Rahul Acharya

Scratching my own itch

[Rahul Acharya @fractalyza]

Sheesh! So many thoughts reading this:

- We got “AI Penn & Teller: Fool Us” before GTA-6
- There should be a “MagicBench” to evaluate this across models (early digging: doesn’t look like anything close exists on https://t.co/F6pikQXaxt)
- There has to be a product hidden here, waiting to be built?
- And can the converse be plausibly true? If LLMs can deduce tricks like these from first principles, they should be able to apply those same principles in other interactions and deliver “magical” experiences that genuinely surprise?

[Stanley Tang @stanleytang]

I have an insane magic trick that so far none of the models can figure out, including Mythos. It's a bullet-proof trick that I've shown to 100+ people, including magicians that couldn't figure it out.

It's not anywhere on the internet. Only way to know it is through first principles reasoning. Told everyone I'll believe in AGI when it can crack this trick.

Well... GPT 5.6 just did 😳

[Rahul Acharya @fractalyza]

Tried something I’d previously built a custom tool for and it worked fairly well just by prompting in the Codex app with this new ability.

1. Hacky way: Have the parent thread monitor the child for completion; it spins up an automation/cron that polls every 5–10 (configurable) minutes and just keeps going.

2. Better way: Callback; have the parent thread ask the child to ping it when done with a useful summary of its run, giving you duplex communication and long-running behavior that’s great for migration tasks.

[Guinness Chen @guinnesschen]

If you ever get tired of managing your Codex threads, just let Codex manage itself! Codex can now create threads, search them, organize them, pin the important ones, and spin up worktrees for parallel tasks.

[Rahul Acharya @fractalyza]

Yo! Artemis II wasn’t the only thing I saw going up and fast to the moon today, my Claude code(2.1.90) usage skyrocketed from 46% to 99% in under 5 minutes.
And I was surprised to see the initial 46% in the first place because I barely recall more than a few exchanges. @claudeai

[Jacob Posel @jacob_posel]

Hey @bcherny @claudeai

I'm on the $200/mo plan and blowing through usage instantly. Doesn't feel right. Is there any way to audit my account?

Unfortunately I have experienced several bugs with the Claude product and I fear my plan configuration is not correct.

Thanks

[Rahul Acharya @fractalyza]

@dabit3 @DevinAI @FactoryAI @droid It's great! Will share A "Copy all bugs", "Copy all flags", and "Copy all issues" button introduced here that copies the corresponding content would be pretty useful.

[Rahul Acharya @fractalyza]

@DevinAI code review is free* for personal repos and seems pretty neat.
Try: https://t.co/lxK17xqTza<owner>/<repo>/pull/<number>

Also tried @FactoryAI's @droid-bot before this, the tokens churned extremely quickly, had to pause the GA.

[Rahul Acharya @fractalyza]

The Xcode MCP popup in both Claude Code and Codex had been pretty annoying lately. Worked around it with an mcp-proxy and Xcode upgrade to 26.3 RC2 (17C528)

[Rahul Acharya @fractalyza]

How long is this in agent years? @grok what's the math?

[Rahul Acharya @fractalyza]

Interesting read. Made me think that the ability to comment on a public @AmpCode thread like on a gist, would be pretty cool!

[Ivan Leo @ivanleomk]

@AmpCode https://t.co/OcaV9Vyz63

Thread here if curious

[Rahul Acharya @fractalyza]

I know I'm going to come back and re-read this quite a few times.

[Taalas Inc. @taalas_inc]

24 dedicated people.
$30M spent on development.
Extreme specialization, speed, and power efficiency.

Today we launch Taalas’ first product. Check it out:
Details: https://t.co/88CA0XAL71
Demo chatbot: https://t.co/ec4ladcKnw
API: https://t.co/M3EkaxEqPj

[Rahul Acharya @fractalyza]

Address matching is an interesting problem by itself. STT struggling with addresses didn’t ever occur to me but I’m intrigued now that I know of it.

[Kaitlyn Zhou @KaitlynZhou]

Text-to-speech models can’t get your address right? Turns out you’re not the only one.

📢New preprint! State-of-the-art speech models get 44% of street names wrong — and non-English primary speakers suffer twice the error impact!

[Rahul Acharya @fractalyza]

cctip: If a claude code plugin expects a certain format, ask cc to do it instead of firing /command directly; it figures out the expected format and uses it.

Eg:
"Create ralph-compliant tasks based on the ralph plugin for each of these tasks and then ralph away" => 30 mins

[Rahul Acharya @fractalyza]

gg @getpeid! The bigger picture is interesting.

[Essential @essential]

Create apps shaped exactly around your specific needs and context.

That's what Essential Apps are.

You describe what you need. AI builds it. It appears on your phone's home screen, ready to use.

One billion apps for one billion people.

Beta starts today on Nothing Playground.

[NVIDIA AI Developer @NVIDIAAIDev]

Kimi K2.5 is now live on GPU-accelerated endpoints for free prototyping, so you can quickly start building with a frontier-scale multimodal model instead of just reading about it.

Want to get started? We’ve got you:

• Step-by-step tutorial
• Ready-to-run GitHub notebook
• First inference in minutes, not hours

📖 All available in the technical blog → https://t.co/UzCfQ6Y32e

[Rahul Acharya @fractalyza]

Wasn't expecting a local Suno this early in the year 🤯

[ACE Music @acemusicAI]

We're releasing ACE-Step-v1.5(2B), a fast, high-quality open-source music model.

It runs locally on a consumer-grade GPU, generates a full song in under 2 seconds(on an A100), supports LoRA fine-tuning, and beats SUNO on common eval metrics.

GitHub: https://t.co/Zydfhw9oUl

Key traits:
Quality: beats Suno on common eval scores
Speed: full song under 2s on A100
Local: ~4GB VRAM, under 10s on RTX 3090
LoRA: train your own style with a few songs
License: MIT, free for commercial use
Data: fully authorized plus synthetic

The music AI space lacks commercial-grade open models. Many creators are forced to rely on closed-source services, and can’t fully own, run locally, or fine-tune their own models. We want to help change that.

[Rahul Acharya @fractalyza]

How to exhaust a month's subscription in a day

[Google DeepMind @GoogleDeepMind]

Step inside Project Genie: our experimental research prototype that lets you create, edit, and explore virtual worlds. 🌎

[Rahul Acharya @fractalyza]

I'm finding myself running git bisect much more often now than ever before.

[Rahul Acharya @fractalyza]

TIL @Stanford has lectures on iOS development with SwiftUI at: https://t.co/rkzj6tz8Oa

[Gregor Zunic @gregpr07]

We open sourced our tiny agent-sdk🧙the easiest way to build Claude Code like agents (with any provider)

All the value is in the RL'd model, not your 10,000 lines of abstractions🦓

pip install bu-agent-sdk or 🔗↓ or put it to Claude Code and 1 shot your own in any language.