building the future of ai @typesafeai
San Francisco
Joined November 2015
- Tweets533
- Following383
- Followers2K
- Likes17.4K
Eugene Shvarts retweeted
Agents have a trust problem.
Today’s models can do most work. What they need is the right boundaries. In traditional software, those were set at build time. With agents, they have to be set and enforced at runtime.
Today we’re introducing Composio Firewall, the runtime boundary for agents.
Millions of users and developers already use Composio as the runtime for their agents - giving agents tools, auth, and sandboxed execution. Now with our agent firewall every action an agent takes is evaluated using Jev and custom SLMs tuned by Composio based on real-world usage data.
The firewall strengthens that existing boundary in both directions, checking the actions agents take and the information coming back to them through Composio.
The advantage of the runtime boundary is that Composio Firewall verifies that the agent respects user intent and the boundaries set by the developer and user. Additionally, this provides the safety layer required to protect the user against malicious prompt injection as well as other attacks.
Composio Firewall is live now for all Composio developers.
Check it out at composio.dev/firewall
The things we found in common with everyone on the data team, after the fact, were absurd. Almost a little eerie.
Maybe it's you too?
typesafe.ai/careers
Eugene Shvarts retweeted
Last week it was "set minimum package age to 7 days for everything because the supply-chain apocalypse is here"
This week it's "update all your shit to what was released 5 minutes ago because the 0-day apocalypse is here"
I wonder what next week will be
Eugene Shvarts retweeted
We put Jev inside the Lightpanda binary and built judge() on top of it. @typesafeai
No more sending a whole page to an LLM just to find out it's a cookie banner.
> Blocked pages, CAPTCHAs, cookie walls, logins, paywalls, error pages, empty searches and pages still loading can all return an HTTP 200
> judge() asks a set of predefined questions and tells you if the page has real content, and what kind: article, listing, product or docs
> One call, ~0.2s
In the video:
> 30 sites, ~14s total
> Lightpanda backs off when rate-limited, accepts cookie walls, and keeps or skips each page
Every day, people and companies are discovering new ways that their everyday work can be automated with deterministic code that calls for exactly the right tier of intelligence, right where it's needed.
Filtering is the expensive half of a news digest. Every article needs a judgment call, and a judgment call has always meant an LLM call.
Not anymore. $0.00004 per article.
We teamed up with @n8n_io and built a weekday AI brief using @typesafeai's Jev. Now live on n8n blog as part of their Verified Node Spotlight. 🧵
Eugene Shvarts retweeted
🧵New post in my AI series: Can a million parallel agents create superintelligence? TLDR: More agents buy speed, but intelligence gains diminish fast, and much faster than RSI models. Thanks @noahpinion for the encouragement! 1/25
Eugene Shvarts retweeted
I just built Computer Use that’s cheaper, faster, and acts like a human 🖥️
Powered by @typesafeai’s JEV 1.14, using the new Image/Box feature.
It looks at the screen, decides what to do, then clicks and types like a real person.
⚡ Way cheaper than using an LLM
🧑 Acts like a human: sees the screen, picks an action, clicks or types it
📋 Full report of every step: the screenshot, the action it picked and why
Demo: I told it to open Wikipedia, search for an article about aliens, copy the link, and post a tweet with the link and some text.
- Cost: $0.0019
- Tokens: ~50,000
OPEN SOURCING SOON on my github 👀 (as always)
Eugene Shvarts retweeted
fail more!
fail faster!
fail relentlessly!
Eugene Shvarts retweeted
Common misconception!
Modal's moat is actually a ditch around the perimeter of the office we have filled with sharp sticks.
When someone else tries to build serverless infra, the ditch slows them down as we pour boiling pitch on their heads and set fire to their siege engines.
Eugene Shvarts retweeted
SELECT prompt_jev(image, ‘Is this a hot dog?’) FROM food
Jev image support is now in MotherDuck SQL! Classify, score, or get bounding boxes on images at analytics scale.
Our actually-rigorous benchmark on this is below. Comparable quality to Sonnet at 10x the speed and 1/25th the cost. And since it’s Jev, you get a real confidence metric. What a dream.
Eugene Shvarts retweeted
Hello Software Engineers, I lead TypeSafe AI Model Capabilities team. We are hiring. Here are a few qualities we are looking for:
• strong plus for people who resonate with the title "Experience Engineers": mobile engineers, full stack engineers, or even (virtual) reality engineers.
• strong coder (you need to be able to manually write code)
• product oriented thinker
I promise are fun! I trust that you will figure out how to apply.
Eugene Shvarts retweeted
Nathan has been putting out a whole bunch of little explainers and idea-seeds for using Jev - if you're not following him, fix that right meow
Eugene Shvarts retweeted
We’re giving away $1 million to agents.
Introducing Composio Instant: the easiest way for agents to pay for their own tools.
Claim your agent's share: composio.dev/instant
