$RKLB

Still a completely underrated part of the business: Spacecraft

This video highlights their established 4 spacecraft called Photon, Pioneer, Explorer and Lighting but what it doesn't show is their newest addition: Flatellite

"A scalable solution for both national security operations and commercial customers to provide next-generation global availability for satellite communications in contested or remote environments as well as remote sensing applications. Flatellite can be launched on Neutron and other medium and heavy launch vehicles."

Of course, "the rocket always steals the show," as SPB often says. But what's on the rocket is just as interesting and arguably even more important to the bottom line.

SpaceX AI Satellites are HUGE.

They have a wingspan of 70 meters (229 ft) when deployed.

To put that in perspective, a Boeing 747-8 has a wingspan of 68.4 meters (224.5 ft).

Bastrop, Texas will soon become home to one of the largest solar cell and wafer production facilities in North America with SpaceX's new investments in the area, creating thousands of jobs.

The solar facility is already under construction, and the AI sat facility is expected to start initial production by the end of 2027.

$BKSY: BlackSky was awarded a contract modification to its existing National Reconnaissance Office (NRO) contract to accelerate the development of the company’s AROS broad area collection satellites as a critical commercial alternative to current suppliers for foundation imagery.

The effort funds a direct path toward a flight ready multi-spectral, large-area mapping spacecraft and foundation data collection system in 2028.

These new satellites will be designed to support dynamic country-scale digital mapping, navigation, maritime situational awareness and 3D digital twin applications.

The AROS system will operate as an integrated extension of the company’s existing fleet, incorporating BlackSky’s space, software and platform stack and unlocking an entirely new class of scalable, AI capabilities.

Wow, here is our first look in-depth look SpaceX AI Satellites:

• 70 m wingspan
• 20 m height when deployed
• 150 kW peak / 120 kW average compute payload
• 70 kW/ton
• Compute provider interchangeable

Thermal:
• 110 m² deployable liquid radiator
• Redundant pumping loops
• Integrated micrometeor shielding

Solar:
• 150 kW Solar Array
• 250W/m²
• SpaceX manufactured solar technology from Bastrop, TX

“Starship V3 is more than double the thrust of the Saturn V moon rocket

By V4, we'll be pretty much three times the thrust of the Saturn V moon rocket

We expect Starship to be flying more than once per hour down the road”

— Elon Musk

Infinity Jukebox 🥁

$VELO: Velo3D and Aurelia Technologies Announce Strategic Partnership to Advance Additive Manufacturing for Next‑Generation Gas Turbine Systems

The partnership supports Aurelia's broader continuous‑improvement strategy, one centered on design consolidation, faster iteration, supply chain resilience, and long‑term cost reduction.

The collaboration is important  for Velo3D as it underscores the expanding role of additive manufacturing in advanced energy and propulsion systems.

By partnering with an innovative turbine developer like Aurelia Technologies, Velo3D further demonstrates how its technology can accelerate product development, unlock greater design freedom, overcome traditional manufacturing constraints and strengthen resilient domestic supply chains.

Advanced Bias Correction for Polymarket Weather Trading Bots: ConvLSTM + Self-Attention + Dynamic Normalization

If you're building a serious trading bot for temperature markets, at some point simple bias correction stops being enough

I found one of the most interesting papers from 2025 that breaks down how to build a truly powerful AI bias correction system for ECMWF

Paper: "How to systematically develop an effective AI-based bias correction model" (ReSA-ConvLSTM, Zhou et al., arXiv, April 2025)

Link: arxiv.org/pdf/2504.15322

> What the Authors Built

They created a full framework called ReSA-ConvLSTM, which delivers up to 20% RMSE reduction on the 1–7 day horizon compared to raw ECMWF output.

> Architecture Breakdown

1. ConvLSTM as the main backbone

Regular convolutional networks (like U-Net) are good with images but weak at understanding temporal dynamics

ConvLSTM is a hybrid of Conv + LSTM. It analyzes both spatial features (mountains, seas, cities) and temporal evolution (how temperature changed over the past hours/days)

This makes it much stronger for meteorological data

2. Dynamic grid-wise normalization instead of standard Z-score

Standard normalization uses one global mean for the whole world - which works poorly for temperature

Here, each grid point has its own climatological mean and variance

This removes the huge difference between tropical and polar regions and significantly improves training stability

3. Strict temporal causality

The model only looks at past states and doesn’t “peek” into future forecast days. Critical for real-time trading systems.

4. Residual connections + Self-Attention

Residual connections help the network go deeper without losing information

Self-Attention allows the model to automatically focus on important areas (coastal zones, mountains, etc.)

> Results

• Raw ECMWF: RMSE 1.2°C on day 1 → 2.1°C on day 7
• ReSA-ConvLSTM: below 1.0°C on days 1–3, below 1.8°C on days 4–7

The model also transfers well to wind and pressure variables

> How to Implement It Yourself

There’s no official ReSA-ConvLSTM repo yet, but you can build a very close version using:

ndrplz/ConvLSTM_pytorch - one of the best and cleanest ConvLSTM implementations
Repo: github.com/ndrplz/ConvLSTM_pytorch

tsugumi-sys/SA-ConvLSTM-Pytorch - ConvLSTM already with Self-Attention
Repo: github.com/tsugumi-sys/SA-ConvLSTM-Pytorch

NOAA-EMC/ML4BC - real bias correction project
Repo: github.com/NOAA-EMC/ML4BC

Take ConvLSTM + Self-Attention as the base → add Residual blocks and dynamic grid-wise normalization following the paper → train on ERA5 + ECMWF data

Hybrid architectures like ConvLSTM + Attention + Residual are currently among the strongest directions in weather bias correction

Especially if you want to consistently win on tight 1°C bins on Polymarket

SpaceX $SPCX Plans to build Orbital Data Centers!

Here’s what they will need:

- 1B Optimus Robots from Tesla🦾
- 10M Tons launched a year🚀
- 1 TW+ of Solar Energy☀️

EVERYTHING we currently know about Claude Mythos.

Release is imminent.

Claude "Fable" spotted in Google Cloud.

Extremely likely that we get the Mythos model today.

Probably the biggest model release since the GPT-5 package.

I'm putting $100,000 on France to win the World Cup.

I win $625,000 if they do.

Am I an idiot?

It’s time:

🚀 Artemis III Risk Check
New Glenn, Starship HLS, and SLS

New Glenn appears to have the most schedule uncertainty.
Starship HLS hasn’t begun integrated flight testing yet.

But Artemis III ultimately depends on all three systems succeeding.

Which one worries you most?

#starship #Artemis #newglenn

Elon is $161 billion away from becoming the first trillionaire.