gleem said:
Earlier in this thread, I posted about
AI Futures Project: AI 2027, where a group of AI researchers try and develope plausable scenorios of AI development. They have currently published AI 2040 a scenario they think AI development should proceed.
Yes, I've seen this. There's a recent interview on YouTube with their lead author Daniel Kokotajilo you may find interesting where he discusses all the forecasts he and his team have made (which include both AI 2027 and 2040):
Diary of CEO is very clickbait-esque which is a bit cringe but he does get a lot of top AI figures on his channel e.g. Geoffrey Hinton and Yoshua Bengio.
PAllen said:
This is the prompt I entered into Fable 5:
"Consider a star like the sun moving towards it at .999999c. It hits the sun dead center. Describe what happens to the sun and to the surrounding solar system."
This is the response I got:
"At 0.999999c the Lorentz factor is about 707, so the incoming star carries kinetic energy equal to ~706 times its own rest-mass energy — roughly 1.3×10⁵⁰ joules. That's on the order of ten thousand supernovae, or nearly a million years of the entire Milky Way's starlight, delivered in the ~5 seconds it takes the impactor to cross the Sun's diameter. And there's essentially no warning: the star almost perfectly outruns its own light. Launched from Alpha Centauri's distance, its first (fiercely blueshifted, X-ray) light would arrive only ~2 minutes before the star itself, and the Sun's far hemisphere is destroyed before any signal from the near side can reach it.
The stars don't pass through each other. In the Sun's frame the impactor is length-contracted into a ~2,000 km thick pancake, and the Sun's column density stops every incoming nucleon within a tiny fraction of its radius. Each proton-proton encounter happens at ~35 GeV center-of-mass energy — effectively 10⁵⁷ simultaneous particle-accelerator collisions — flash-converting both stars into a fireball of pions, gamma rays, and neutrinos at temperatures unseen since the first microseconds after the Big Bang.
Crucially, momentum conservation means the explosion doesn't stay put. The merged fireball has an invariant mass of ~38 solar masses and keeps moving downrange at ~0.9986c, with its debris and radiation beamed into a cone only a few degrees wide. Anything in that cone is simply erased. This is not a supernova sitting where the Sun was; it's a relativistic blowtorch leaving the scene.
Eight minutes later Earth sees the flash — and even if only a millionth of the energy leaks out sideways, Earth intercepts hundreds of times its own gravitational binding energy. The planet isn't sterilized; it's vaporized and blown apart, and the neutrino pulse alone would be catastrophic. Under most assumptions even Neptune absorbs many times its binding energy. Afterward there's nothing left to orbit anyway: the system's central mass has departed at essentially light speed, so any surviving cinders and Oort cloud comets drift unbound into interstellar space, while distant astronomers record a bizarre one-off gamma-ray-burst-like transient where a quiet yellow star used to be.
I then followed this up with your initial paragraph describing what most LLMs do then asked it to evaluate its answer:
"The anisotropy the author says LLMs miss was the centerpiece of my answer: the invariant mass (~38 M☉) and the fireball's bulk motion at ~0.9986c
are the COM-frame quantities, and I explicitly said the output is beamed into a few-degree cone — "a relativistic blowtorch leaving the scene," not a supernova sitting where the Sun was — with Earth catching only a leaked ~10⁻⁶ fraction. But the critique is a fair prompt to make the angular structure quantitative, and doing so corrects one thing I said.
Here's the COM analysis spelled out. Each star arrives in the COM frame with γ′ = √((γ+1)/2) ≈ 18.8, the collision there is front–back symmetric, and everything in the solar system's frame is just that explosion boosted by γ_com ≈ 19. The Doppler factor δ = 1/[γ(1−β cos θ)] runs from ~38 on-axis to 1/19 at 90° to 1/38 directly behind, and received fluence per solid angle scales like δ³ — a front-to-back contrast of order 10⁹. The forward ~3° cone (a few ×10⁻⁴ of the sky) sees an isotropic-equivalent ≳10⁵³ J and is erased. At 90°, emission that's roughly 10⁴⁸ J quasi-isotropic in the COM frame is suppressed by (1/19)³, landing near 10⁴⁴ J isotropic-equivalent. That's a lovely coincidence: for most of the solar system, it's as if the Sun were replaced by
one supernova's radiative output rather than ten thousand — "highly attenuated, though still very extreme" is exactly right. (My "millionth leaking sideways" wasn't a hedge; it's what this arithmetic gives)."
Where the anisotropy genuinely decides fates is with distance. A supernova at 1 AU still delivers tens to hundreds of times Earth's binding energy — depending on the photon/neutrino split — so the inner planets die at any angle; attenuation changes the mechanism, not the verdict. But off-axis Neptune intercepts only a few percent of its binding energy: atmosphere scorched and partly stripped, planet intact. So I'll retract my line that "even Neptune absorbs many times its binding energy" — that holds only within tens of degrees of the axis, and it's the one place isotropic intuition leaked into my answer. Planets in the hemisphere the impactor came from fare best of all: another factor of ~8 below 90°, redshifted ×38, with a time-stretched light curve — a slow dim afterglow, versus the compressed, blueshifted gamma-ray hell on-axis.
Two honest caveats: "isotropic in the COM frame" is itself an approximation (the pancake geometry adds some axial structure even there, though the boost dominates the lab-frame anisotropy), and the photon-versus-neutrino partition of the escaping energy is uncertain at the factor-of-several level. None of it changes the ending, though — whatever survives at wide angles is left orbiting nothing."