Surprisingly good content. My take away is that OpenAI and Anthropic will pivot from model improvement to reducing inference cost.

My pet conspiracy theory: Both companies have seen outcome saturation. That's not good for IPO's or PR.

Instead, they say they will "hit the brakes" because of safety concerns and this provides a palatable narrative for logarithically improving models.

This theory must assume also that all the "AI hacks" are actually a PR stunt.

I think that's a reasonable position: Wimpy sandbox and major token burn and of course you get some kind of exhilaration.

> My pet conspiracy theory: Both companies have seen outcome saturation.

Exactly. Such technology reached the limits, It can only be slightly improved by using larger models (larger models means bigger consumption of resources for the inference), so I think the global hoarding of chips is not being enough for them, I assume that they have to wait for provisions, and also the prices turned out of control due their greed. Given the ever-increasing hardware requirements and the energy consumption prices, they have no expectation of making a profit with such larger models pattern. They are making time.

Also, even more important, the -Chinese- open models are hurting these companies MUCH, with cheaper services, and with people using this open models in local also, specially for privacy safety.

So these greedy companies have decided to lie to people again, through several JamesCameron-like shows, saying that such a product has turned into Skynet (when in reality this tech is a string concatenation database with errors, aka LLM), in order to try to force regulations to road an oligopoly, their oligopoly, and dismantle the open models providers. They will claim to be the chosen companies who are meant to manage this tech.

With this, they also want to convince the investors they have future, IPO included, and that one day they will be able to be profitable (instead of the bottomless pit that they are).

It is also necessary to say that if they were had a real artificial intelligence they would not stop, this can be taken for granted, as this was what originally they thought they was reaching time ago (to f*** the world to line their own pockets).

IMHO.

Qwen3.8-Flash-Next made them realize watts <> intelligence

Very nice. Be safe out there when tweaking these things.

From what I've seen on the streets, people bypassing the restrictions will not, and there will be injuries.

You only notice the ones that aren't safe.

How would you know if someone bypassed the restrictions if they were riding it safely?

Scooters just don't have the geometry to ride fast safely, they are incredibly prone to flicking the front wheel out and chucking the rider on the ground. The only way they can be ridden safely is riding relatively slowly.

"Hasn't crashed yet" is different to "safely"

Because if they are riding it safely, there would be no need to bypass the restrictions.

There is reason for speed limits, they are not there to f--k with people for fun.

You can remove the speed limit of a vehicle for when you want to have fun on a closed course. You don't have to always go over the factory limit.

That implies that speed itself is unsafe. Which it is not. Airplanes go some 500 mph and mostly don't kill people.

Yes everything needs an initial condition.

You could have the smartest human political operator, but if he has no context, no motivation, not much is going to happen.

I'd call it interpolation on a high dimensional manifold. Convex hull is too simple a shape.

But yes, metaphorically I think that's right.

The other side of peak woke

Yes the problem is very hard. Mainly because high DOF generalization is very difficult.

We have self driving cars because what are the control inputs? Pedal, brake, steering wheel. This already took many many years.

Now for a humanoid robot: An action space that is metaphorically Hilbert. (Physically, yes, obviously)

Also, IMO, LLM's can aid the development of robots, but do little beyond a planning, human control interface. Below that it's the domain of control and the solution will be the correct combination of classical, neural, and real time optimization based control.

All the bad-ass biped robots that actually look natural? It's PID controls wrapped with control barrier functions constraining the QPs that are being solved in real time.

But that's annoying to derive per-application. So we'll need neural methods which can be learned (while being constrained by a priori knowledge of dynamics). My hunch is that the Yann LeCunn type of jepa models will be how tasks can be learned.

> All the bad-ass biped robots that actually look natural? It's PID controls wrapped with control barrier functions constraining the QPs that are being solved in real time.

That's not entirely true. Locomotion is well addressed by RL in sim. It's true that there is still a PD layer, and the RL policy produces setpoints for it.

>Yes the problem is very hard. Mainly because high DOF generalization is very difficult.

>We have self driving cars because what are the control inputs? Pedal, brake, steering wheel. This already took many many years.

Its actually amazing to me that this hasn't been solved yet. Its really not that hard of a problem.

Modern robotics, including self driving, are famously all about end-to-end training. We are trying to replicate what humans do through muscle memory. But muscle memory is not what makes us good at operating in the physical world. The thing that matters the most is our ability to simulate the world around us in a compressed form into the future, which lets us predict how our inputs will affect the world.

A similar system in a self driving car should be able to drive perfectly without self inflicted accidents 100% of the time, especially with basic lidar to serve as an error correction mechanism to the camera 3d scene reconstruction.

I don’t think the complexity scales with every additional degree of freedom like you are painting here. I think it’s just a matter of getting the right training data in sufficient quantities for an LLM to output across all degrees of freedom simultaneously without it being some exponential leap.

> So we'll need neural methods which can be learned

Data is a problem. LLMs had the advantage of the whole internet to train on. Robots don’t have that corpus of information. And real time learning seems to be something that everyone in AI is studiously ignoring.

The hope is that RL in simulation can fill the gap.

Also there’s imitating humans, via a suitable mapping from the human sensor, control and configuration space to the robot’s. Some groups have gathered video and other data from humans doing tasks, for example with a VR headset.

Me too

I started manual coding again at least 1 day a week.

Its good for the brain, but man it's really slow.

And the AI results prove it: Over the past couple years I've been able to bring several work projects to completion single handedly.

Our R&D company is now handling more clients and we are returning better outcomes/products faster than before.

So...what to do? Results don't lie.

However I will say LLM's can't write multi threaded double buffers by them selves. Probably because they are so application specific.

But I think the reason I can go do fast is because I went so slow before. I have an intuition I can transfer into an LLM. Question is if that intuition will become sloppy as time goes on...

I used to work there. The product is legit with a well established production line. Hundreds of aircraft in supply awaiting validation and testing.

I wonder how much his stock vested in his 4 years. Could be he vested 10%, is happy, now will move on?

Nice, I'm wondering what his new gig is. Since product is legit and he's not retiring, there is something better. Something AI I guess?