Shibel K. Mansour (@SHxKM) header
Shibel K. Mansour (@SHxKM) profile picture

TV news anchor gone software developer. Incubated entrepreneur. Build @musicbutler_app.

Joined March 2013
Photo by Shibel K. Mansour (@SHxKM)Photo by Shibel K. Mansour (@SHxKM)Photo by Shibel K. Mansour (@SHxKM)Photo by Shibel K. Mansour (@SHxKM)Photo by Shibel K. Mansour (@SHxKM)Photo by Shibel K. Mansour (@SHxKM)Photo by Shibel K. Mansour (@SHxKM)Photo by Shibel K. Mansour (@SHxKM)Photo by Shibel K. Mansour (@SHxKM)Photo by Shibel K. Mansour (@SHxKM)
ethos can now manage agents in other projects too, including scheduling. This is going to be very useful for @mattpocockuk style /retros, regularly improving test suite runtimes, and triaging simple tickets.
After 48h I realized a flip side of operating a single agent in a never-ending session is that you're constantly jumping between topics. The obvious solution (lacking even in Anthropic's and OAI's apps) is being able to quote instead of clarifying. So I built it into ethos.
66
After 48h I realized a flip side of operating a single agent in a never-ending session is that you're constantly jumping between topics. The obvious solution (lacking even in Anthropic's and OAI's apps) is being able to quote instead of clarifying. So I built it into ethos.
I couldn’t wait so I told Claude to implement the spec. It’s…an interesting experience so far.
1
1
136
@VictorTaelin you should consider for UniiChat. I only pass the agent the ID of the message and it can `zoom` on it if it needs to.
35
UniiChat’s (previously OptChat) spec was updated yesterday. One decision is to use Haiku at xhigh effort instead of Sonnet @ medium for compaction. An Opus agent Ethos dispatched to test this on 300 messages hints that may have been about costs rather than quality.
ethos (my custom harness built on @VictorTaelin’s UniiChat spec) can now spawn agents, schedule tasks (executed by ethos or dispatched), update itself, send me notifications, receive voice messages and reply with audio as well. 48 hours from 0 to this. What a time to be alive.
106
ethos (my custom harness built on @VictorTaelin’s UniiChat spec) can now spawn agents, schedule tasks (executed by ethos or dispatched), update itself, send me notifications, receive voice messages and reply with audio as well. 48 hours from 0 to this. What a time to be alive.
1
1
2
263
So looks like I have my own harness now. ethos. It’s (by now loosely) based on @VictorTaelin’s spec for OptChat. The product he released is probably much better. But this one is mine.
1
94
Glad to see me and Victor took the same approach on some of the things. I told my ethos about his spec which was updated today:
1
1
70
And some things I should consider
14
Ok maybe I’m taking this a step too far but is there such a thing nowadays? It is indeed liberating not having to think about context at all.
I couldn’t wait so I told Claude to implement the spec. It’s…an interesting experience so far.
1
35
I couldn’t wait so I told Claude to implement the spec. It’s…an interesting experience so far.
Hi, I'll soon open my harness This is the last text wall about it. I promise! (but not the last in general, I have plenty of time for that now :D) First, sorry for this silly side-quest, I know some expect more from me. But is it really stupid if it works? I really needed this to exist and I think many of you will love it too ????-Chat is like ChatGPT, except the chat itself is OptMem: an append-only log, compressed as a nested memory tree of varying resolutions. This allows the AI to always see a constant-space overview of your entire life, and it can zoom at any part of it when it needs more details, giving you a practically endless chat. To me, this is the closest thing to real infinite context I've ever experienced. Don't get me wrong, though; this is not actual infinite context, it isn't a new model, nor any deep breakthrough. It is just an embarrassingly simple memory management system, that turns out to be surprisingly effective in practice. You can keep talking to it for years and years, and it won't get overwhelmed ("context-rot"), because the context never grows large. You can ask it to do different tasks, and it will absorb the complexity and deliver via sub-agents, 100's if needed. Years later, when you ask something, it will recall everything from any campaign, with remarkable accuracy. It just works. Also, since there is only one chat, every connected device (browser, terminal, mobile) is unified. You can give it a task from your phone, and it will execute that task on your computer, as long as ????-Chat is open there. This is super convenient. And it improves on several other small things that I found bafflingly stupid on other harness, but that's not the place for this :') ???? is great for work, but it is also incredibly convenient for mundane stuff. For example, I can ask it to buy my groceries today, in the same tab where all my work is being done, and the manager will do so, without being overwhelmed. I can keep talking over and over, and it absorbs everything. When I pass it an instruction on how to behave, it sticks. If I revoke it later, the last word wins. Then, 1 year later, I can just ask it to remember anything about me, and it will. Because it is all in the memory tree, a few hops away. In short, think of it as the manager of your agents. It is the centralized point of contact between you and 100 agents. The bridge between your life and the AI. Forgo your temptation to have many open tabs, as this intuition comes from limitations that no longer hold. Just keep dumping problems at ????, let it learn about you, absorb the complexity, and execute your tasks. I wish I had this while developing Bend, and I'll use it as my only harness moving forward. I'll never use other harness again, unless they incorporate this (and I suspect they will, as certain one already did...) You can use it on ????Chat[.com], with Bend Credits (it charges API costs + 10% cc fee), or you can bring your own keys. I also posted a prompt below so you can replicate the setup in your own harness and use your own local agents for free. That said, using our credits allow you to unify all costs in one place, and I'll be happy and have numbers to show when selling my soul to VCs in SF 🥳 The site is actually already up, but ????-Chat is adding some last features I want it to have before launching. I've vibe-coded it in the last 2 days using itself, and everything just works! I'm commanding it on my phone rn, since I want to visit an Aquarium with her today. I'm free!!!! That's all. I hope this silly thing improves your life too :D
1
128
Shibel K. Mansour retweeted
Hi, I'll soon open my harness This is the last text wall about it. I promise! (but not the last in general, I have plenty of time for that now :D) First, sorry for this silly side-quest, I know some expect more from me. But is it really stupid if it works? I really needed this to exist and I think many of you will love it too ????-Chat is like ChatGPT, except the chat itself is OptMem: an append-only log, compressed as a nested memory tree of varying resolutions. This allows the AI to always see a constant-space overview of your entire life, and it can zoom at any part of it when it needs more details, giving you a practically endless chat. To me, this is the closest thing to real infinite context I've ever experienced. Don't get me wrong, though; this is not actual infinite context, it isn't a new model, nor any deep breakthrough. It is just an embarrassingly simple memory management system, that turns out to be surprisingly effective in practice. You can keep talking to it for years and years, and it won't get overwhelmed ("context-rot"), because the context never grows large. You can ask it to do different tasks, and it will absorb the complexity and deliver via sub-agents, 100's if needed. Years later, when you ask something, it will recall everything from any campaign, with remarkable accuracy. It just works. Also, since there is only one chat, every connected device (browser, terminal, mobile) is unified. You can give it a task from your phone, and it will execute that task on your computer, as long as ????-Chat is open there. This is super convenient. And it improves on several other small things that I found bafflingly stupid on other harness, but that's not the place for this :') ???? is great for work, but it is also incredibly convenient for mundane stuff. For example, I can ask it to buy my groceries today, in the same tab where all my work is being done, and the manager will do so, without being overwhelmed. I can keep talking over and over, and it absorbs everything. When I pass it an instruction on how to behave, it sticks. If I revoke it later, the last word wins. Then, 1 year later, I can just ask it to remember anything about me, and it will. Because it is all in the memory tree, a few hops away. In short, think of it as the manager of your agents. It is the centralized point of contact between you and 100 agents. The bridge between your life and the AI. Forgo your temptation to have many open tabs, as this intuition comes from limitations that no longer hold. Just keep dumping problems at ????, let it learn about you, absorb the complexity, and execute your tasks. I wish I had this while developing Bend, and I'll use it as my only harness moving forward. I'll never use other harness again, unless they incorporate this (and I suspect they will, as certain one already did...) You can use it on ????Chat[.com], with Bend Credits (it charges API costs + 10% cc fee), or you can bring your own keys. I also posted a prompt below so you can replicate the setup in your own harness and use your own local agents for free. That said, using our credits allow you to unify all costs in one place, and I'll be happy and have numbers to show when selling my soul to VCs in SF 🥳 The site is actually already up, but ????-Chat is adding some last features I want it to have before launching. I've vibe-coded it in the last 2 days using itself, and everything just works! I'm commanding it on my phone rn, since I want to visit an Aquarium with her today. I'm free!!!! That's all. I hope this silly thing improves your life too :D
every memory solution I tried sucks, so I rolled my own it is a plug and play prompt that gives any agent infinite memory db is an append-only log. compression on the spot (nap, don't sleep). nothing is ever deleted, memory context is constant-sized, details fade with age
74
29
8
868
39,306
Surely Silo would’ve made it to the top 5 if it was over?
Andor was amazing and it’s so cool seeing it in the top 50.
117
Shibel K. Mansour retweeted
Replying to @dhh
1. Celever Hans effect. LLMs are excellent brainstorming mates but if you let them take over the wheel in production, you're in for a rude awakening. The models are actually pretty bad at architecture. You're good at it. 2. Architecture was never really about *writing* the code.
1
1
261
One thing I didn’t take into account when I moved a production app from Heroku to @Railway is that they literally have 0 support. Their forum is now filled with reports on apps being down for more than 12 hours because of a Redis issue. 0 response. Don’t move to Railway.
1
2
512
Amazing customer support by @posthog so far, and I’m on their free tier. Most bugs fixed within 48 hours. (I’m on the free tier)
Really liking @posthog’s fresh attitude towards product analytics, but it seems that the product (at least the hosted solution) is still buggy. Even submitting a bug report is buggy.
1
266
Really liking @posthog’s fresh attitude towards product analytics, but it seems that the product (at least the hosted solution) is still buggy. Even submitting a bug report is buggy.
1
1
1
503