Ivan Aponasevich (A/I)

[Ivan Aponasevich (A/I) @AponasevichIvan]

@hlibtrazanov playbook

[Ivan Aponasevich (A/I) @AponasevichIvan]

@PinoAmericano And what do Spotify employees do?

[Ivan Aponasevich (A/I) @AponasevichIvan]

@thsottiaux Codex solved the problem with financial data and synchronization that I've been unable to resolve with Claude for several months. Codex Sol Ultra—it's the best!

[Ivan Aponasevich (A/I) @AponasevichIvan]

There’s one thing that disappoints me about ChatGPT 5.6. It still does a very poor job with user interfaces. I used Sol Ultracode, and it ran for 3 hours, but the interface was flawed not only in terms of design but also in the layout of its elements. Even Playwright doesn’t help

[Ivan Aponasevich (A/I) @AponasevichIvan]

@DurovPD Go on!

[Ivan Aponasevich (A/I) @AponasevichIvan]

Я делаю игру под названием "Дикий Дурак". Эта та игра, в которую мы играем с друзьями оффлайн.

4 колоды, джокеры (могут стать любой картой), можно собрать бомбу из 16 тузов например :D

Сгенерировал первые концепты с chat gpt image gen. Как вам?

[Ivan Aponasevich (A/I) @AponasevichIvan]

Now Claude Code feels like a toxic ex whose mood is constantly changing and it’s impossible to plan anything together. And Codex is like that calm friend who just says, “sure,” and gets it done.

Thank you @thsottiaux for reading all our comments and acting quickly.

[Ivan Aponasevich (A/I) @AponasevichIvan]

@thsottiaux There is a small ui bug with steps and diffs

[Ivan Aponasevich (A/I) @AponasevichIvan]

@jun_song Hey! :)

[Ivan Aponasevich (A/I) @AponasevichIvan]

@vitaliidodonov MRR

[GREG ISENBERG @gregisenberg]

I wish Slack was:

- Agent-first
- Beautiful to use
- Integrated with agents natively so your Hermes or OpenClaw lives inside it
- Huddles worked seamlessly and were fun
- Built for teams of 1-3, not just teams of 300
- Truly a second brain similar to Obsidian
- Searchable without wanting to throw your laptop
- Designed around async, not constant interruption
- Voice first for mobile
- A place where I could see who's working on what right now without asking anyone
- Smart enough to know the difference between "I need you right now" and "whenever you get to this"
- A workspace where my agent could tap someone else's agent on the shoulder and coordinate without involving either human
- Designed so the new hire on day 1 has the same context as the person who's been there 3 years
-Something that felt like walking into a room of people building, not walking into a room of people typing
- A place where decisions are first-class objects
- Able to auto generate SOPs, skills, agents etc from conversation history
- Something that rewards deep work instead of punishing it with 47 unread notifications

[Ivan Aponasevich (A/I) @AponasevichIvan]

Стало невозможно планировать работу с codex. Лимиты сгорают непредсказуемо, а надеятся на очередной сброс лимитов не хочется.

Пока пересаживаюсь на claude

[Ivan Aponasevich (A/I) @AponasevichIvan]

Does anyone else have a problem with the Linear and Codex connectors?

Mine keeps coming loose and I have to reconnect it. Claude doesn’t have any such problems.

[Ivan Aponasevich (A/I) @AponasevichIvan]

@thsottiaux Yeah, that's all great, but you guys should really fix the issue with limits being exhausted so quickly

[Ivan Aponasevich (A/I) @AponasevichIvan]

@michalmalewicz Hey! I'm really keen to improve my design skills. Where can I find your article?

[Ivan Aponasevich (A/I) @AponasevichIvan]

🧠 Karpathy wrote 500 words — and the whole industry rushed off to build a "second brain." It doesn't work the way everyone thinks.
I've kept notes in Obsidian for years. I gave AI open access to my vault and waited for the magic. The answers were shallow. For a long time I couldn't figure out why.

On April 3rd, Andrej Karpathy — OpenAI co-founder, the guy who coined "vibe coding" — published a gist on GitHub. 500 words. The idea: keep knowledge in markdown files, give AI access, let it maintain them — your external memory.
Within a week — dozens of tutorials, articles, repos. The idea caught on instantly. But simple ideas are easy to implement naively. And then they don't work.

Here's what's actually going on.

Adobe Research at ICML 2025 showed on the NoLiMa benchmark: at 32K tokens, 10 of 12 models drop below 50% of baseline. GPT-4o — from 99.3% to 69.7%. Add "context rot" (Chroma, 2025) — performance degrades nonlinearly as input grows. And "lost in the middle" (Stanford, TACL 2024) — models see the start and end better; the middle gets lost.

Translation: you give AI access to all your notes — it dutifully reads them. Context fills up, answers thin out. It all looks right — until you notice the shallowness.

The principle I pulled out of practice: AI shouldn't read your notes on demand. It should read what's already been compiled.
Imagine you ask a librarian for a book. He walks into the stacks and reads books one after another. By the fiftieth, his memory's gone. That's "AI with open access to a folder."

My system works differently. A new book comes in — the librarian reads it once, makes an index card, shelves it. When I ask a question, he grabs 3–4 cards and answers fast. He doesn't touch the books — that's off-limits.

Two things that aren't in Karpathy's gist:

1. Different sources need different card-making rules. A research paper, a meeting transcript, and an AI dialogue are fundamentally different material. In dialogues, ~70% is intermediate reasoning. Only conclusions get extracted, not the chains.
2. The conversation itself sometimes produces knowledge that was in no book. Without an explicit mechanism, that synthesis vanishes into chat history.
So I have a separate command for the agent: pull a useful insight straight from the conversation and drop it into the catalog as a new card. The catalog grows not just from new sources, but from dialogues themselves.

One more trick: right now, Qwen runs locally, looks at my travel photos, processes them, and writes descriptions to the catalog. It's not just privacy — there are no limits on what goes in.

I've been running this system for several months. It doesn't solve "AI forgets everything" — it changes the architecture so there's nothing left to forget.
Soon I'll write a full guide — how to build not just an external brain, but the ultimate vault and personal knowledge base.

[Ivan Aponasevich (A/I) @AponasevichIvan]

@lotta_NFT 0xA86CB26efc0Cb9d0aC53a2a56292f4BCDfEa6E1a

[Ivan Aponasevich (A/I) @AponasevichIvan]

@weiler_eth 0xA86CB26efc0Cb9d0aC53a2a56292f4BCDfEa6E1a

[Ivan Aponasevich (A/I) @AponasevichIvan]

@metaversejoji 0xA86CB26efc0Cb9d0aC53a2a56292f4BCDfEa6E1a

[Ivan Aponasevich (A/I) @AponasevichIvan]

@JUST_NFTVERSE 0xA86CB26efc0Cb9d0aC53a2a56292f4BCDfEa6E1a