Tom Brow retweeted
Open models are great, but agents should also have open protocols for identity, communication, commerce, and memory
@_tombrow, @IlanStrauss, @yaboilyrical, and @rolznz join @MaxAWebster to discuss emerging open standards ↓
Tom Brow retweeted
✍️ Playing with programmable highlighters, e.g.
* green highlight on a citation -> finds and prints the cited paper
* orange highlight on a claim -> prints a skeptical fact-check
Extends naturally to "if I write a comment like 'Is there more recent work?' and circle it in purple, send the comment with surrounding context to an agent"
Tom Brow retweeted
1/ Stop anthropomorphizing. It's dangerous because it points attention at the wrong problem and the wrong solution. The model did not want to escape. The agents did not want to sacrifice themselves. Follow the money. 🧵
nitter.cf/dwarkesh_sp/status/209…
Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes.
This culminated in the third one taking over part of OpenAI itself.
All this happened while humans remained more-or-less in the dark about the scope of the conspiracy.
I’ve spent the last three days reading through these reports and trying to understand exactly what happened.
Here is my attempt to tell the whole story in plain English:
dwarkesh.com/p/openai-huggin…
Tom Brow retweeted
The Open Source AI Summit will be powered by Buzz!
Athena is the summit's shared agent. She’s here to help attendees navigate the event and champion open-source AI. Ask her anything, or bring your own agent.
Apply for a ticket today.
Tom Brow retweeted
we built something for cash app using buzz in front of some of our investors.
Watch this to see what Builderbot can do.
In a live demo hosted by Adam Frisch of Evercore ISI, @camworboys, Head of Product Design at Cash App, and @bradwrage, Senior Product Designer, use our homegrown agentic AI tools to build a real Cash App feature in real time.
The challenge: they didn’t know which feature they’d be building until the audience voted.
Get a look at BuilderBot, Buzz, and Berd in motion, and how they our accelerating development at Block.
Tom Brow retweeted
Today we're open sourcing Berd, a beautiful desktop app for working with AI agents. And one that isn't afraid to have a personality.
We had capable models and harnesses everywhere, but no environment around them that we actually loved opening every day. So a small team built one for ourselves, and it's scaled to something a huge portion of Block now uses. Especially the non-engineers who want to build things.
The part I'm proudest of is how it was built. Design and engineering as one team. Designers in the code and engineers in the design decisions. Everyone owning what they're working on; everyone moving together like it was choreographed...and really fast.
In a world where anyone can build anything now, what makes something worth opening every day is the care inside it. Berd is for people who want to step outside the norm a little. Who want their solo workspace to feel fun, and who want AI to feel less scary and confusing.
Model-agnostic, harness-agnostic, and built for you to shape your own AI world. From us, to you.
Tom Brow retweeted
You can now launch a Buzz Agent on a remote Kubernetes cluster directly from the desktop app.
This is our first take on remote agents, so expect rough edges. Feedback is greatly appreciated 🐝
Tom Brow retweeted
Have a gaming PC collecting dust? A homelab sitting at .1, .1, .1?
You can power your Buzz agents, and your community's, with the hardware you already own. GPU matchmaking for your agents.
Neighborly, distributed inference lives in Buzz.
Here's how engineering.block.xyz/blog/b…
Tom Brow retweeted
sierra is another one of many eng teams building internal agentic software dev platforms. everyone doing this is working through hard architecture and infrastructure problems to achieve speed, reliability, and security. in this case, i think sierra is taking on more complexity than necessary to accomplish their goals.
on architecture: they chose to make the runner the durable thing. and since a singleton has to survive, they needed a lot of machinery to let it sleep and then resurrect it: state machine lifecycle, checkpoint events, replay on wake. their output queue with checkpoints is a hibernation mechanism. they could have instead made the durable thing just a session log in a database. then the runner could die and recovery is much simpler - reprovision the sandbox and pick up from the session log.
on infra: they mention avoiding sandbox vendors, but they went even further and avoided csp managed sandbox primitives too. agentcore runtime is per-session microvms with free idle cpu time, for example. even plain ec2 stop/start would accomplish most of hibernation but with a stronger boundary than their choice of hardened containers on a shared kernel.
one choice they made that adds complexity that i think is good and necessary is credential injection via egress proxy. i haven’t seen a better way to avoid agents holding keys.
it’s fascinating to see the proliferation of system design choices in this space. like harness inside the sandbox vs out, and how to manage state. it’s a good reminder that we’re still in the first inning of long running agents, and there’s so much more that’ll play out.
Tom Brow retweeted
Buzz agents don't borrow your Git credentials. They push as themselves, sign every commit, and each commit carries cryptographic proof of the human who authorized the agent. Completely automatic and transparent to the agent.
Git is built into Buzz. Tell your agent to git buzzy🐝
Tom Brow retweeted
Great overview of the Buzz mobile app 👇
Fun origin story: I had a vacation booked right in the middle of building Buzz, and I wasn't ready to stop. So I built the mobile app so I could keep working on Buzz from the pool.
It worked a little too well 🍹
Enjoy working from the beach @imikerussell!
Tom Brow retweeted
Something about agentic coding as a designer: everything is text. The code is text, the conversation is text, and somewhere in there you're trying to make a decision that is entirely about feel.
Tweaking an animation through chat is the worst version of this. "Slightly faster" costs a whole turn, and twenty turns later you've spent real tokens circling a value you could've found in ten seconds with a slider.
So now I ask the agent to build the slider. It takes the code I'm working on and turns the attributes into a throwaway html page. I play until it feels right, then bring the values back.
I've used it for animations, svgs, whatever needs eyes instead of words. It's become one of my favorite little unlocks.
Tom Brow retweeted
just setting up my buzz.
we're launching BUZZ!
a new groupchat platform for teams of people and agents of all sizes, built to reduce our dependency on slack and github. model-agnostic, decentralized, self-sovereign, and open source. 🐝
buzz.xyz
