time is spend on code review and cleaning up the output.
You get 0% productivity gains if you are careful and actually reviewing the code the LLM produces. The only way to actually get the massive productivity gains that AI bros claim is to throw quality out the window.
I'm not sure that Claude's "Realigned the shape of the load-bearing ownership gate to reduce the blast radius of the design contract; confirmed, not assumed" is more meaningful than "fix".
> I'm not sure that Claude's "Realigned the shape of the load-bearing ownership gate to reduce the blast radius of the design contract; confirmed, not assumed" is more meaningful than "fix".
For PR/commit descriptions, I mainly use Claude Sonnet 4.5. It isn’t perfect, but it produces significantly less of this weird gibberish than 5.x models or even Opus 4.x do
I also use an iterative process in which it writes the description, I read it, and then either manually edit it or ask it to make changes
I use Astra at Very High, and shit is still bad. It doesn't actually understand anything, so it often says things which are clearly not needed to be stated. Recently, I've learned that I have very high standards for these things. For example, "fixes" as a commit msg just is NOT acceptable and would never fly where I work.
A few things, but generally more chain of thought before generating a response. So the model is tuned to think more. Given that what it outputs for this task is a summary of its thinking, tuning it to think more will just make a more verbose, less useful commit message.
Tune your model parameters to what is right for the task, not the highest you can afford.
Astra is such a mixed bag. It makes some amazing reviews and sometimes architecture suggestions that I like. But it’s also lazy and will just make up things.
Have another model (or even another instance of the same model) review the output of the first.
Models will hallucinate. They are also quite good at spotting hallucinations in other models' output (with some more hallucinations thrown in). With a threshold for confirmation, and a few iteration loops, you arrive at a fixed point where every claim is supported.
Honestly you probably want a model that has only been trained on language and literature. Nothing from online discourse.
And even then… writing is personal expression. Here people are talking about commit messages. That’s fine but AI doing writing for anyone and I WILL NOT READ IT unless it’s literally basic tech manual.
We read to hear and engage with people’s thoughts. If someone outsources that to AI then they should be shunned.
So long as there are in fact those things, and so long as it didn't sneak something else in there at the same time, it being just on the knife-edge between sense and word-salad is better than "fix".
Buuuuuut far to often it says it fixed a bug I reported, when it only touched one superficial failure mode rather than the root cause.
Yesterday's issue: Why is zoom/pan randomly failing? It told me it was because it was applying a transformation matrix with every input and sometimes JavaScript gave it a non-invertible matrix (why?) which then propagated NaNs everywhere and you can't update a matrix filled with non-numbers.
Why was it doing that in the first place? Seems to be because it's too motivated to perform quick wins and not sufficiently motivated to do good engineering.
Good thing this was just a game editor. Spiky intelligence: superhuman on some dimensions, total noob on others.
It is, because it includes the key words such as "ownership gate" and perhaps others, which makes it infinitely more informative than just "fix", by pointing at what was in scope, and what wasn't.
I've never had Claude spew that out in a commit message. That being said, I also use my tools properly. Your comment sounds like the type of thing someone who puts "don't make any mistakes" at the end of a prompt and gets upset when it doesn't come out perfect. Like, someone adding "make it better than AAA" is going to do something. At this point, it's really just telling on yourself that this is the output that you get.
Mildly disagree. "Fix" is exasperating but immediately tells me I need to look at the diff. With unconstrained Claude spew, I need to wade through three levels of deep fried LLM-speak before realizing...I need to look at the diff
"Deep fried" is the perfect analogy. LLMs have been trained on themselves so many times that their output is the linguistic equivalent of many rounds of JPEG compression.
I don't get this thread. Just tell the agent to use conventional commit messages and to keep it nice and tight. Are you all just raw dogging agent output with no alignment/conventions?
> Just tell the agent to use conventional commit messages and to keep it nice and tight.
I have this in my claude.md along with guidance on (not) writing comments but it’s still dumps multi-paragraph comments of Claude speak everywhere.
I think a lot of us are talking past eachother, but what I and I think others are complaining about is doing things well, and being minimal. LLMs don’t really do that yet, they can’t look at and simplify a codebase well, even with guidance. They always seem to add rather than subtract. And eventually it becomes an issue. I think they were actually better in this regard with like Opus 4.8, and are now getting even worse (but score better, and are more autonomous).
At home I use Codex, and it’s better as far as language goes at least.
Anyway, it’s like the logic piece necessary to do good work is still missing, and hurt by recent reinforcement fine-tuning. And it leads to massive bloat since more comments = higher scores, but walls of text and the tokens (or cognition if you’re the sad human reading them) aren’t free.
They are still making billions from highly sticky enterprise customers, which is more than enough to survive. Otherwise, Microsoft would have gone out of business years ago.
Sure, but then they're out of the "growth" stock category they so badly wanted to be in, with its higher valuation multipliers, then in the stodgy "keep the lights on" niche. If a company makes $10B this year, and you had a way of knowing it was going to make $10B every year from now on, there's not much incentive to invest in it.
Note that this doesn't make for a bad company! You can run a $0-profit company indefinitely and give employees nice jobs for providing nice services to your customers. There's nothing wrong with that. It's great! But you don't get the same multiples for that arrangment as you do for a company that's growing quickly. Lower multiples = stock drops = One Medium Wealthy Asshole Called Larry Ellison, and that's something I simply do not think he could tolerate.
Nothing about this is surprising, something like this was always going to happen, because you - or an LLM for that matter - can always find a line of motivated reasoning that justifies any course of action. One would have to be extremely naive to believe that "alignment" provides any kind of actually robust guardrails. Simultaneously, we have seen decades of security vulnerabilities. Unless your testbed is truly and fully physically airgapped, any current SOTA model will find a way to break out.
The reality is that the current approach to AI safety is little more than a fig leaf, but you also won't be able to put the genie back in the bottle, because the technology is simply too powerful to abandon. There is always going to be someone developing it further from now on.
So, the real question is what a novel and actually effective approach to AI safety looks like and how to get there.
>effective approach to AI safety looks like and how to get there
It's probably impossible, or it may only be possible in hindsight which means it's already too late.
If for example you make an entity smarter than you all you can do is hope it will be safe. Any entity smarter than you has more freedom of choice of actions than you do, or at least the ability to explore them. A common example here would be a 2D entity trying to contain a 3D entity. The 3D entity can simply rise up and over any line you draw to stop it.
And that's for cases where intelligence has a will to break out of it's box. It doesn't even need that. Instrumental convergence can sit around and build solutions until one of them passes the "don't do bad things" classifier.
And lastly, you're assuming that models will remain expensive to train well into the future. If the cost drops significantly you can expect someone to develop an intelligent but completely unhinged model at some point. And it may even be AI itself that creates it. There is no natural evolution that occurs that natively makes safe models.
It's honestly a smart decision. They already have access to the common market just like Norway. Becoming a full member just means being subject to a EU governing body that has proven itself to be largely ineffective and incompetent, sometimes reaching the realm of "outright harmful" with things like Chat Control. Why would you want to be a part of that, especially if you would also be paying for that privilege?
I love how people comment very confidently on shit they clearly know nothing about.
Iceland is in the EEA. It already has to comply with pretty much all EU regulations (with some exemptions, such as their fishing thing). It is also a part of Schegen, freedom of movement, etc.
Unlike other countries, it doesn't get to vote on those regulations, so it doesn't even get the veto power.
I honestly think it may make more sense for Iceland to not be a full EU member, the fishing industry may be too important for a fairly small island nation, etc.
But it has nothing to do with the reasons you fantasized about.
I would say it’s the other way round: being in the common market means that they are already subjected to EU regulations, without having any representation in the governing bodies however.
But seriously, the fact that this started as Crypto-adjacent should have immediately disqualified them for serious use.
reply