Zsolt Ero

Building and

[Zsolt Ero @hyperknot]

@petergostev Isn't it because they get dumber and dumber as their context window grows? Did you try asking them to regularly trim it?

[Zsolt Ero @hyperknot]

@yongfook Listmonk + AWS SES. Was a massive pain to setup before agents, not anymore.

[Zsolt Ero @hyperknot]

@damonchen How do you end up with domains like https://t.co/v2BMWihXEP and https://t.co/kgH7kF5h2j?

[Zsolt Ero @hyperknot]

@teortaxesTex How are they able to offer Sol 5.1 and Opus 5.5 at current prices (in subs)? Is it just a financial hack before IPO, or they really found a way to serve these models so cheaply?

[Zsolt Ero @hyperknot]

@theojaffee Wow,
OpenFreeMap usage really exploded with agents. Nice work!

[Zsolt Ero @hyperknot]

@zeddotdev Its the best editor, but Git is really unusable. Your number one issue on GitHub, please look into Git, we cannot use an editor without reliable Git in 2026.

[Zsolt Ero @hyperknot]

@mitsuhiko Even before, I explicitly asked it to use one-off python commands for complex edits before, it was way more token efficient.
Actually, happy to see this development.

[Zsolt Ero @hyperknot]

@pidotdev @mitsuhiko @badlogicgames https://t.co/E1IQnZ5pZJ

[Zsolt Ero @hyperknot]

Dear @pidotdev team (@mitsuhiko @badlogicgames)

I opened my 11th meaningful GitHub issue: auto-closed like always.

I've spent at least 10 minutes on each of those issues and many have contributed to meaningful discussion or PRs.

No matter how much I love pi and the core team's values, its such an extremely negative experience to try to contribute to this project.

[Zsolt Ero @hyperknot]

Opus 5.5 is the first model in a very long time which I really enjoy using. It doesn't want to over-engineer, happy to remove and simplify, and is generally an nice partner to work with.

Last model which I enjoyed so much was Opus 4.6 from almost a year ago.

[Zsolt Ero @hyperknot]

Ok, a question I wanted to ask: what monetisation do you see working for agenic harnesses / chat apps, other than selling tokens? OpenCode seems to have found a PMF on Go, but what if someone doesn't want to compete with you? Is there a market for staying small in this field today? Like selling a desktop app as a single developer? Or its all big fish eats the small fish today, peak capitalism and small players either die or get aqui hired to a frontier lab?

[Zsolt Ero @hyperknot]

That's how you create generational wealth in 14 hours.

(After spending a decade of your life tackling difficult problems.)

Thank you for the inspiration @tldraw!

[tldraw @tldraw]

Here you go, link in the next post

[Zsolt Ero @hyperknot]

@elithrar @Cloudflare That GitHub issue is a goldmine for custom outdoor styles and the PoC by Joe Hawes is open source.

[Zsolt Ero @hyperknot]

@mitsuhiko I hope there'll be an open model by a lab who cares about code quality. I wonder if a fine-tune might be enough, like Cursor did with Kimi 2.5.

[Zsolt Ero @hyperknot]

Opus happily finished the UI refining step, with visual feedback in a Playwright loop.

Do LLMs not feel pain when looking at this?

[Zsolt Ero @hyperknot]

@zephyr_z9 What is capability doubling? Have you talked with software developers who care about the code quality? Show me one person who'd say that capability doubled over Opus 4.5, release almost a year ago!
It got maybe 10-20% better over a year, and actually got worse in some aspects.

[Zsolt Ero @hyperknot]

@mitsuhiko Exact same experience here. TL full of videcoded 3D demos meanwhile it writes worse code than most GPT models before it. Also it writes terrible prose.

[Zsolt Ero @hyperknot]

@scaling01 How much of it is dedicated to vision tasks in your opinion?

[Zsolt Ero @hyperknot]

OpenAI researchers be like: 25% of the human cortex is dedicated to vision, why not do the same in GPT-6?

Mixture of Experts could also let vision parts stay dormant for non-vision tasks, so it doesn’t add much inference cost.

[SkalskiP @skalskip92]

Astra can do segmentation

this is pure VLM result. no expert models (like SAM) were used

No other VLM even comes close to this quality

- high effort
- avg input tokens / image: 2,052
- avg output tokens / image: 4,685
- avg cost / image: $0.255
- median time / image: 78.3 s

↓ GPT-6 Astra segmentation deep dive