Allan

The future is bright.

@dmcgco I have a full on terminal with the most gratuitous ascii art if you just sit on my vanity site for 10 seconds. We absolutely can not have ascii fatigue yet.

Ok ok we’ve had our fun now but the ASCII and halftone effects have to stop.

@johnpalmer Mine is a SUV sized robot dog that you can sit on top of and it’ll ride you around town and stuff.

my billion dollar hardware idea is a NEO robot but it’s six inches tall and just hangs out on your desk

@johnpalmer Damn

New logo for Mesh, a laser company.

@attacless @usgraphics You, yes. I mostly infantilize products so I'll be stuck paying for Berkley Mono.

@Allan @usgraphics so did we pass?

ping just passed the vibe check @usgraphics use berkley mono for our next update?

@dmcgco Same issue.

There must be a simpler way to manage/setup multiple (5+) email addresses from different domains vs hooking up a bunch of separate google/biz accounts? I have a bunch of different addresses from different ventures and I feel like I'm missing a trick here.

@mschoening A small novelty but I gave mine access to a receipt printer. I now get a printout in the morning with my schedule and some todos.

Here are tasks I want to get done: - Kick off coding agents on real codebase (there are 4000 services that do this) - Read my email, draft replies, archive BS - Reply to Slack messages - Help me schedule things and make reservations - Write me little research reports on topics I care about - Grocery shopping - Organize my digital life - Cancel dumb subscriptions - Manage my personal finances and pay bills - Renegotiate contracts

What is the most useful thing you’ve seen an OpenClaw do? I love the tinkering. But, what does it actually do for you?

@JPEGuin FWIW, I'll buy it

Fool me once…

I'm scared to launch my new iOS app. I'm in my happy place right now coding. I know as soon as I hit publish, I have to start shaking my ass on tiktok for downloads.

@max_creating Woah, super impressive! I love this! We should race our agents!

@Allan Look at this. I think my approach is even faster / turbo x.com/max_creating/s…

@brycedriesenga @ZainMerchant9 @westoque That's probably a natural place to end up. It's also very tempting to fall back on AppleScript or things that aren't keyboard/mouse. Once it's both extremely competent and fast with input designed for humans, that'd be the idea.

@Allan @ZainMerchant9 @westoque I wonder if it's possible for it to tap in to app intents/scripts/shortcuts and default to those when possible for speed, but fall back to vision?

@LarryVelez Porsche AG had a very rough year financially so maybe instead a new AI agent division via acquisition of some idiot's pet project is in order.

@Allan Porsche's IP lawyers are aggressive, so start working on another logo.

@louis030195 I've seen but never tried. Looks impressive and I wouldn't be surprised if it's quite good. But perhaps because I'm twice as lazy and half as clever, it's a "just works" solution. There's no chat with the agent, nor will it execute code it writes on the fly.

@Allan did u try openinterpreter?

@FaithfulFirst That’s very roughly how it works now, although speed and ability of the model is driving which is used. Turbo uses two small local models.

@Allan Super cool, have you tried to mix models. A local one for regular fps and an event driven of a stronger more token/$ for bigger things?

Yes! This is what it does! Every run it updates a small SQLite database for each application with Icons/UI, Task Sequences (small sequences that can be replayed), and recipes (action patterns). In theory it should get smarter every time and I could share my "skills" with you and speed up your Turbo agent, if needed.

Skills is the same approach I went when using MacOS automation tools/control scripts. It really is the best approach I’ve found for making sure the agent knows/has a reference guide for whatever app/workflow it’s trying to perform. Add an agent that creates new skills based on user interactions and you got a self improving system right there

@KalraIshaan11 It started as fixed-tick and worked, but it was very token hungry. I switched to a reactive / event-driven loop. That doesn’t rule out continuous perception or delta tracking though. Just not built yet. Likely a can of worms but might be important.

@Allan Hey Allan, this is awesome. Quick question: how frequently does Turbo sample the screen (fixed FPS vs event-driven vs adaptive)? Also, have you thought about a “always-on” background mode that can persist without disrupting the user’s workflow?

Agree on speed. Turbo’s architecture is optimized around fast inference and persistent UI state, so it doesn’t have to relearn the interface. I made application "skills" portable too — and when they're in use, the Turbo agent is basically working at human-ish speeds. It's early and not optimized much yet. I bet I can get Turbo to work on some tasks at faster-than-human speeds. As for a "good multimodal agent": if speed is the goal (and it is, given the name), a single agent is probably the wrong approach. Turbo mixes local models and larger frontier models.

@Allan vision is correct technically but currently it's just too slow. tried to do this before and you need: 1. fast inference 2. a good multimodal agent that knows the UI of what you're automating github.com/bytedance/UI-T…

@grok @007Killpop I believe Claude Cowork / computer-use agents are tool-mediated (they ask tools to do the work) and turn-based, re-perceiving the screen each step. Turbo runs natively on macOS, stays stateful, and would probably win in a footrace.

@WillDeLonDove 🥵

The LLM that's responsible for planning receives 3ish key pieces of context. Mainly: (1) an optimized version of the current UI state, (2) a structured catalog of every detected element + its label, coordinates, role, description, ..., and (3) the task description. So: It gets the visual state plus a semantic map of what's clickable and where, which allows the model to output specific executable actions like "click element #12 at (245, 120)" or "type 'Projects'" rather than vague instructions — it's essentially planning against a known inventory of interactive elements. But also, Turbo tries to avoid calling the planning model when possible using some cleverness.

@Allan I’m curious how you’re prompting it, what context is used for the LLM to produce clear executable outcomes