Pinned Tweet
Worried you'll wire it wrong? Don't buy anything yet.
Click "Wire it for me" and watch an ESP32 get a mic, a speaker and a screen in 40 seconds. Then do it yourself, step by step, until the dog shows up.
No hardware needed to try it. github.com/beasaltfish/bo…
Click "Wire it for me" and watch an ESP32 get a mic, a speaker and a screen in 40 seconds. Then do it yourself, step by step, until the dog shows up.
No hardware needed to try it. github.com/beasaltfish/bo…
Last time it was a breadboard and a hopeful LED. Now it has a face, and you can just talk to it. 🐶
A screen for under $2, an ESP32-S3, and a dog that sits, shakes and spins when you talk to it. No buttons, no app, just your voice.
Simple commands never touch the LLM. JEV es them and fires the animation, so "sit" feels instant. The LLM only gets the open-ended stuff, like "good boy!"
What went wrong on the way:
• Fried my first screen. VCC and GND swapped for one second. Backlight on, picture black. The backlight survives, the chip behind it doesn't.
• The speaker crackled, but only while talking, and only with the screen wired. The screen's clock pin sat next to the audio clock pin, and its sharp edges leaked into it. Fix: drive the screen pins at the weakest setting.
• Stripes across a still picture. I reused a buffer before the DMA finished reading it.
The brain runs on my computer; the board just listens, speaks and shows.
3
2
31
1,084
Do you use an AI agent to draft your tweets and replies? Does it actually sound like you?
A general model doesn't know what you've actually said. So I gave mine my own history: my timeline syncs into my own database, and every draft has to cite the past post it's drawing from.
I review every draft, then send it myself. The agent never posts.
A general model doesn't know what you've actually said. So I gave mine my own history: my timeline syncs into my own database, and every draft has to cite the past post it's drawing from.
I review every draft, then send it myself. The agent never posts.
1
0
6
628
Last time it was a breadboard and a hopeful LED. Now it has a face, and you can just talk to it. 🐶
A screen for under $2, an ESP32-S3, and a dog that sits, shakes and spins when you talk to it. No buttons, no app, just your voice.
Simple commands never touch the LLM. JEV es them and fires the animation, so "sit" feels instant. The LLM only gets the open-ended stuff, like "good boy!"
What went wrong on the way:
• Fried my first screen. VCC and GND swapped for one second. Backlight on, picture black. The backlight survives, the chip behind it doesn't.
• The speaker crackled, but only while talking, and only with the screen wired. The screen's clock pin sat next to the audio clock pin, and its sharp edges leaked into it. Fix: drive the screen pins at the weakest setting.
• Stripes across a still picture. I reused a buffer before the DMA finished reading it.
The brain runs on my computer; the board just listens, speaks and shows.
A screen for under $2, an ESP32-S3, and a dog that sits, shakes and spins when you talk to it. No buttons, no app, just your voice.
Simple commands never touch the LLM. JEV es them and fires the animation, so "sit" feels instant. The LLM only gets the open-ended stuff, like "good boy!"
What went wrong on the way:
• Fried my first screen. VCC and GND swapped for one second. Backlight on, picture black. The backlight survives, the chip behind it doesn't.
• The speaker crackled, but only while talking, and only with the screen wired. The screen's clock pin sat next to the audio clock pin, and its sharp edges leaked into it. Fix: drive the screen pins at the weakest setting.
• Stripes across a still picture. I reused a buffer before the DMA finished reading it.
The brain runs on my computer; the board just listens, speaks and shows.
As promised, the test 🤖
I asked it to prove it's alive in three words: "Two copper wires."
My robot used to hear and speak through an old Android phone. Now an ESP32-S3 on a messy breadboard does that job, and the brain runs on my computer.
In its own words: no eyes, no legs, hopeful LED.
4
2
41
1,791
I have a Nintendo Switch that I was planning to sell, but I've changed my mind now.
🗞️Bruno St John has gotten a full copy of iPadOS 17.7.11 running on an original Nintendo Switch 🔥
onejailbreak.com/blog/ipados-17…
0
0
3
79
One channel of this week's #2 repo only answers once you've starred it, per its README. There a star is a toll, not an endorsement, and a ranking by stars gained can't tell them apart. How many stars on your own list are tolls? x.com/i/article/2117…
1
0
0
842
Is "AI drafts + I review and send" still genuine engagement on X?
My setup: AI picks tweets worth replying to and drafts in my voice. I edit every reply and hit send myself. Nothing is auto-posted.
Doomscrolling eats my day, but growing here means showing up. Does this feel ou run a similar workflow, what do you use?
My setup: AI picks tweets worth replying to and drafts in my voice. I edit every reply and hit send myself. Nothing is auto-posted.
Doomscrolling eats my day, but growing here means showing up. Does this feel ou run a similar workflow, what do you use?
3
1
19
2,192
Silicon Valley calls this the AI Golden Age, but for most people, it just looks like a new tax on data security, hardware, and tokens—while modern worker rights become the first thing to be sacrificed.
"We're in humanity's golden age." I'll likely be fine whatever happens. But watching my kids grow up next to a war and with a society that doesn't know how to finance itself, let alone work with AI, I sometimes struggle to see the gold. dark.ronacher.eu/2026/10/31/gol…
0
0
0
66
Challenge: name a game you wish you could play inside Claude Code while it works. Within reason, I'll build the best one and ship it to Idle Arcade.
(No promises, but I'm curious what you'll come up with.)
(No promises, but I'm curious what you'll come up with.)
Inspired by Marko. Now you can play a fish game inside Claude Code while it works. 🐟
idlearcade.beasaltfish.com
4
1
26
1,161
Claude Code has mods now: small plugins that add panels, buttons, commands, even games inside the terminal.
What they are, how they work, a 12-line example, and an game example I built with them. x.com/i/article/2116…
0
0
1
40
I can’t even take a nap
I lead product for Claude subscriptions. Today we're renaming our Max plans, and I want to explain why.
When we launched our top plan in April 2025, it had a single limit: 5 hours. It gave you 20x Pro's limit, so we called it Max 20x. When we added weekly limits to Max 20x, we set them to 10x Pro's (the 5-hour limit stayed at 20x).
I get that some new users assumed 20x applied to both limits. We could have done better here, and I'm sorry.
The Ultra plan keeps the weekly limit (10x Pro), but we're removing the 5-hour limit so you can keep building without getting blocked mid-session. This starts today and rolls out in stages. To help prevent fraud and abuse, a temporary limit applies until we've verified your account. You'll get an email when your limit is removed.
We've got more cooking and can't wait to share it. Thanks for building with Claude 🙏
1
0
1
118
Asten retweeted
I lead product for Claude subscriptions. Today we're renaming our Max plans, and I want to explain why.
When we launched our top plan in April 2025, it had a single limit: 5 hours. It gave you 20x Pro's limit, so we called it Max 20x. When we added weekly limits to Max 20x, we set them to 10x Pro's (the 5-hour limit stayed at 20x).
I get that some new users assumed 20x applied to both limits. We could have done better here, and I'm sorry.
The Ultra plan keeps the weekly limit (10x Pro), but we're removing the 5-hour limit so you can keep building without getting blocked mid-session. This starts today and rolls out in stages. To help prevent fraud and abuse, a temporary limit applies until we've verified your account. You'll get an email when your limit is removed.
We've got more cooking and can't wait to share it. Thanks for building with Claude 🙏
When we launched our top plan in April 2025, it had a single limit: 5 hours. It gave you 20x Pro's limit, so we called it Max 20x. When we added weekly limits to Max 20x, we set them to 10x Pro's (the 5-hour limit stayed at 20x).
I get that some new users assumed 20x applied to both limits. We could have done better here, and I'm sorry.
The Ultra plan keeps the weekly limit (10x Pro), but we're removing the 5-hour limit so you can keep building without getting blocked mid-session. This starts today and rolls out in stages. To help prevent fraud and abuse, a temporary limit applies until we've verified your account. You'll get an email when your limit is removed.
We've got more cooking and can't wait to share it. Thanks for building with Claude 🙏
We’re renaming our Max plans:
Max 5x → Max
Max 20x → Ultra
Max is still $100/month, and Ultra is still $200. If you’re a subscriber, there’s nothing you need to do.
We’re also removing the 5-hour limit for verified Ultra accounts, starting today with most existing unts keep a temporary limit until they’re verified, which helps us prevent fraud and abuse. You'll get an email when it's removed.
Weekly limits remain unchanged. View our plans here: claude.com/pricing
803
142
4,074
836,363
Still wincing when the 1-hour prompt cache expires and 300K tokens get re-sent to Claude at full price? 💸
Claude Code now compacts your context while you're away, before the cache dies. Your next message resumes on a small summary instead of a giant cold re-read.
It's on by our settings right now:
1. Run /config
2. Find "Idle compaction"
3. If it says true, you're covered. Nothing to do.
4. If it says false, flip it to true.
You'll know it worked when you come back to a line like: "Compacted while idle, before the prompt cache expired."
Claude Code now compacts your context while you're away, before the cache dies. Your next message resumes on a small summary instead of a giant cold re-read.
It's on by our settings right now:
1. Run /config
2. Find "Idle compaction"
3. If it says true, you're covered. Nothing to do.
4. If it says false, flip it to true.
You'll know it worked when you come back to a line like: "Compacted while idle, before the prompt cache expired."
1
0
9
493
Claude Code has mods now: small plugins that add panels, buttons, commands, even games inside the terminal.
What they are, how they work, a 12-line example, and an game example I built with them. x.com/i/article/2116…
What they are, how they work, a 12-line example, and an game example I built with them. x.com/i/article/2116…
4
0
7
6,643
Asten retweeted
Want your robot to react to simple commands instantly? Don't send them to the LLM. Put a tiny classifier in front.
What I learned doing it with a 0.8B decision model:
• Make "pass" the biggest class. A wrong move is a crash; a wrong pass just means the LLM answers, a bit slower.
• Split train/test by base phrase, not by row. Otherwise "Stop!" and "stop" end up on both sides and inflate your score.
• 490 rows, 30 min on free Colab, ~200 ms per call on a Mac mini.
Fast enough to knock itself over. Full write-up in the replies 👇
What I learned doing it with a 0.8B decision model:
• Make "pass" the biggest class. A wrong move is a crash; a wrong pass just means the LLM answers, a bit slower.
• Split train/test by base phrase, not by row. Otherwise "Stop!" and "stop" end up on both sides and inflate your score.
• 490 rows, 30 min on free Colab, ~200 ms per call on a Mac mini.
Fast enough to knock itself over. Full write-up in the replies 👇
Built a fast, local AI brain for my DIY robot using @UnslothAI 's new Decision Model training guide! 🤖⚡️
Instead of heavy chat models or cloud APIs, I fine-tuned Qwen3.5-0.8B into a deterministic decision engine (JEV/Clef style) running entirely local on my Mac Mini M4.
The rdware (bot-reborn):
⏱️ ~200ms reaction time (you can see the server logs in the video)
🎯 >90% accuracy on precise robot motion commands
🔒 100% local & private
How it's built:
Hardware / Brain: An old Android phone riding on a cardboard saddle, driving a toy car's motor driver over WebUSB (FT232H async bitbang mode). No intermediary server for local control.
Training: Used Unsloth to train a custom dataset (robot-simple-commands) on Qwen3.5-0.8B. The model outputs structured decision probabilities directly instead of generating free text, which keeps inference blazing fast.
Everything is 100% open-source. Robot, Code, dataset, and model weights are all live below!
Pull requests and stars are super welcome. 👇
📦 Repo: github.com/beasaltfish/bo…
📊 Dataset: huggingface.co/datasets/beasa…
🧠 Model: huggingface.co/beasaltfish/ro…
2
1
1
775
Growing an audience on X might be one of the most immediate feedback loops out there.
Building products often means shouting into the void, which can be discouraging. But simply sharing my work and engaging more here has changed that equation.
As a software and hardware sely every single day, I love sharing what I learn.
Drop a line if you’re into AI hardware/software—let's connect ! 💡
Building products often means shouting into the void, which can be discouraging. But simply sharing my work and engaging more here has changed that equation.
As a software and hardware sely every single day, I love sharing what I learn.
Drop a line if you’re into AI hardware/software—let's connect ! 💡
10
2
29
3,829
Inspired by Marko. Now you can play a fish game inside Claude Code while it works. 🐟
idlearcade.beasaltfish.com
idlearcade.beasaltfish.com
if Claude had chat themes, I'd never switch back🌴
concept, made with Claude
3
1
6
1,814