swe_dima

Would love to see benchmarks on visual tasks.

UPS is also full of holes. No verifications whatsoever.

You can purchase a label from a reseller and then ship using that label to a different place at a different weight several times. Battled with this at a previous job.

Was considering buying ads on ChatGPT. A damning thing is that yours audience then are users too cheap to buy a sub...

does this mean RAM prices can go down now? Please?

According to the metrics the "fast" mode is not any faster...

I don't understand why they don't update the status page. Been going on for 1+ hour. Is it counted as "uptime"?

my prompt (doesn't help):

> Go easy on the comments, only add comments if there's a big gotcha that is not clear from the code itself, or if something in another place is going to cause a side effect. Code should be self-documenting. When in doubt, don't add a comment at all. If you do have to add a comment, make it short and on point. Comments should show history of code changes or functionality, only comment on the current state (or not at all).

My feeling that as it becomes a better coder it becomes a worse communicater. It's overfitting for coding benchmarks, while communication style is harder to quantify during training.

And no matter how often I tell it to stop adding comments it just can't help itself.

It's why I love React Native and over the air updates. So I don't have to deal with this roulette.

In my experience SVGs are still too hard for LLMs.

I gave Fable a jpeg and asked to draw an SVG, using a loop that renders the SVG into an image so Fable can inspect it.

Results looked like drawing of a 5 year old.

I think by default they check out a new connection when obtained a job and then release it. In comparison with some async languages like JS where a connection is only checked when a query is about to be executed

My concern is number of database connections. In the example it's 100 fibers per worker, at that rate you are going to exhaust db connections sooner. Happy to be wrong.

I also love them - makes spotting AI written content easier. People debated implementing complex algorithms how to watermark AI content - but luckily AI just settled on it by itself.

(sorry for a bit of gaslighting)

Gemini flash lite family of models currently has the best ratio for price/speed/intelligence for understanding images, no real alternative AFAIK

It's scary relying on Google's models.

I have a very price sensitive workload that used to run on flash 2.5 lite - it's deprecated now.

The replacement 3.1 flash lite is a lot more expensive, but now also has a sunset date.

3.5 flash lite is even more expensive.

So the price is rising and you have no choice but to keep paying more and more.

About the same, using cloudflare. For some very bizarre reason Vertex has a very low non-adjustable quota.

the writing is on the wall for it, I have switched to gemma-4-26b-a4b.

At least in benchmarks, it scores higher and is faster.

"You were the chosen one! You were supposed to destroy them, not join them!"

if I travel and expect to be working I now bring a CO2 meter

If agent is asking a question it means it stumbled on something tricky and often an important architectural decision has to be made.

Even if I am staring at the screen at the time, it's often impossible to study code and make a judgement call that quickly.

My point being, even for opt-in it can be too short - highlights how little consideration they put into it, just releasing slop in production.

Is it possible to use their subscription pricing with Opencode?

many years ago when Hanami was just getting popular I remember doing benchmarks against Rails when it comes to SQL and was unpleasantly surprised when Rails' ActiveRecord ended up being much faster, despite "speed" being advertised as one of the advantages :-(

Not sure what niche it's going to occupy: too expensive for it's intelligence category.

Pleasantly surprised that it costs as GPT 5.5, thank god for the competition.

I found Ruby LLM to be surprisingly good - in terms of usability it's close to Vercel's AI framework.

It tries to strike a balance between working out of the box and being flexible... which has its challenges, still nice overall.

One big real-life pain I experienced is that caches don't always work, e.g. for xAI, since it only supports completions API and thought signatures are returned wrong.

Love Vite, but always felt sorry for them because it was not clear how they can make money, the whole VoidZero thing felt like a stretch.

It's one of those things that always stopped me from building cool tools - you have to make a living somehow.

So I am happy for the team of builders that they were able to receive the deserved payout and sustainability.

I still don't understand how people use git worktrees with Docker. You need a full database and etc. For me it's simpler to have multiple checkouts.

the video on this post looked like a Tiktok - everything is jumping, I could barely grasp what was going on

You may remember the argument that you can build an AI app and it continues to improve as models improve and costs go down?

Well, looking at OpenAI / Google / Anthropic we see crazy cost increases, such that it might invalidate your unit economics.

Cheering for Chinese models!

Flash family but costs like a Pro. $9 vs $12 for output.