The input and output of a GPT it not a sequence of messages in turns: fundamentally the input is the entire context (both from the user and the model), and the output is a probability distribution for 1 single next token.

The chat is only one of many interaction patterns possible with that primitive: one where the "most likely" output token is concatenated to the previous input and this is repeated until a limit is reached or the model produces a special "end of sequence" token.

Representing a more complete interaction pattern as a chat is impossible because the chat abstraction is lossy and could never represent it. Sometimes the abstraction leaks, as with the concept of "temperature", that's not a parameter of the model but of the chat.

In technical terms it lets you stop the transformer at any token, sample alternatives from the probability distribution at that point, pick your desired next token and resume the generation.

This UI lets an expert guide and control the output of the tool, instead of having to fight it. I wish there was a frontend for large GPTs that worked like this instead of the ubiquitous-yet-useless chatbot.

It is possible to protect those two uses you mentioned, and other reasonable uses of land, without at the same time spending state resources in enforcing perpetual, absolute, hereditary, exclusive titles on plots of land.

The green-sensitive cones overlap with the red-sensitive cones, and to a smaller extent also with the blue-sensitive.

Full saturation red and blue are possible by emitting light on the edges of the visible spectrum.

Full saturation green, however, also activates the red and blue cones.

To cover the whole gamut is impossible, but you can approximate it with ~three green tones: a 490nm deep cyan that hits blue and green but not red, ~510nm that hits red and blue equally, and ~540nm the peak of the green cone.

You may have heard of football clubs: If it's too expensive you can pool resources with your friends.

In a great coincidence, just months after being put into operation, a new type of detector meant to study if protons are stable, detected instead the neutrinos coming from the supernova in a way that could be individually timestamped. No other supernova has been visible with the naked eye ever since, or is likely to be seen in our lifetimes.

https://sci-hub.st/10.1103/PhysRevLett.58.1494

The neutrinos arrived 3 hours before the supernova was first seen in the sky. This could mean that neutrinos travel faster than light: that they negative mass. It has not been ruled out yet, but it would take extraordinary evidence for any physicist to admit to a faster-than-light particle.

The core of supernovae are predicted to take around 4 seconds to explode, and during that time they release 99% of the energy that was binding the star together as neutrinos. Once this has happened, however, it takes several hours for the explosion to be visible outside.

If they had 0 mass, you'd expect the burst of neutrinos to last for about 4 seconds. The detected spread was of around 6 seconds.

It was likely that the model for supernovae was missing something. But perhaps neutrinos are slowed down as they interact with other particles on the way here from the supernova. Massless particles always travel at the speed of light, only massive particles can be slowed down. Perhaps neutrinos have mass.

In order to test the basic theory of nuclear fusion in stars, the "Homestake experiment" was set up to count neutrinos from the Sun.

https://sci-hub.st/10.1126/science.191.4224.264

The experiment consists of a big tank of perchloroethylene. Electron neutrinos (and only electron neutrinos) are expected to collide with the Cl-37 atoms to form Ar-37 in the reverse reaction to that of radioactive beta decay. Then the individual Ar-37 atoms can be separated into a gas and counted.

According to the brightness of the Sun, it was expected that around 50 such Ar-37 atoms would be detected every 100 days. Only 17 were observerd, on average. A third of the neutrinos went missing.

A new round of experiments was made, culminating in the fantastically sophisticated Super-Kamiokande, which is sentitive to all other "flavours" of neutrinos, and also capable of detecting the direction they are coming from.

https://arxiv.org/pdf/hep-ex/9807003

It was confirmed that the missing 2/3 of the electron neutrinos expected to come from the Sun had somehow transformed into the other flavours. What's more, muon neutrinos coming from cosmic ray decay in the atmosphere were detected half as frequently in an upwards direction, coming from the other side of the Earth, than in the downwards direction.

The conclusion is that neutrinos must change flavour as they move through space.

Particle accelerators were built in several sites, pointing towards the Super-Kamiokande across the earth's crust, sending a beam of neutrinos of a known kind to measure this "oscillation".

https://arxiv.org/pdf/hep-ex/0606032

Since they have mass, the artificial neutrinos are slightly dispersed by their passage through the Earth's crust. Their arrival time forms three different overlapping peaks, corresponding to three different masses.

The proportion of each kind changes along the path in a sinusoidal fashion, but the mass peaks remain.

The current interpretation is that the mass and the flavour of neutrinos are (almost) "conjugate" properties of the particles, in a similar way to the well-known uncertainty between position and momentum, or between time and frequency in Fourier analysis.

As a neutrino moves through space, its mass can be measured more and more accurately from the dispersion in arrival times. The better we know the mass of the particle, the more uncertain we are of its flavour. If because of the great distances the mass is well known, as is the case with the solar neutrinos, then the three flavours are almost completely entangled.

On short distances the flavour is well determined but the masses are uncertain.

Original NTSC cyans are more saturated than even DCI-P3 cyans.

Typical CRTs use the cheaper, brighter phosphors specified by SMPTE C (the basis for the sRGB gamut) and a circuit that pumps the saturation to compensate.

It's likely your screen uses the better phosphors instead of a colour correction circuit.

It's certainly blurrier than Netflix, but dark scenes, grain and textures are usually much more detailed on a DVD nonetheless.

The case is all thick ABS.

It weighs like 2.4 kg, and the weight is unbalanced.

The USB-C charge only works at 20V, nothing less.

While charging it overheats and spins up the fans.

It came with a TN screen with terrible viewing angles, that could not be used in a brightly lit room. I didn't use the laptop for two months while I waited for a replacement screen from aliexpress.

Keyboard is much thinner, the trackpoint drifts easily.

Camera quality is worse, somehow it cannot handle sun-lit scenes. Microphone and speakers are similar to the T14.

It stopped receiving firmware updates after two years.

It uses about 0.5 W while suspended, so its tiny 48 Wh battery typically doesn't last the weekend with the lid closed.

The motherboard has design issues, a missing protection diode in the headphone jack microphone input ended up frying the CPU due to a ground loop. Meanwhile the T14 has eaten the same ground loop and even a 48V passive PoE in an accident and dealt with it by rebooting. A T450 from 2015 is still running.

If these are JPEGs with a grey or green lower half, it's likely only a few 16x16 macroblocks are corrupted and you can recover the rest.

This cannot be done programmatically because you have to guess what the average colour of the block was, but it can be worth it for precious pictures.