Danny Ki

AI revenue automation for sales, marketing, analytics & ops

[Danny Ki @thatsgonnah]

i only merge prs that delete more lines than they add

[Danny Ki @thatsgonnah]

anthropic really put fine print on a free reset

5h limit resets, but only fable usage gets wiped from your weekly total

resetting the whole counter might bankrupt them lmao

[Danny Ki @thatsgonnah]

@JamesMalsawm paying for 6 ai subs across different providers rn and it’s always some bullshit

one day they’re messing with the limits, the next they tweak a config and your model wakes up lobotomized

[Danny Ki @thatsgonnah]

@Corporate_CJP absofuckinglutely

[Danny Ki @thatsgonnah]

fable 5.2 is stupid fast

makes every openai model feel like it’s on dial-up

[Roni @ronisarkar_exe]

Opus 5.5 in the past two weeks.

[Danny Ki @thatsgonnah]

@proximalcoast my chatgpt dot spent 2 weeks playing email ping pong with bybit’s bot over my locked account

asked it to find emails for actual humans in compliance instead

sorted in 2 days lmao

[Danny Ki @thatsgonnah]

@MystiqueMide @NousResearch congrats, safety first

[Danny Ki @thatsgonnah]

turns out the cheaper model caught 45% more prompt injections

ran 391,678 requests across 8 Jev-models for $49.62

Perplexity Decider 1.2 was 5.4x cheaper than jev 1.13, but had 80 false positives vs 0

Jev 1.14 still wins on long inputs tho

would you ship one model or route between them?

[Danny Ki @thatsgonnah]

an agent can parse a webpage value perfectly and still have no proof where it came from

[Danny Ki @thatsgonnah]

same prompt, same context, same blind spot. a reviewer agent can turn one bad assumption into two passes of agreement

https://t.co/yVG1WLyfPf

[Sergii Makarevych @sermakarevich]

A research/review of how to fix common agentic issues with architectural tricks:

- early failure
- hand-off efficiency
- noisy context
- more agents same miss-takes
- big context
- LLMs tend to agree
- faster/cheaper
- learning https://t.co/hvzobBxcK3

[Danny Ki @thatsgonnah]

an AI agent that needs the same context pasted five times is a clipboard with a subscription.

[Danny Ki @thatsgonnah]

@eliadele been using @orca_build for 3 months, best agent IDE imo

project folders, 1-click Linear/GitHub issues, worktree delivery, browser, mobile emulators, easy multi-sub switching

custom orchestration lets agents monitor workers live and redirect bullshit mid-task

[Danny Ki @thatsgonnah]

@liornedogarsky likewise bro 😎

[Danny Ki @thatsgonnah]

@patrick_kuhnke_ kimi is goated, qwen is solid too

[Danny Ki @thatsgonnah]

@soham_nayak04 ram altman log off

[Danny Ki @thatsgonnah]

@724real in some of our projects, we stopped stuffing the entire chat history into context and switched to background memory updates that extract key facts instead

cheaper, faster, and in some cases we’re seeing up to 2x faster responses

[Danny Ki @thatsgonnah]

opus has an insanely good design harness and is great at orchestration too

but let it write code on its own and it’ll happily leave shit unfinished, then confidently tell you “all done bro”

that’s why you absolutely need codex for code review, it always catches something opus missed