Scaled Veeam 0→$1B. Building everworker — AI does the work, humans decide. Physics major who still codes Haskell. Type theory, FP, piano.
Greenwich, Connecticut
Joined February 2008
- Tweets2.3K
- Following2.7K
- Followers2.7K
- Likes197
Pinned Tweet
I scaled Veeam from $0 to $1B in sales, then started running AI startups.
Here's what that decade taught me that the AI hype is getting wrong:
- Buying is neuroscience. Selling is economics.
- How humans buy hasn't changed — latent pain → pain → vision, filtered through risk and trust. That's constant.
What's collapsing is the cost of selling — AI is driving the price and cycle-time of every step toward zero.
- So the winners won't fire their teams and let agents run wild. That's just faster chaos. They'll keep the operating manual and change the cost of running it:
→ Automate the volume.
→ Augment the judgment.
→ Keep the boardroom human.
AI doesn't replace your people. It makes each one 10x more powerful.
That's what I'm building at everworker together with Ratmir Timashev, Veeam's founder and Ohio Philantropist — and what I'll write about here.
Follow along. 👇
1/ TypeSafe, founded by one of the original ChatGPT developers, just introduced Jev—a “System One Model” that gives up text generation entirely.
Basically: frontier intelligence packaged as a very fast, fuzzy if statement. 🧵
6/ The blueprint is simple:
Unstructured input
↓
Jev classifies and scores
↓
Application code enforces policy
↓
LLM, human or deterministic handler takes over
Use LLMs for language and open-ended reasoning. Use Jev for millions of narrow decisions.
7/ I immediately see applications in actionable-signal filtering, priority matching, urgency scoring and skeptic passes.
We’re planning to bring Jev into OneRingAI as soon as we get early access.
Agentic harnesses will need to be rewritten—but this is for the better :)
typesafe.ai/blog/introducing…
Ok I’m starting Aicoholics Anonymous. For all you poor souls who go „aaa, my token limit is reached, where can I get another dose ASAP?!“
Those who are on unlimited API plans are beyond help anyway.
This is hilarious. "Congratulations on reaching AGI!" -- in the meantime, AGI can't exit #vim for the whole night.
That's the AGI we deserve, apparently.
GPT-6 ASTRA CHANGES THE AGENT LOOP.
Tools no longer have to block reasoning.
OneRingAI v1.1.6 is out with:
→ async tool calls
→ mid-turn steering
→ reasoning updates + compaction
→ misalignment monitoring
Open source: npmjs.com/package/@everworke…
oneringai.io
1/15 AI AGENT POWER = CONTEXT × TOOLS.
We obsess over tools. But most production failures I see start in CONTEXT: the agent gets too much, too little, too late—or loses the result of its own work.
Here is the native agent-context architecture we built in OneRingAI: 🧵
14/15 Layer 6: SECURITY.
Memory scope comes from trusted host identity—not model arguments.
Every new entity/fact needs an owner. User/group/principal permissions are enforced at storage. Read and write are enabled independently.
A shared vector DB is not a permission model.
15/15 The result is not “infinite context.”
It is GOVERNED CONTEXT:
right information
right fidelity
right time
right permissions
measurable cost
Models will commoditize. Context architecture is where an LLM becomes a reliable worker.
What should I unpack next?
oneringai.io
