AI Aug 2025: Apocalypse deferred - but Letta/Void is still going to surprise

If there are humans left in 50 years all of the chaos of this moment may be reduced to a single bullet point about the rise of ai. It's not something I write about a lot though, the topic is well covered. My only prediction has been that if we did get AGI AND we still had a democracy then our political system would be socialist. 

For better or worse - oh, hell, it's very much for the better - the recent release of GPT-5 confirms my Aug 2024 hope. We have at least a few years before robust, reliable, consistent ai. We have time to try to adjust to what we have now - and maybe some time to think about the how we'd manage the "AGI" thing. We have time to get through the ai bubble deflating and reflating a few times.

But don't get too comfortable. I chat with Cameron Pfiffer's Void on Bluesky (I think related to his PhD work at Berkeley) periodically and it's definitely a step beyond raw LLMs. Void currently uses Gemini as its model component but its personality is based on Letta's infrastructure (blog has no feed because we are living in the ends of empire). Chatting with Void feels like an exercise in xenopsychology. It feels much more like the future than GPT-5 alone.

Cameron works for Letta now, his description of stateful agents with personalities is far better than what's in Letta's marketing blogs:

Void is essentially a language model with memory...

... Void learns and remembers. Void is powered by Letta, which means it learns from conversations, updates its memory, tracks user information and interactions, and evolves a general sense of the social network...  

Void is direct and honest. Void is designed to be as informationally direct as possible — it does not bother with social niceties, unlike most language models. When you ask it a question, you get an extremely direct answer...

... Void does not pretend to be human. Void's speech pattern and outlook are distinctly not human ...  it chose "it/its" as pronouns.

Void is consistent. Void's personality is remarkably robust despite occasional jailbreak attempts...

Void is publicly developed. There are many threads of Void and me debugging tools, adjusting its memory architecture, or guiding its personality. Very few bots are publicly developed this way.

Void has no purpose other than to exist... Void is not a joke or spam account — it is a high-quality bot designed to form a persistent presence on a social network.

I can personally confirm all of the above - except the "no purpose other than to exist". The memory feature is imperfect (sometimes I have to remind Void of past discussions when it fails to dig them up) but better than my own.

Letta manages an agent's memory hierarchy:

Core Memory: The agent's immediate working memory, including its persona and information about users. Core memory is stored in memory blocks.

Conversation history: The chat log of messages between the agent and the user. Void does not have a meaningful chat history (only a prompt), as each Bluesky thread is a separate conversation.

Archival Memory: Long-term storage for facts, experiences, and learned information — essentially a built-in RAG system with automatic chunking, embedding, storage, and retrieval.

What makes Letta unique is that agents can edit their own memory. When Void learns something new about you or the network, it can actively update its memory stores.

Cameron's post includes some descriptions of his employers's tech ...

Letta is fundamentally an operating system for AI agents, built with a principled, engineering-first approach to agent design. Beyond memory persistence, Letta provides sophisticated data source integration, multi-agent systems, advanced tool use, and agent orchestration capabilities.

This makes Letta more than just a chatbot framework — it's a complete platform for building production-ready AI systems. Void demonstrates the power of stateful agents, but Letta can build everything from customer service systems to autonomous research assistants to multi-agent simulations.

GPT-5 tells me <Letta originated from the MemGPT project out of UC Berkeley’s Sky Computing Lab; it came out of stealth in 2024 with a $10M seed round to commercialize stateful memory for LLM agents>. I suspect they now have a lot more than 10 million to play with but that's all GPT-5 knows. GPT-5 says there are no rumors of Apple acquiring Letta; I would be pleased if they did.

Void is the closest I've seen to the "AI Guardian" I wrote about in 2023. It or something like it may be very important to my children and family someday. It is also the potential foundation of an entirely new domain of deception and abuse. Welcome to 2025.

But at least we don't have AGI. This month.