Pinned Tweet
when vercel dropped websockets and tom lienard dropped hints on how it worked,
I was hyped and got to work to build a tunneling service that is just a vercel deployment, so introducing
β² πππ π πππππππππππ -π
- deploy in one command
- supports both HTTP and WS ection and IP allowlist
- stable domains with auto configuration through vercel CLI
- monorepo support
I was hyped and got to work to build a tunneling service that is just a vercel deployment, so introducing
β² πππ π πππππππππππ -π
- deploy in one command
- supports both HTTP and WS ection and IP allowlist
- stable domains with auto configuration through vercel CLI
- monorepo support
I really enjoyed using ππππππππ by @ctatedev β http was a non-starter for local wp development!
11
5
102
15,634
Sree retweeted
Introducing Rivet 3.0 (preview)
Built on S3 with SSD speeds
ποΈ 8ms actor coldstarts
πͺΆ 0.7ms SQLite writes
π€ 78 KB of RAM per actor
π± Embed, self-host, or BYOC
π S3 is the only dependency, no database
Built on S3 with SSD speeds
ποΈ 8ms actor coldstarts
πͺΆ 0.7ms SQLite writes
π€ 78 KB of RAM per actor
π± Embed, self-host, or BYOC
π S3 is the only dependency, no database
34
45
674
44,187
Sree retweeted
Congratulations Bryan Johnson for having the highest uptime in the industry right now.
16
16
424
19,772
Sree retweeted
private keys from 8 intermediate CA certs have been compromised (eMudhra)
Happy "Update everything" day to all my pager-carrying friends
emudhra.we-have-a-problem.com
Happy "Update everything" day to all my pager-carrying friends
emudhra.we-have-a-problem.com
93
999
5,171
4,669,122
mate, @Bungie, weβve been asking this for ages
just 2 of the places that will be showcased in tomorrows video :)
0
0
1
100
my jubilation can only be expressed by this glorious picture at what's being cultivated in rivet labs
2
0
6
1,531
Sree retweeted
Training large models with RL typically runs on a fixed homogeneous cluster of the latest accelerators for numerical stability and faster throughput. This throws away inference optimization work done for other vendors and generations, limits the capacity you can utilize, and is .
So we trained a model on agentic coding tasks using fully asynchronous RL and mixed-precision, mixed-accelerator samplers. With off-policy correction, training stayed stable through convergence, on par with homogeneous samplers, at a lower cost.
So we trained a model on agentic coding tasks using fully asynchronous RL and mixed-precision, mixed-accelerator samplers. With off-policy correction, training stayed stable through convergence, on par with homogeneous samplers, at a lower cost.
32
129
1,665
165,500
Sree retweeted
can you run pip inside agentOS? (tiny wasm sandbox)
with c extensions?
and train a language model?
yep
with c extensions?
and train a language model?
yep
@NathanFlurry The real test for wasm sandboxes is an agent running pip install on something with C extensions, that is usually where you end up back on a microVM. Cover that path and the 100x holds for real agent workloads.
5
1
56
2,326
kingdom: animalia
phylum: containeria
genus: agentOSa
phylum: containeria
genus: agentOSa
what if sandboxes did slightly less but were 100x more efficient?
microVMs get the job done today
but new forms of sandboxes are on the horizon
this is the vision of agentOS (wasm linux sandboxes)
~~~
as agents own more design decisions β they will chose more efficient eady seeing this with so many projects defaulting to Rust today -- trading flexibility for performance
agentOS is like the Rust of sandboxes -- a bit more rigid, but hundreds of times more efficient
agentOS will not replace existing sandboxes
freestyle, e2b, daytona, modal are all building phenominal infra and aren't going anywhere
~~~
agentOS's goal is to be the most:
- efficient
- portable (embed as library or call as api)
- flexible
for whatever agents need to sandbox
today, it:
- runs on webassembly and v8 isolates
- maintains linux sys call compatiblity
- assumes most code is written in in js or python
but there are paths forward that move being this:
- there are faster langues than javascript
- there are safer languages than javascritp
- there are more efficient alternatives to wasm
- there are forms of hardware isolation to try (opt-in)
and agentOS will evolve to meet those goals
~~~
x86 sandboxes isn't going anywhere
but there are new things coming
1
0
4
251
Sree retweeted
Introducing agentOS 0.4 & Hosted VMs
WebAssembly Linux sandboxes, now as a library or a managed HTTP API
β‘ ~5 ms boot & ~22 MB RAM
πΎ Durable, sleeps when idle
π₯ Multiplayer, crons, workflows, & JWT
π Many many bux fixes
WebAssembly Linux sandboxes, now as a library or a managed HTTP API
β‘ ~5 ms boot & ~22 MB RAM
πΎ Durable, sleeps when idle
π₯ Multiplayer, crons, workflows, & JWT
π Many many bux fixes
4
13
180
16,488
Sree retweeted
100% this
most agents just need a workspace, not an expensive sandbox
we're building agentOS for this:
- single digit coldstarts
- posix compliant filesystem
- connect to "workspaces" anywhere (s3, potentially archil)
using webassembly for:
- cheap sandboxing
- MB of RAM
- supports real linux programs (including networking)
it's not built to be a full x86 sandbox replacement
it's built to be the most efficient way to run 90% of agent workloads
and when they need a sandbox, it can escalate compute
most agents just need a workspace, not an expensive sandbox
we're building agentOS for this:
- single digit coldstarts
- posix compliant filesystem
- connect to "workspaces" anywhere (s3, potentially archil)
using webassembly for:
- cheap sandboxing
- MB of RAM
- supports real linux programs (including networking)
it's not built to be a full x86 sandbox replacement
it's built to be the most efficient way to run 90% of agent workloads
and when they need a sandbox, it can escalate compute
seems obvious to everyone around me that sandboxes are on their way out. we hear too many stories about reliability, capacity, broken sandboxes that need manual intervention -- plus they are way too expensive for what they provide.
the future for agents? workspaces.
your agent doesn't need a computer. it needs a place to store it's context, and the *tools* to manipulate that context each turn.
we're in the very early stages of the AI era. the simplest version of this pattern is just handing each agent a time-multiplexed VM, since that's the only thing that we have at hand right now.
over the next 12 months, you're going to see a rapid shift away from this model and towards what i call the "magic compute box".
rather than giving an agent a VM, you give it a set of tools (could be "bash", "python", "js", "read_file", "write_file") -- and you send those commands to a magic service in the cloud that manages to run them for you -- using the cheapest technology available (just-bash, wasm, or even Linux if really required) for each command invcation.
your agent doesn't want a computer that it can set a desktop background for, it wants a place where it can fire off an unlimited number of tool calls and have it update the context correctly.
12
7
124
5,956
Sree retweeted
here's why i'm excited about executor v2:
take the Gmail API, you can start with something as simple generating tools from their api spec for your agent
from there you can send a prompt like "make it so gmail sends emails with a 5 minute cancelation period" and executor v2 will update your gmail app
you can then say "make me a ui that shows me my unread emails and pending sent emails" and it will give you back a url for apps you can bookmark and actually revisit
then after that, you can send a prompt like "read all of my emails and distill it into a writing style skill" and it'll create a skill
then after that you can send a prompt like "when i get a new email, classify it with pangram and jev and if it's spam delete it" where executor will create a webhook handler for you
lastly you can send a prompt like "i want to be able to take notes on the emails i receive" and it'll extend the tools with returning notes, with durable storage
it lets you vibe code personal software on top of all of your connections, it's so good
you can then also share this app for anyone to use
take the Gmail API, you can start with something as simple generating tools from their api spec for your agent
from there you can send a prompt like "make it so gmail sends emails with a 5 minute cancelation period" and executor v2 will update your gmail app
you can then say "make me a ui that shows me my unread emails and pending sent emails" and it will give you back a url for apps you can bookmark and actually revisit
then after that, you can send a prompt like "read all of my emails and distill it into a writing style skill" and it'll create a skill
then after that you can send a prompt like "when i get a new email, classify it with pangram and jev and if it's spam delete it" where executor will create a webhook handler for you
lastly you can send a prompt like "i want to be able to take notes on the emails i receive" and it'll extend the tools with returning notes, with durable storage
it lets you vibe code personal software on top of all of your connections, it's so good
you can then also share this app for anyone to use
23
10
280
15,792