The model card says:

> Post-trained from Qwen 3.5 397B

The model card also says that they use an inference framework based on "SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs" by Shi et al.:

https://arxiv.org/abs/2510.05069

So the sources seem properly attributed.

They only claim that what they did to "Qwen 3.5 397B" has improved the LLM, including, as expected, with "strong performance in Portuguese".

Model Card:

https://huggingface.co/prefeitura-rio/Rio-3.5-Open-397B

Probably the greatest mistake was that arrays of different sizes were different types in Pascal. While C had extremely poor support for arrays in comparison with older languages like Fortran, PL/I or Algol 68, Pascal was even worse, because in Standard Pascal it was pretty much impossible to write a library implementing linear algebra algorithms.

Actually in Standard Pascal it was impossible to write any kind of library, because separate compilation was impossible.

Turbo Pascal was a decent programming language, but only because it had a lot of essential extensions over Standard Pascal, including the ability to write multi-file programs.

The languages designed by Wirth have become very well known, because he and others have written some very good books about them for beginners and they were used in teaching in many places, but all of his languages were quite bad in comparison with the languages that he wanted to replace, because he thought that they were too complex, e.g. Algol 68 and Xerox Mesa.

The incoming power is not the electrical power generated by the solar panels, but the entire power of the light that is absorbed by the solar panels and by the body of the satellite.

Even with a perfectly reflecting body and with SOTA solar panels, the amount of incoming power is at least double in comparison with the electrical power consumed by the datacenter.

Also, the heat radiated is smaller than in TFA, because no radiator is perfectly black at the radio waves in the frequency range corresponding to the ambient temperature.

I am too lazy to make the correct computation, but there was another article linked on HN some days ago where a more plausible computation was done and the conclusion was that the minimum area of the radiators is slightly larger than the area required for solar panels.

This would still be feasible, but in reality the area would have to be even larger, because the radiator cannot have a uniform temperature, the parts where the cooling fluid is incoming will be hotter than the parts where the cooling fluid is outgoing. Moreover, the pumping of the cooling fluid requires extra power that must be added to the power budget.

There is no doubt that it is possible to build a space datacenter, if much more GPUs are installed in it than necessary, to enable to correct the more severe transient errors and to preserve enough capacity after many GPUs become permanently defective, but the cost will not be competitive with terrestrial datacenters any time soon.

Nuclear fission reactors, similar to those used on submarines or ships, would enable very different applications.

Until now, they have not been used for fear of what would happen after a failed rocket launch, when the reactor would fall back on Earth.

This could be mitigated by sending only components of the reactor and assembling it in space.

I do not think that routine exploration of the Solar System beyond Mars will ever be possible without using at least nuclear fission reactors, because it is too slow with chemical sources of energy.

However, the number of slots that are available in Sun-synchronous orbits with permanent view of the Sun is limited, and many potential users want them. So those who desire to build datacenters would have to compete for such orbital slots. There are much less such slots than for geosynchronous orbits. Other countries would certainly be outraged if USA occupied all the available slots with datacenters.

Improved control of the satellites for collision avoidance could allow smaller slots, but maneuvering heavy datacenters would require a lot of fuel, so they might require periodic refueling, greatly increasing the costs.

Genome duplication has happened in many animal groups. Wherever it happens, it enables the descendants to evolve into having bodies that are bigger and more complex, because each ancestral gene is replaced by multiple copies and then each copy evolves differently, becoming able to accomplish additional functions than in the ancestors.

Without a genome duplication, accumulating a similar number of genetic innovations would require a much longer time.

Another example besides the vertebrates is that the ancestor of spiders and scorpions also passed through a genome duplication event.

In animals, the genome duplications are very rare events, because the development of animal embryos is very complicated and after a genome duplication the embryos will usually fail to develop correctly and they would die.

On the other hand, in plants genome duplications and also hybridizations when a genome is doubled by combining together 2 genomes of some closely related plants, are very frequent.

Some of the most important crops have genomes that have been multiplied, either from the same species or by hybridization, i.e. which are called tetraploid or hexaploid, to mark the fact that they are doubled or tripled in comparison with the original diploid genomes. In cultivated plants, this genome multiplication has resulted in a higher productivity in comparison with their wild ancestors.

HBF was initially announced by SanDisk, early in 2025, then early this year Hynix has announced that they have joined SanDisk in producing HBF, and that the common specification will be standardized under the Open Compute Project.

With HBF, it would be easy to make a GPU card with 4 TB of HBF, which could run the biggest existing open weights LLMs in their native unquantized form.

NVIDIA Nemotron 3 Ultra is a relatively big LLM for which a part of the training data is public, but not all of it.

Nobody who has trained a really good and big LLM can afford to make public all the training data, as much of it must have been copyrighted.

The weights for GLM 5.2 will be published in a few days on Hugginface.co.

While I would want very much to have access to the entire training set of a big LLM, I would want that in order to be able to run traditional search tools on it, to get accurate answers, instead of possibly hallucinated answers.

I could not use that dataset to perform the training myself, as that requires too expensive hardware.

On the other hand, with the open weights of even a very big LLM like GLM 5.2, I can run inference on any computer, with the weights stored on SSDs. Obviously, inference will run slowly, probably at less than 1 token per second at the size of GLM 5.2, but that is still useful in some cases.

Because they lose neither capacity nor charging speed at low temperatures, like the lithium-ion batteries, they expect that in the future sodium-batteries will be the best choice in the countries with cold climates.

Another poster has mentioned that BMW also uses EESMs instead of permanent-motor magnets.

BMW uses EESMs as the main motors, on the rear axle, while they use induction motors as auxiliary motors on the front axle.

Besides being cheaper, the induction motors have the advantage that if they are used only as auxiliary motors, you can cut the power supply to them at any time, in which case they will consume nothing.

So their lower efficiency does not matter, because most of the time they are turned off.

Brushes are typically made of graphite mixed with some binder. The graphite conducts the electrical current, but it also acts as a lubricant.

The metallic part that is in contact with the brush is called a slip ring, if it is continuous, like in synchronous motors, or a collector ring if it is segmented, like in DC motors or single-phase motors with brushes.

In 1891, the three-phase induction motor was invented by Mikhail Dolivo-Dobrovolsky, combining the principles of the three-phase synchronous motor previously invented by Mikhail Dolivo-Dobrovolsky with the principle of the induction motors invented by Nikola Tesla and Galileo Ferraris.

Like any inventions, the induction motors of Nikola Tesla and Galileo Ferraris had not sprung out of nothing, but they were based on the experimental observation that had been known for many decades that if you rotate some magnets around a disk of copper, the disk will rotate, even if the magnets do not have any action on the disk when stationary.

Because of the symmetry, it is easier to generate electromechanically three-phase currents than two-phase currents where the phase difference must be precisely of one right angle.

The Dolivo-Dobrovolsky motor is the ancestor of all high-power induction motors, while the Tesla motor can be considered the ancestor of the single-phase induction motors that have been used (more frequently in the past than today) for several household appliances, like washing machines (or reel-to-reel magnetic tape recorders, a half of century ago).

The lower efficiency means a lower range for the same battery, which is why the companies that have used them in the past, like Tesla, have abandoned them.

Permanent-magnet motors have the highest possible energy efficiency, followed by electrically-excited synchronous motors, than by the induction motors mentioned by you.

Both permanent-magnet motors and induction motors do not contain parts that need frequent maintenance, while this property is more difficult to achieve for electrically-excited synchronous motors.

The electrically excited synchronous motors have been known forever, but they had not been used in EVs because of 2 disadvantages.

The first is that traditional EESMs require brushes, i.e. sliding electrical contacts, which are worn out by friction, so such motors require frequent maintenance for changing the brushes.

It is possible to make brushless EESMs, but they require a rotating transformer and a semiconductor rectifier inside the rotor.

The second disadvantage is a lower efficiency than with permanent magnets, which cannot be improved so much as to match PM motors, because the electrical currents that circulate through the rotor windings must generate heat. The lower efficiency also makes cooling more difficult.

Renault says that their EESMs have an efficiency of 92%. This is a good efficiency, even if not as good as attainable with permanent magnets. Losing a few percents in efficiency is an acceptable compromise for avoiding the use of expensive and supply-constrained chemical elements.

What I wonder is whether Renault reaches this 92% efficiency with EESMs having brushes, or with brushless EESMs, and this is what I would have liked to read on the parent Web page.

Brushless EESMs usually had a lower efficiency, so 92% would be impressive for them, while it would look normal for EESMs with brushes.

If Renault has succeeded to make a brushless EESM (i.e. maintenance-free) with an efficiency of 92%, that is something worth to brag about. Otherwise, making a traditional EESM would not be great news, because everybody has avoided those because of the maintenance problem.

However, when a company sells a device, as opposed to providing it for lease, I do not believe that it has the right to not document any feature of the device that is relevant for its usage, like it should also not have the right to impose any constraints on how the owner should use what has been bought.

Obviously, the owner of any kind of things may not use them to perform illegal acts, but that is a constraint imposed by the valid laws, not by the seller of the things.

Today, far too many companies claim to sell things, but they also attempt to control what the owner may do with them. I avoid to buy such things, but my choices are limited by those who buy them, allowing these policies to be beneficial for the sellers.

Because of this lack of documentation, every release of a new version of Apple hardware or software may require the restarting of the reverse engineering work, like in this case, just to keep working the alternative operating system.

The wavelength is the ratio between velocity and frequency, so it changes proportionally.

If you multiply 36 kHz by 8326 m, you get a value only slightly less than the speed of light in vacuum, which is true for the propagation of electromagnetic waves in most gases.

On the other hand, with 170 m, you will get a speed of VLF radio waves in sea water that is much lower than in vacuum.

The speed of electromagnetic waves in most media depends strongly on frequency.

At frequencies corresponding with visible light, only in few materials the speed is lower than half of the speed in vacuum (i.e. the refractive index is greater than 2).

On the other hand, for low frequency radio waves, speeds that are 10 times slower or even 100 times slower than in vacuum are not unusual.

Such antennas and transmitters cannot be installed in a small submarine.

Here a new kind of antenna is used, which is efficient under water even at small dimensions, so it can be installed in small submarines, for communication at distances of up to a few hundred meter.