once you dive into this and there are more then 2 people and 5 tickets, the challange shifts from this gimmick feature to company knowledge recall for the agents. I built this feature in a demo 3 months ago with the gemini live model + internode. That was already good enough if you know what you're doing. The conclusion was that correcting bad human recall during the meeting is more important than shifting tickets around. Updating tickets after the meeting async is much better.
Balazs gives a glimpse of how our day to day is impacted by having an organizational memory.
It is quite challenging to pinpoint the greatest values when you have it, as it improves almost every aspect of a work day.
Still my favorite sentence he wrote: "The agent isn't impressive because of the model - it's impressive because the context underneath it is real."
We wrote a blog post about how our own team uses Internode day to day, from turning meeting transcripts into tickets, to weekly change logs drafted in two minutes, to asking the company brain who owns what.
Read it here: https://t.co/R56qdmOEEX
We built Internode because we kept forgetting what we said in our own meetings.
In the last 2 days we made ourselves the first client. 3 ideas, 2 decisions, 4 new tasks captured all of which we would have lost before. The company is starting to remember itself. DM us for a demo!
Happy Labor Day! 🎉
Humans forget 70% of new information within a day. The future of work separates memory from humans. That's what Internode is built for - and we just launched for free. 👉 https://t.co/RSnbqQE82O
"Wait, didn't we talk about this last week?"
Most teams hear it weekly. The reflex is to blame someone's memory. The real problem is that every tool we use treats humans as the database.
Memory-native flips it.
Happy Labor Day! 🎉
Humans forget 70% of new information within a day. The future of work separates memory from humans. That's what Internode is built for - and we just launched for free. 👉 https://t.co/RSnbqQE82O
@victorcardenas would love to exchange notes (what works what doesn't, short term v long term, cost, scale, etc). this is what we focus on and build daily for everyone @internode_ai
there exists an experiment to answer who is right, Eric or Stuart: Eric would be right if the information evolution can be measured without interruption from the photon's properties all the way to the state where it becomes "color" the experience. Currently both guys are making their "bets" on whether this can be done or not. Eric assumes that the final state where experience exists can be deterministically described with math, Stuart would disagree. Careful though, there is nothing that proves that the measurable state of the nervous system equals experience. Experience itself is not defined, and that is exactly where @ericweinstein makes a shortcut by calling it a consequence of physics, which is just an assumption (or his "bet").
Levie is right that context is the dominant challenge. But most people hear "context" and think "give the agent access to more files."
That's not the hard part. The hard part is making sure the agent doesn't redo the same reasoning every single time it runs.
Right now, every agent call starts from scratch. It searches, reasons, searches again, reasons again. Even if your team already made the decision three weeks ago. You're paying frontier model prices to rediscover your own conclusions.
Access is table stakes. Memory is the moat.
Agents getting the right context to do their work will be the dominant IT challenge over the next decade. Every agent strategy is at the mercy of how effectively agents can access the right data and systems to make decisions. Huge opportunity for those that get this right.
@levie access is just the first step. you need to make it work at scale. Current systems re-derive the same conclusions every time an agent runs...burning tokens to rediscover what the team already decided last month. The missing layer isn't access, it's memory: https://t.co/nhoIMwtoii
@kimberlywtan@a16z Coding is so far ahead because the whole relevant context is always in the reach of the agents. General enterprise AI needs the same: org knowledge modelled as an ever-changing system, encodes reasoning of creation and is activated without an LLM. @internode_ai does this.
@pmarca think in these terms: encoded temporal reasoning is the new software, while LLMs are the chips. Everybody is using their LLMs without having proper software and prompts are NOT that 🤦♂️ you do NOT want to make the LLM go through the same reasoning every single time you ask smthg
running these things without a proper memory layer is like rewriting the entire Excel software every time you want to do a SUM 🤣 of course it's going to cost a LOT 🤦♂️🤦♂️
@himanshustwts the problem with these setups is that the knowledge becomes stale super fast. it's the same as when you're using confluence or notion and things change daily. nobody has the time to keep things updated. you must have a system that models knowledge with change at the foundation.