jascha_eng

tbh AI is great at vague strategic decision making with a low risk bias. I wouldn't mind working for Claude

huh? because im curious what they used? Claude Code and codex take completely different approaches. The core loop of feeding generation and having a bash tool is entirely the same. If you want to start building your own thing I'm sure there is good setups out there to start from rather than going from zero.

Ofc I can just fork gemini-cli if I want my own version but like I don't think that's what OP meant by build your own thing and customize it.

How did you build your own coding agent? What language/framework did you use?

It used to be that DevOps was the movement of merging Ops leftwards so that Devs own the Operation of their software.

And that doesn't work with a simple prompt?

Can you give an example? And more curious about what you do with the resulting code afterwards I imagine its gonna be a big chunk then?

Is this useful? I feel like the problem is usually not that the model isn't capable of achieving what I give it, but the way it does it. Especially if originally I didn't 100% know how I would do it myself the model often takes weird paths through the code base, takes shortcuts that end up in weird feature interactions or pulls in a dependency without weighting if it could've been done without that.

I haven't really found a good way to solve this other than:

1. Produce an initial PR fulfilling all the requirements I knew at the start

2. Chat with the model about any weird snippets I notice and talk through alternatives

3. Simplify anything that I think is overengineered or plain unncessary

Sometimes I restart all over with more precise requirements but then it sometimes makes different mistakes/takes different shortcuts.

In practice the earlier I review the better the end result imo, so /goal seems very unproductive to me?

Until the classifier is wrong or also prompt injected. the classifier is just as vulnerable as the model itself is. Yes it is harder to break but trying to make a nondeterministic tool deterministic by adding another nondeterministic one on top just reduces the chance of something going wrong.

Tbf as long as that chance is low enough it doesn't matter in practice, but I have definitely seen the classifier approve things that were questionable, and I've also seen it decline things that were obviously okay.

But a pixel is quite a bit more expensive no? At that point you can consider an iPhone?

Sad I have a 6 year old oneplus and was looking for a new phone somewhat soon, would've considered them again for sure. Any alternatives? They always had a reputation for me for being a great no fuss, little bloat and simply fast android phone.

I mean it might lead to better performance on the model side. So the tokenizer is better but more expensive.

Aside of the claudeisms and the obvious AI smell, it overexplains everything and doesn't come to any useful conclusions. It's just not a good post.

The nudge to think about both "tokenization as variable" as well as actual tokens consumed per task is still good.

Slowly improving the UX on my SQL review/approval tool: https://github.com/kviklet/kviklet

Also finally closed the first real customer on it recently!

I want to get through a large chunk of the open issues the next few weeks and then spend some time building agentic capabilities for it. I believe a central place to configure database access for your dev team without having to share passwords and with sensible review policies should also help e.g. if claude needs to access production data to validate a premise.

Still have to figure out the right UX though not sure the agent should have the exact same review requirements that a human does. Maybe it needs to be configurable separately

im talking about anthropics pricing

Apparently not? At least our it guy said after 150 people or so you have to pay for enterprise which is pay per token for everyone.

Not including their best model in a max subscription would otherwise be truly a good reason for once to consider going back to openai for me. I'll at least try it.

It's great software in the sense that it makes a shit ton of money though. In the end software that doesn't get used and doesn't make any money but has no bugs is not valuable either.

Not saying that this is the trade off you have to make but if you have a working mode in place that achieves usage and money somewhat consistently i can understand being hesitant about changing it to optimize for less bugs instead.

I mean these were all solved before I assume so 100% not the same human ofc but models are expected to be good at a variety of code bases while human can specialize in one and learn. I think it's fair to compare to an individual that is used to working on a product.

I'm more interested in how fable would do

Can't wait for China to pull ahead so we have an end to this bullshit

Shorter sessions more often doing a /clear etc. save a shit ton of tokens. I pay 100 bucks a month but barely use 30% of it most weeks.

I mean if you have an idea or project that convinces you its worth to found a company, a thousand bucks is not a great hinderance. If the 1000 bucks are a hinderance you probably shouldnt found a company. It's a decent filter if you are serious about it.

idk I founded a UG within a few weeks recently. Yeh it costs me maybe a thousand or so in notary fees but that's it.

This mostly reads as a comparison between Opus 4.7 and 4.1 it would be more interesting if they reran the experiment against a team of humans with 4.7 and see how much the humans still improve the results today.

yes I agree with this, more granular going back, letting me interrupt where it went off the rails, or even editing file reads myself etc would be lovely. Ingesting parts of other conversations would also be cool!

The benchmark improvements actually look pretty damn nice tho!

They kind of have a little diagram explaining the steps I imagine every single step in that to basically be it's own Claude code session.

yup its all vibes. And anthropic is winning on those in my book still

None of his math really checks out. Building a piece of software is or at least was orders of magnitudes more expensive than maintaining it. But how much money it can make is potentially unbounded (until it gets replaced).

So investing e.g. 10 million this year to build a product that produces maybe 2 million ARR will have armortized after 5 years if you can reduce engineering spend to zero. You can also use the same crew to build another product instead and repeat that process over and over again. That's why an engineering team is an asset.

It's also a gamble, if you invest 10 million this year and the product doesn't produce any revenue you lost the bet. You can decide to either bet again or lay everyone off.

It is incredibly hard or maybe even impossible to predict if a product or feature will be successful in driving revenue. So all his math is kinda pointless.

FWIW TJ is not your average vibe coder imo: https://www.linkedin.com/in/todd-j-green/

In september he burned through 3000$ in API credits though, but I think that's before we finally bought max plans for everyone that wanted it.

Doesn't look as bad as I expected tbh. Sure some stuff could be better but I've seen much shittier vibe coded projects (including my own). I'd be more interested in their workflows and testing pipeline though. They ship pretty often but Boris still says he has 10+ PRs a day. I would be really curious what triggers a release, since it doesn't seem like every PR is released. I'm also curious how large their PRs really are.

There is a big difference between:

Build plugins

and:

Add 3px padding in line 5

if you claim "No code is written by humans anymore"