BADAS 2.0 Launch Brief

01 -- Core Positioning

David vs Goliath.

NVIDIA spent $26 billion building open models. Their Cosmos 2 has 2 billion parameters. BADAS 2.0 defeats it with 300 million parameters, and our smallest model does it with 22 million. That is 100x fewer parameters. Same training data. Fundamentally superior architecture.

There is nothing that understands risk better than BADAS 2.0. Just imagine the amount of effort already put into creating Cosmos versus our investment and our team. This is totally a David versus Goliath story.

Eran Shir, CEO

94%

mAP on Kaggle (BADAS 2.0)

~85%

mAP Cosmos (fine-tuned)

2.5x

Lower error rate

02 -- Model Family

Smaller. Cheaper. Better.

BADAS 2.0 ships as a family of three models. Every size outperforms BADAS 1.0. The smallest runs on a CPU. No GPU required. Literally any device.

BADAS 2.0

300M parameters

State-of-the-art collision prediction. Best time-to-alert. Exceptional on long-tail and out-of-distribution cases. The flagship.

BADAS 2.0 Flash

87M parameters

Quarter the size of the flagship, still outperforms BADAS 1.0. Lower false positive rate than the full model. Ideal edge deployment.

BADAS 2.0 FlashLight

22M parameters

Runs on CPU. 2.5ms per window on Jetson. On par with BADAS 1.0 at less than 10% the size. IoT-grade, ultra-light.

03 -- The Smoking Gun

World Model vs Autoregressive.

Same data. Two architectures. BADAS uses a world model approach. Cosmos uses autoregressive prediction. On standard benchmarks, Cosmos performed adequately after fine-tuning. But on out-of-distribution scenarios, Cosmos completely failed to generalize. BADAS handled them natively.

This is the validation of Yann LeCun's thesis. The world model learns generalization capabilities that autoregressive architectures cannot. And we have the numbers to prove it.

Same data, two different architectures: autoregressive versus world model. The world model learns generalization capabilities. Autoregressive Cosmos does not. Although Cosmos has two billion parameters and we have just 300 million. This is the smoking gun.

Ronnie Goldshmidt, Head of Research

300M

BADAS parameters

Cosmos parameters

100%

Cosmos fail rate on OOD

04 -- Explainability

It shows its work.

BADAS 2.0 generates attention-based heatmaps that reveal exactly what triggered an alert. The model's attention spreads across the entire road, then focuses sharply on the dangerous object the moment risk increases. Like a radar locking onto a target.

This was not trained. It emerged natively from the architecture. And it comes with text descriptions and voice narration of the risk scenario. All model sizes support this, including FlashLight.

The quality of the attention maps of the Flash model is even better than the original BADAS 2.0 model. I believe it is because we trained the small models using VJA method on large scale of our own data.

Ronnie Goldshmidt, Head of Research

Attention maps

Visual explainability

Voice + Text

Narrated risk alerts

All sizes

Flash attention is even sharper

05 -- Narrative Pillars

Five stories to tell.

1. David vs Goliath

NVIDIA spent $26B on open models. Our small team built something better with orders of magnitude fewer resources.

2. World Models Win

Same data, two architectures. BADAS generalizes. Cosmos does not. Validation of the world model thesis at the center of AI discourse.

3. Compounding Data Advantage

BADAS 2.0 was built using BADAS 1.0. No one else has access to our prior models, Atlas, or the dataset. The moat deepens with every iteration.

4. Smaller, Cheaper, Better

A full model family from 22M to 300M parameters. The smallest runs on CPU at 2.5ms. The edge story writes itself.

5. The Pepsi Challenge

Open challenge to the industry. Release benchmark data. Let anyone try to beat it. No one will, and no one will open their closed models to try.

06 -- Deliverables

What needs to ship.

Click any deliverable to expand sub-tasks and dependencies.

Headline Claim

Marketing

One-liner positioning statement. "Best collision prediction model on the planet" or equivalent. Must be defensible, quotable, and legally clear.

Draft 5 headline candidates with supporting evidence for each

Legal review of top 2 claims for defensibility

C-staff sign-off on final wording

blocks Landing Page, Social Content, Investor Package

Landing Page / Microsite

Marketing

The public face of the launch. Hero claim, interactive demo embed, research numbers, benchmark dashboard.

Wireframe and page structure (Marketing)

Copy deck aligned to headline claim

needs Headline Claim

Embed API Playground widget

needs API Playground

Integrate benchmark charts

needs Benchmark Charts

Embed comparison + explainability videos

needs Video Assets

QA, mobile optimization, performance audit

Deploy to production URL

API Playground

NAP Team

Deploy all 3 models on NAP infrastructure. Public API endpoint where visitors upload videos and get predictions. End the "send us your video" loop.

Deploy BADAS 2.0 (300M) to NAP inference endpoint

Deploy Flash (87M) and FlashLight (22M) endpoints

Build video upload + results UI widget (embeddable)

Rate limiting + abuse protection for public access

API documentation for developer access

Load testing at expected launch traffic

Benchmark Dataset Release

Product

Curated subset of long-tail edge cases: animals, cyclists, snow, infrastructure. Enough to prove the claim without giving away competitive advantage.

Research team selects candidate edge cases from long-tail test set

Product + C-staff determine release scope (volume, categories)

Legal review of data licensing terms

Package dataset with documentation and evaluation scripts

Host on public repository (HuggingFace or similar)

blocks Open Challenge credibility

Comparison Videos

Marketing

BADAS vs Cosmos side-by-side on identical scenarios. Two tiers: mobile-optimized for social and detailed for solution architects.

Source externally-sharable video clips (Marketing to provide)

Run BADAS 2.0 + fine-tuned Cosmos on approved clips (Research)

Produce Tier 1: social-optimized (minimal annotation, readable at phone size)

Produce Tier 2: detailed technical version (full annotation for large screen)

Captions, branding, format for each platform (LinkedIn, X, YouTube)

needs Externally sharable footage clearance

Explainability Demo Videos

Marketing

Attention maps, text descriptions, voice narration. Demonstrate the "radar lock" behavior. Same two-tier approach as comparison videos.

Select best attention map examples (animal, pedestrian, vehicle scenarios)

Generate attention visualizations via API (Research)

Record voice narration overlay

Produce Tier 1 (social) and Tier 2 (technical) versions

needs API Playground deployed for generation

Research Presentation

C-Staff

Board-ready deck with all benchmarks. Designed for board forum distribution and Yann LeCun network amplification.

Compile all finalized benchmark numbers into presentation format

World model narrative framing (LeCun thesis context)

Visual design aligned to brand (Marketing support)

Board forum distribution strategy and timing

needs Benchmark Charts finalized

Investor Narrative Package

C-Staff

Timed before investor conversations. Tease capability in meetings this week, full launch within two weeks.

Teaser talking points for immediate investor conversations

Full narrative deck integrating research presentation

One-pager leave-behind with key claims and links

needs Headline Claim, Benchmark Charts

Performance Benchmark Charts

Product

mAP on Kaggle, long-tail AP, false positive rates, model size vs performance. Add Cosmos to all existing charts.

Finalize remaining benchmark runs (Research)

Calculate Cosmos metrics on same test sets for apples-to-apples comparison

Build OOD generalization benchmark (new test set)

Design publication-quality chart visuals (Marketing)

Create model size vs performance scatter plot

blocks Landing Page, Research Deck, Investor Package

Open Challenge

Marketing

Public invitation: "You have a better model? Prove it." Release benchmark data. Reinforce confidence and frame real-world data as the winning approach.

Define challenge rules, evaluation criteria, and submission process

Build challenge page (section on Landing Page or standalone)

Link to public benchmark dataset

PR/social campaign around the challenge launch

needs Benchmark Dataset, Landing Page

07 -- Timeline

Two weeks. No more.

Investor conversations are active now. The launch must land by March 26 or tease credibly enough to carry momentum into those meetings.

Today

Deliverable

Benchmark Charts

Headline Claim

API Playground

Benchmark Dataset

Comparison Videos

Explainability Videos

Research Deck

Investor Package

Landing Page

Open Challenge

LAUNCH

NAP / Marketing

Product / C-Staff

Dependent / phased

Milestone

Open Decisions

Dataset Scope

How many edge cases to release publicly. Enough to prove the claim, not enough to give away the competitive advantage.

Cosmos Framing

Team comfortable being direct. "Not that Cosmos is bad, just not close to BADAS." Need precise tone for public comms.

Yann LeCun Timing

Board presentation timing and network amplification depend on readiness of research deck. C-staff to coordinate.