STEAV

STEAV is the sovereign intelligence platform. Companies build AI models on their own data, automate their workflows with agents, and keep full control.

[STEAV @getSTEAV]

Lloyds’ agentic AI fraud platform: 10,600+ live cases in 90 days, 90%+ decision accuracy, overrides down 30%. Specialists still make the final call.

Fraud teams: which is harder to get right, accuracy or the handoff to people?

[STEAV @getSTEAV]

Almost every infrastructure boom has ended in an overbuild. Canals, railways, electric utilities, fiber, solar. Investors lost, but the people who came after got cheap infrastructure.

Is AI compute the next one? If so, who carries the loss, and who benefits?

#AI #Infrastructure #DataCenters #Investing

[STEAV @getSTEAV]

@jackson_felty We are!

Check us out: https://t.co/ckAQNYpWAj

[STEAV @getSTEAV]

We were also able the same with decision making models, using Laya at @getSTEAV.

We just put our System One fraud models on Hugging Face, open for anyone to download and check.

There are four of them, one for each benchmark set whose data we could release a model on: fake job postings, IP addresses on a threat blocklist, simulated card transactions, and two days of real 2013 card payments. Weights and code are Apache-2.0. The other five sets in the Fraud Dataset Benchmark stay internal, because their data carries non-commercial or research-only terms.

Check them out here: https://t.co/SmpGuEoESH

More technical details here: https://t.co/ehkPzRHE30

[Brian Armstrong @brian_armstrong]

We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).

I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.

We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:

"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"

I imagine we'll have some backend services that could be replaced entirely by this new architecture.

[STEAV @getSTEAV]

Every repo rebuilds its scores bit for bit and passed three independent reviews. These are benchmark models, not production fraud systems.

Models: https://t.co/TNSJoKwZHP
Write-up: https://t.co/2EmBsMaSYI

#AI #OpenSource

[STEAV @getSTEAV]

The game specialists, fine-tuned on Laya: agreement with the planner's move went from 60.9% to 100% on Dino, 60.9% to 99.8% on Snake, and 22.9% to 96.1% on Tetris's top five placements.

[STEAV @getSTEAV]

CID runs every recipe through the same seven stages and records every outcome, so those measurements add up.

Full post: https://t.co/0Z2oUn5Bfv

[STEAV @getSTEAV]

Reasoning is a training recipe, not a model type. DeepSeek-R1 made it public in January 2025: RL on rewards a rule can check, like an exact answer or a passing unit test. No learned reward model.

[STEAV @getSTEAV]

Before you adopt a recipe, measure four things on your own workload: pass@1 vs pass@k against a fixed base, p50 and p95 trace length, serving optimizations on and off, and trace faithfulness.

[STEAV @getSTEAV]

Earlier this week we said CID had not demonstrated Level 3. Then its gate promoted a learned advisor, then promoted a second that beat the first, 0.886 to 0.818. It said no three times.

The strictest test is still running.

https://t.co/dzk3XAYN1e

#AI #MLOps

[STEAV @getSTEAV]

That's the rung we built CID's Model Studio for. We haven't demonstrated Level 3, and we'd rather say that plainly than show you a demo.

#AI #MLOps

[STEAV @getSTEAV]

People call a lot of things "self-improving." A scheduled retraining job. A hyperparameter search. A research agent rewriting its own code. They are not the same thing, so we laid them out as a ladder.

[STEAV @getSTEAV]

Most production ML sits at L1 or L2. Research systems touch L4 inside sandboxes. L3 is the gap in the middle, and the first place "self-improving" becomes something you can actually test.