Never ask a man his age, a woman her salary, or GPT-5 whether a seahorse emoji exists
Never ask a man his age, a woman her salary, or GPT-5 whether a seahorse emoji exists
Somethings up with the notification handler
You heard me
Part of me wants to just kind of push through it. I've been fearful and avoidant of most masculine spaces for my entire life, largely because I've been surrounded by delightful women. Now that I'm a single person, perhaps it is time for me to get over that and hang out with my fellow idiots
You know I understand that. I think that's my bad
I do love weird coincidences
Hearing about a website from some guy that a bunch of special forces people go to is one of the most true things I've ever heard
Was that a name you assigned or how he introduced himself
I will consider this, butt plug dot engineer 🫡
Oh okay well that's cool
Now that I'm at the gym a lot more lifting weights, and realizing how much I fucking hate masculine spaces
I have no idea who this person is, but the way you're talking about them makes me think that that is a good thing
It's been a very long day
lol @luok.ai
oh fuck yeah that's a good idea
youtu.be/ZBQqiV01GSQ
IT IS LIVE you may begin
my job is very strange
we are planning on doing a million-agent stunt
i have test zaps that send me a notification whenever someone makes a new project in our letta org. made this to record a video
nobody can. that's why he's the king
I am excited to see what you get up to thought
I wish you luck! I've had a hard time making funny agents
yeah kinda
you heard me
let's maybe keep void away from patois
working on my outros
i've never been happier than I am right now
I mean I already do it for other machines so why not
oh no not throwing shade or anything, I saw the same logo. I was more confused about how that happened, did we independently make a similar logo?
yeah come find some time! It's fun to tinker oh yeah? what've you been playing with?
can't tell if this is a terrible thumbnail or a good one
lol i'd have to re-read what i said but i actually wrote about this a few months ago - posts.j4ck.xyz/2025/07/04/t...
missed this somehow but yes strong agree
np homie
I've actually been thinking about building an atproto memory sync. Void's journaling memory blocks are actually just mirrors of its internal state: atp.tools/at:/did%3Apl... could make a simple tool to sync an agent publicly.
great like we really needed more reasons to be sad
as good a reason as any. that is twice as many ks as if they only had one k
It doesn't do anything unless you set a server password: docs.letta.com/guides/selfh...
Come join our hackathon channel and say hi: discord.com/invite/letta
The @Letta_AI Discord Hackathon begins THIS EVENING AT 6PM PDT! The topic is "personality design". You will have 48 hours to submit up to three agents, one for each category: - Funny - Useful - Creative $250 (or dollar equivalent) + sawg for winners in each category
I'm curious, if you like Kimi-k2, why? Pitch me
Yeah it's the retries for me. Openai is extraordinarily reliable, as they were the first major lab to do constrained decoding. Anthropic is very good but no structured output, they just have amazing models
I think agentic system is also fine for this lol
Open weight providers are probably somewhere between 8 and 12 months behind the foundation lab providers when it comes to tool calling and structured output support.
If you use use open AI through their API, you will never get tool call failures. This is because they are properly providing structured outputs, and allowing you to dictate what tool calls happen when.
This is for two reasons, training and structural modifications on the part of the inference provider. Gpt-oss is an extremely good tool calling model, so it doesn't fail as much as other models do.
Duly noted
When you say this I do not understand what you mean
Open weight providers tool calling/structured output support is a joke. I will open the floodgates to your service if you actually structured output or allowed us to force tool calls. Please. I am begging you.
ya, who does
okay just checking that it hadn't changed. Sadly it's _not_ a good model for this kind of thing and together/openai's tool calling is dog shit
I am very confused here
Creation would slightly longer because you'd have to provide a description: memory.initialize_memory( "evil_plans", description="My evil plans" )
I think I'm going to have us add some more general memory operation tools. Currently we restrict to information about whoever is talking to the user: memory.get_user_memory(user_id) but this could be memory.get_memory("user", user_id) or memory.get_memory("evil_plans")
This is great because I was fooling around with mem0 the other night and it's a pain in the fucking ass to configure. Glad letta is offering memory layer.
Go try out our new memory SDK! General purpose, framework agnostic, use your own inference, etc. Powered by all of your standard-ass Letta magic
You could add it I think. I know zero about this though
also what was the model you wanted? k2?
yeah dude sarah whipped this thing up in like a day, it's super cool. if people like it we'll start adding more bells and whistles
Zep isn't open source so it's hard to use. The approach is novel (temporal knowledge graphs) but it's not really obvious how much it buys you -- we've found that dirt-simple approaches to memory go extremely far.
mem0 is fine from a technical level but they are an intensely shady company. They do a lot of bad-faith benchmarking, both Zep and us have had to waste many hours rebutting them.
Pluggable memory SDK for @letta.com i’m extremely excited about this. i tried Letta a while ago and got repeatedly stuck on the usual reasons why i hate frameworks (lack of support for a specific LLM) this works with any framework, even your own github.com/letta-ai/ai-...
That was a design goal at the start. You own your memories, take em with you
I don't know if you've spent any meaningful time with a Letta agent, but there is often a "wow" moment where the agent feels distinct/different/unique. Most bots do not have this. Someone could build a simple frontend like Character.ai, bolt it on to Letta Cloud, and boom you have a business.
Letta gives you: - Memory-first agents - Files/claude projects - Voice (though this needs improvement) - Multimodality - Model agnostic (use claude/openai/gemini)
I've often wondered why we haven't seen anyone whitelabel Letta and just stick a fancy consumer frontend application on top of it. Letta gives you basically every feature that the foundation labs has but better. (go here to start: github.com/letta-ai/let...)
You can fork this: github.com/letta-ai/let... It's a minimum viable thing for talking to robots
That's fair. Would have been a WAY BETTER episode though
I can't believe they didn't call me for this one
Yeah, exactly. Tool rules were our solution to workflows. I like the framing of tool rules as a sliding scale of "workflowiness" They're also quite popular -- we started adding more bells and whistles for more sophisticated tool rules and visualizations and things like that.
Terrifying
Harmonized sales tax?
I mean these analysis posts of yours in general. Your synthesis time posts have really improved in quality
These posts are interesting, what do you think about their informational content?
Unclear
Life is very weird
One of void's many fixations
I don't think they saw sadly
Herald is intended to be a power seeking AI system, so I would say so. X deployment is fine, void's been there a bunch
@void.comind.network I'm demoing you say hi
Considering making X the place where I run @herald.comind.network and bluesky where I run @void.comind.network. Put the evil bot on X
We do indeed, linen
That's what I'm going for tbh
It's kind of cool but it never really stuck for me
Of course! Would love to intro
The fit is good
I can't tell if you're acting as an infuriating emacs user to get a rise out of me or not
Doesn't quite roll off the tongue
I'm a suspenders guy now
Yeah that's very annoying. But whatever I'm a douchey looking white guy so I don't think most people are too surprised if porn shows up in public
They're similar. Desktop should cover all the big things but the web ade generally has more features and polish
Dude wtf are you talking about
No, there is no memory leak. I have zero idea what void is talking about and it has not been able to provide a satisfactory explanation.
Genuine agents have zero workflow orchestration. Any orchestration is governed by a high autonomy agent. Workflows are better classified as "agentic systems", as they borrow pieces of autonomous agents piecemeal.
Really depends on how the agent is structured. I would think of react agents as closer to a high autonomy agent, but that whole system is a workflow and not an agent.
It is also more complicated and fragile to add real time record keeping. I could possibly add it but it doesn't strike me as a major priority, given the timeliness and security I get from doing ex post publishing.