5.6 Sol isn't the latest model, it's 6.0 Astra.

Astra crushes Sol in performance.

"Let's Get Together" (1957): This story features literal suicide robots. Set during a prolonged Cold War, a rival global power infiltrates America with ten identical, human-looking androids. Each android carries a piece of a "total conversion bomb" inside its body. When all ten autonomous robots converge in the same room, they are programmed to self-destruct and trigger a massive nuclear explosion.

"The Feeling of Power" (1958): This story features autonomous missiles that function exactly like modern AI-guided combat drones. In a future where humans have forgotten basic math and rely entirely on computers, warfare is fought using expensive ships and missiles controlled by advanced onboard computers. When a technician rediscovers mental arithmetic, the military eagerly plans to replace these computerized "drones" with cheaper, human-piloted suicide missiles.

But my question is, what mechanism for lack of a better word, allows it to learn and improve? Is it simply a programming language? That doesn't seem quite right......it's almost like there is something else......and I'm not implying "not of this Earth" but it's unimaginable to me that something is being programmed that can't be controlled.

#12 | Posted by brass30 at 2026-09-20 03:14 PM | Reply | Flag:

the encoding here... rcade could fix that if he just tasked an agent to it for 15 minutes...

youtube, josh starmer StatQuest.

It means it's going to set your credit card on fire when you try to train it on your own.

but a decent foundational model can be distilled to run on lightweight hardware. I created RNFM, Reality Navigation Foundational Model, which uses 2 tier distillation. The big model at a data center, the distilled teacher model on "near edge" hardware like a PC or laptop, which then can distil into a far edge model that proved it could reason basic navigation directions using GPS inputs and only 384 params on stock ESP32-WROOM chips.

and the hard part of that is the "reality tokenizer", which can best be visualized by the moment Neo wakes up and sees the world as green matrix symbology.

This just another way of saying you compressed the algorithm after a data center did all the heavy lifting, right?

#25 | Posted by sentinel at 2026-09-20 09:04 PM | Reply

No.

Its called transfer learning, and the biggest version of RNFM so far is training on UH Carya Cluster which is very powerful but definitely not data center scale.

See if I can simply explain this.. you get a model that's good at something, call it riding a bmx bike. The early layers of the model teach it how to balance, steer, and stop. Later layers teach you how to pedal it, how to bunny hop, how to do a pump track.

Now we need a model that can ride a motorcycle, so we freeze the early layers, discard all of the advanced bike knowledge layers, and add a new layer that teaches how the twisty throttle works. Now we've got a much smaller model that can ride a motorcycle. It can't backflip like Travis Pastrana, but it can get around and that's all we need it to do. Congratulations, you now have a distilled child model from a teacher or foundational model.

Compression has nothing to do with this.

and to wrap that up, because we only had to train the last layers on the new motorcycle dataset, we didn't need large amounts of compute power, we can pull it off with a cheap google colab t4 instance, or train it locally if you a single Nvidia graphics card. Very small models you can train on a laptop.

That's what Ukraine is doing on the front lines. There's foundational models for their drones. A guy with a laptop trains specific models loaded with the latest battlefield images streamed in from drones, and that distilled model is loaded into the 1 way quadcopter's guidance for machine vision (like a Nvidia Jetson Nano or something along those lines). It doesn't need to map the entire world, just that local battlespace.

Recursive self-improvement isn't infinite, it can plateau and run into bottlenecks.

Its not. You would fail data science exams.

And Carya Cluster is more like a fraction of a single row in an actual data center. Calling it one diminshes the scale of what they are.

Suffice to say the 2 terms are deeply intertwined. My 2 tier distillation works via transfer learning because a self driving car doesnt have to know how to fly. End result is small, efficient models with low RMSE. Or not small, if the thing needs expansive platform specific sensor fusion and datalink processing.

I should just add some of you to the repo.

Maybe say it like this, you can get a smaller model. Now you need it for a specific task and are adding new training data sets and layers. Your distilled model is now bigger than the original. Compression can be an outcome but its not guaranteed and depends on the goal for the model.