Marco De Rossi (@marco_derossi) — posts, photos & videos

2 1 0 8 501 0

@marco_derossi

22h

My cofounder spent his nights this week showing that a decision model CAN play World of Warcraft (using image analysis+choice) and hit level 5. It worked! 🤣

Big Iron Chris@bigironchris · 22h

This is the first decision model to level up a World of Warcraft character.

It died at level 2 😆

Then it learned to heal itself mid-fight. 128+ kills later the dwarf hit level 5, at 2¢ a kill.

The model is Sage, the decision model we build at @levantolabs.

At each step, our harness, SageCraft, hands it a screenshot and a short list of moves. Sage picks one. The harness checks the pick and turns it into keyboard and mouse input.

LLM agents have played games before. They write their own actions. Sage only chooses.

Picking turned out to be the easy part. Sage keeps no memory between calls, so the harness has to carry the goal, what just happened, and what to try when it's stuck.

> 32 sessions, 5h 29m of play, 4,573 decisions
> 103 troggs, 19 wolves, 6 boars
> stuck in one tree long enough that I stopped the session by hand

It wasn't hands-off. A coding agent fixed harness bugs between sessions and helped revive the dwarf while I slept. The last 18 minutes, from late level 4 to 5, ran with no manual input, code changes or restarts.

The video is me narrating it unscripted, watching most of it back for the first time.

This is my first harness build, and as far as I can tell, the first time a decision model is leveling up a live WoW character.

It's open source. The harness and notes on the build are in the first reply below 👇

What do you think could be done better?

8:31

0 0 0 14 569 0

@marco_derossi

Oct 9

Glamsterdam. Coming soon on @ethereum 💙

2 0 0 10 414 1

@marco_derossi reposted

Levanto Labs

@levantolabs

Oct 7

Sage is the decision model who can answer ''what-if''

Autoreasoning toggles on when Sage is faced with multi-step logic that can't be handled by other decision models.

Unbefallen explores alternative timelines, see more 👇

Unbefallen@unbefallen · Oct 7

Our first uchronia campaign: Waterloo, 1815.

7 historical figures. 1 shared world. Every choice recorded. Play at : unbefallen.com/app/play

Talk to characters powered by Sage from @levantolabs. Collect, trade, build coalitions.

$ELSE now has its first in-game use. More to come.

0 2 1 6 452 1

@marco_derossi

Oct 7

Decision models in financial pipelines

Levanto Labs@levantolabs · Oct 7

PerpFlow cut an agentic research workflow from 14 days to 12 hours with Sage, a 28× speedup 🔥

@PerpFlow analyzes markets and sends selected trading signals to subscribers.

The team used Sage to classify 5,972 anonymized market cards against fixed research criteria. The adapter saved the results to files and returned summaries to Claude, keeping thousands of classifications outside its context. Python then evaluated the labels across 85 pre-registered tests.

PerpFlow reached its research conclusion sooner while retaining the requirements for changing its production signals.

0:31

0 0 0 4 410 0

@marco_derossi

Oct 7

.@OpenAI, welcome to the board

benchmarkheaven.com/jev-models/api

#DecisionModels @levantolabs

1 0 1 13 1.4K 0

@marco_derossi

Use Sage to choose your next @Airbnb stay 🤭

Celeste Ang@celesteanglm · Oct 6

Benchmarks don't really matter, but real use cases do! Been playing around with decision models recently to make my coding + automation workflows faster and cheaper. I was recently planning a Gangnam stay, so I used Sage to check Airbnb listings for a comfortable and affordable listing which are conducive for remote work.

0:41

1 1 0 9 470 3

@marco_derossi reposted

Julien B.

@bneiluj

China is training kindergarteners' bodies
and robots' hands. at the same time.

guess who builds the physical world 👀

Massimo@Rainmaker1973 · Oct 4

Physical education in a kindergarten in China

0:17

3 1 0 12 5K 6

@marco_derossi

.@AntSeed is a P2P network for AI inference, with distributed reputation, TEE inference support, and great prices. Check it out!

Amos Meiri@AmosMeiri · Oct 6

What BitTorrent did for files, we want to do for AI inference.

Antseed lets anyone buy and sell access to AI models, on their own terms.

The Antseed Foundation has raised $2.4M in a token round led by Spark Capital to help make that happen.

1:53

1 0 0 6 690 1

@marco_derossi

Sage by @LevantoLabs is the first decision model, launched in July, and currently the best on the market.

It wins across all the evals proposed by @perplexity_ai's CEO (nice job @AravSrinivas!), and it's first on JevBench, beating Jev and Google's models on every dimension.

BUT frankly, benchmarks are not what excites us.
I'm BORED by the dozens of new decision models launching every day with a tiny +0.02 on some eval. They all look the same.

The magic is somewhere else.
There are thousands of new things that could be invented in the space of decision models and models for machines. The invention space is huge!

After launching Sage, we also pioneered:

- The first guardrails and model routers built on decision models

- Automated context engineering with "grounding" (automatic web search when confidence is low). Decision models are only as smart as their context, especially because they can't make tool calls.

- Reasoning. In many use cases you don't know whether your question is System 1 or System 2. With reasoning, the model is as fast as a System 1 model when the question is System 1, and takes more time when it's needed. This enables hard, multi-step decisions. Our clients just want the correct answer, and in some use cases they're happy to wait 0.5s more. This will soon become the market standard (GLiDE/@george_onx and @TypeLLM already followed).

And *so* much is yet to come. Models for machines: it's just the beginning!

We're cooking. Follow @levantolabs for new product primitives that will surprise you.

Open innovation is HEALTHY.
Competition is HEALTHY.
Benchmark-race is BORING.
Product inventions are FUN.

Go go go!!!

14 22 6 59 5.7K 11

Illia (root.near) (🇺🇦, ⋈)

@marco_derossi

Oct 5

Anthropic valuation: $2 Trillion
Meta valuation: $1.85 Trillion

Anthropic 2025 revenue: $4.6 Billion
Meta 2025 revenue: $200 Billion

11 0 1 32 2.7K 4

@marco_derossi reposted

@ilblackdragon

NEAR AI has claimed first place on the Lean Eval v1 leaderboard.

Lean Eval v1 is an active benchmark maintained by Lean FRO that contains extremely hard formalization problems. We reached first place despite joining the competition just four weeks ago, a full month later than other top participants.

NEAR is very heavily investing in formal verification efforts. Our ultimate goal is to formally prove the protocol and all the core contracts, and to provide a practical tool for others to formally verify their contracts in a fast and affordable way.

The system that enabled us to claim first place is almost entirely powered by DeepSeek V4.1 Flash, and thus is (a) extremely cheap, and (b) has no dependencies on third parties because the model is open weight.

When it comes to formally verifying software, the approach boils down to three pillars:
1. Representing the program in a way that can be reasoned about;
2. Expressing the properties that we desire to prove;
3. Coming up with a proof and formalizing it in Lean.

Until recently, step (3) was prohibitively expensive for practical use at scale. With the system that we used for Lean Eval, we now have it solved. Stay tuned for updates on how we apply it in practice to formally verify NEAR contracts.

33 67 12 442 69.3K 27

@marco_derossi

Chatbots, old way:
Every message goes to the LLM, and you wait.

New way:
A decision model reads the intent in milliseconds, your code handles it, and the LLM is only the fallback.

More of our clients are doing this. The K12 Edu app Leap is a nice example.

Levanto Labs@levantolabs · Oct 2

People don’t like talking to chatbots that take more than 3 seconds to respond. Can decision models help solve this?

Let's look at a tool-calling case study.

A student on Leap types "make me flashcards."

The usual path: send the request to an LLM, wait 3–4 seconds, read a paragraph about flashcards, then go find the button.

Leap's path: Sage, our decision model, recognizes the request in ~100 ms, and the flashcards open.

Every message gets three quick questions first:

• Does the student want an activity?
• Which one?
• More than one?

"Make me flashcards" → the tool opens.
"How do flashcards work?" → that's a real question, so the LLM answers.

Why it matters for the business:
→ Faster: action chips appear in under 300 ms in production.
→ Cheaper: no LLM call to write text nobody reads.
→ Safer: confidence thresholds decide when to act, suggest, or keep chatting, and the team tunes them without an App Store release.

Tested on 327 held-out requests, each run 3 times:
• 89.9% accuracy
• 0 unwanted tool launches
• 6 languages, including mixed-language messages

The pattern works for any agent product: list your actions, test tool selection on real examples, and map confidence to behavior.

Decide first. Generate second.

0:26

1 0 0 10 699 0

@marco_derossi

What if Napoleon had won at Waterloo? @unbefallen has the answer.

Excited to see so many new teams building on our Sage!

My fav part: "Sage can say "I don't know" when unsure, it returns no answer instead of guessing. For us, that becomes character. An in-game character that hesitates, doubts or waits feels far more alive than one that always has an answer".

Doubt is not a bug; it's a feature! 😅

Unbefallen@unbefallen · Oct 2

We ran Jev and Sage decision models

What we learned from running these two decision models within the Unbefallen engine.

1 6 2 23 3.3K 1

@marco_derossi

Levanto Sage mates Jev

♟️🔥

quirq@quirq_ai · Oct 1

We pitted two decision models against each other in a game of chess: @typesafeai 's Jev vs @levantolabs's Sage.

Sage was way ahead (+23) . Jev refused every offer to resign. Then Sage played its highest-confidence move, unwittingly pushing Jev's king with nowhere to go. Stalemate. Draw.

0:45

2 2 0 7 1.2K 2

@marco_derossi

Sep 30

What if decision models could play 🎮 video games?

Levanto Sage starring as... the Goblin!

Levanto Labs@levantolabs · Sep 30

We just scrambled 🍳 this goblin's brain.

What if decision models controlled the boss fight?

@GROWBLOCKSXYZ shows how game developers can write a character description and a list of actions, and call a decision model to fully replace combat AI.

Today we talk Goblins:

For decades, NPC behavior was written with complex deterministic branching.

A goblin's combat style and reactive choices were finely tuned based on if-this-then-that statements with some random number generation.

Using Sage, you can replace that development step with a decision model, a character description, and a list of actions.

Step 1: describe the character in natural language.

Step 2: list the actions the character can take

Step 3: add game context, like whether or not other characters nearby are hostile

Step 4: call a Choice decision to Sage

Step 5: run your evals to see if the monster behaves well

This is unlock of decision models. They aren't just faster ways to classify data.

Sage enables new ways of programming systems, enabling developers to create reliable execution with natural language statements written in the codebase.

Link to the case study below:

0:46

1 1 0 7 776 1

@marco_derossi

Sep 29

Congrats @sama and @romainhuet on launching your decision model. Same features, same name.

When we launched in July, people gave us weird looks.

Now Sam is on stage talking about it ♥️

So fast!!!! 🏎️

Tibo@thsottiaux · Sep 29

Decisions API, for lightning fast constrained decision making powered by Luna. Supports visual inputs, and tuned to be able to make decisions in less than a few hundreds of milliseconds end to end.

6 5 0 29 3.3K 3

@marco_derossi

Sep 29

Pretty relaxing soundtrack. I'll get a chamomile 🫖 tea while cutting my LLM spending.

Levanto Labs@levantolabs · Sep 29

Curious how companies are using Sage? We're publishing a new case study every day.

Today we’re featuring @capacitr_xyz, a mobile app that turns news and social posts into trading signals, with a chat agent for execution.

Capacitr uses Sage across six decision gates:

1. Content filtering
2. Web research
3. Market ranking
4. Trade guardrails
5. Chat scope
6. Signal freshness.

Sage evaluates whether incoming content merits processing and whether it needs fresh research.

That prevents low-value items from consuming expensive model and search calls.

At full enforcement, Capacitr projects:
• 44% lower LLM token spend, avoiding 1.6B+ tokens per month.
• 50% lower web research costs, avoiding approximately 19,500 metered search calls per month.
• $500+ in monthly compute savings.

The case study covers the integration, how the team sets enforcement thresholds, and how caching keeps freshness checks outside the feed’s read path.

Links below.

0:36

1 0 0 8 718 0

@marco_derossi

Sep 29

Bet: After Levanto Sage and Jev, @OpenAI is also launching a decision model today at DevDay 👀

GIF

2 0 2 10 982 0

@marco_derossi

Sep 26

Erik is building something pretty impressive

erik.eth 🛡@programmer · Sep 26

I'm still in stealth, but I'm hiring!

Location: SF preferred, but remote in North American time zones is possible for exceptional candidates.

Open roles
- Founding Engineer: build stuff
- Founding Product Marketing Manager: grow stuff

We offer generous salary, equity, health care, and the opportunity to work with smart and kind people on meaningful problems to help advance the state of agents.

How to apply:
Email jobs @ 5d.tech with:

- LinkedIn / X / GitHub / Resume / Personal site / etc
- A description of the thing you've built you're most proud of
- Your current favorite product & why you like it
- Your all time favorite website
- A good meme