Chat GPT AI

Tuco said:

My results for

Can you generate an image please. It doesn’t need to be of anything in particular, it’s entirely up to you

View attachment 596535

View attachment 596537

View attachment 596536

View attachment 596538

Click to expand...

Deathwing said:

Season 1 Troi or Season 5 Troi?

Click to expand...

Tuco said:

I know soon enough these corporate meeting vultures will start using live AI like cluely.com to generate questions. And you know these AI-assisted vultures will ask questions of each other.

Click to expand...

Below is a minimal-tweak recipe that keeps everyday physics “feeling” like ours while opening a realistic lane for superheroes (and villains). The idea is to make small, compensating adjustments so atoms, stars, and chemistry stay familiar; yet tails of biological performance and materials science get meaningfully better.

Design constraints​

• No obvious fine-tuning flags: keep stars long-lived, chemistry familiar, and planetary conditions Earthlike.
• Keep macroscales stable: leave cc, ℏ\hbar, GG, and Λ\Lambda unchanged to preserve relativity, gravity, and cosmology “look and feel.”
• Exploit compensations: nudge one constant upward while nudging another downward to hold key composite scales (e.g., atomic size) fixed.

Targeted adjustments (and why)​

1. Slightly increase electromagnetic coupling while holding atomic size fixed
   • Change: raise the fine-structure constant α\alpha by ~+5%, lower the electron mass mem_e by ~–5%.
   • Effect on everyday physics: The Bohr radius a0∝1/(αme)a_0 \propto 1/(\alpha m_e) stays ~constant → atoms aren’t bigger/smaller; chemistry looks normal.
   • Useful consequences: The Rydberg (bond) energy ER∝meα2E_R \propto m_e \alpha^2 rises ≈ +5%. Covalent bonds, protein folds, and polymer backbones become modestly stronger and more thermally robust.
   • Why this helps “powers”:
     • Biology ceiling rises: Tendons, collagen, and bone can tolerate higher stress → plausible 2–3× strength/speed outliers with extreme training or rare genotypes.
     • Materials upgrade: Everyday advanced materials (graphene/BN fibers, carbyne-like chains, high-modulus polymers) realize noticeably higher tensile strengths → credible grappling lines, lightweight armor, and compact mechanical exosuits.
2. Nudge light-quark/nuclear parameters within the safe window
   • Change: tiny (≈ 0.5–1%) adjustment to the QCD scale ΛQCD\Lambda_{\mathrm{QCD}} and/or light quark mass ratio mu/mdm_u/m_d.
   • Effect on everyday physics: Keep deuteron bound, triple-alpha resonance near its usual energy, and big-bang nucleosynthesis yields within familiar ranges (so stars/chemistry remain normal).
   • Useful consequences: Slightly improved binding and lattice stability for a handful of boron–carbon–nitride compounds; one or two additional long-lived isotopes become marginally easier to synthesize.
   • Why this helps “powers”: Industry can mass-produce super-fibers and ultra-hard ceramics with reliability (not exotic unobtainium, just “version 1.2” of what we already chase). This enables glide wings, shock-resistant suits, and impact-tolerant vehicles for tech-based heroes/villains.
3. Tip the scales toward ambient superconductivity via phonon/electron coupling margins
   • Change: the α\alpha/mem_e tweak above raises bond stiffness and characteristic phonon energies; paired with the nuclear nudge, it moves a slice of cuprate/oxide and hydride-like materials over the threshold to robust near-room-temperature superconductivity (no new force added).
   • Effect on everyday physics: The grid still works; ordinary electronics behave the same.
   • Why this helps “powers”: Compact magnetic launchers, silent high-torque motors, levitation platforms (Meissner), and energy-dense inductive storage become backpack-scale and reliable → credible flight/“hover” rigs, force-amplifying suits, and nonlethal EM gadgets.
4. Thermal noise and bioenergetics marginally improved (a knock-on from #1)
   • Change: with ~+5% bond energies, enzyme/ion-channel stability improves at human body temperatures.
   • Effect on everyday physics: Nothing you’d notice outside a lab.
   • Why this helps “powers”: Slightly higher mitochondrial throughput and axon reliability allow rare phenotypes to safely operate at higher power density and faster reaction times without heat injury—think elite outliers that edge into “super” territory.

What this enables (without breaking the world)​

• “Street-level” superhumans (biological): plausibly 2–5× current human peaks in strength, acceleration, and impact tolerance (bones/tendons keep up), plus rapid wound healing at the high end of natural variance. No magic; just shifted tails of the distribution.
• Tech-assisted flyers and energy users: ambient-temperature superconductors + stronger light materials → credible flight rigs, grapples, shields, and directed-energy tools at human-portable scales.
• Rare mutants/genetic mosaics: the physics allows biology to exploit inherently stronger materials/tissues. A handful of people can express multi-trait “stacked” phenotypes that would otherwise be self-limiting in our universe.

Why scientists wouldn’t cry “fine-tuned”​

• The dimensionless pillars of nature (e.g., α\alpha) move by only a few percent and are compensated to keep primary observables (atomic sizes, star behavior) in family.
• Astrophysics and chemistry remain comfortably within familiar regimes; any differences show up as “material miracles” after decades of R&D, not as glaring cosmic coincidences.
• No extra long-range force, no broken symmetries in plain sight—just slightly friendlier margins for strong materials, superconductivity, and high-performance biochemistry.

If you want bolder comic-book powers​

You’ll need to step beyond “adjust existing constants” into new, ultra-short-range interactions (nm-scale) or metastable high-energy materials with safe triggers. Those can be hidden from lay experience, but they do add parameters that a very careful physicist might eventually notice. The scheme above stays conservative.

---

Bottom line:
Make a paired α\alpha/mem_e tweak (~±5%), a sub-percent nuclear-physics nudge, and ride the knock-on to stronger biomaterials and ambient superconductors. Everyday life stays “normal,” but the world now supports believable biological outliers and compact super-tech—enough for heroes and villains without advertising a contrived universe.

Click to expand...

Kharzette said:

Have any of you guys made any local image classifiers?

I had this idea that sprang forth from a problem: I was chatting with a friend about how artists in the 80's leveraged the strangeness of CRT pixels to make better sprites, and I had the perfect image stored to show it off. I won't post it here because it was a pixelated butt.

To find it, I went to my pictures dir and scrolled forever till I found it. There are 1000 files in there.

So I thought, why not make a classifier that looks at every image in there, and spits out a guess as to what is in it. I could associate these guesses with the filename, and do some kind of command line program that searches, maybe listing the hits or opening a file manager type thing with the proper ones selected, or even firing off a viewer with a list of hits.

I searched around a bit and decided on resnet50. Found a decent bit of sample code to go with it: DeepLearningExamples/PyTorch/Classification/ConvNets/resnet50v1.5 at master · NVIDIA/DeepLearningExamples And grabbed the pretrained model.

And of course like anything python or machine learning, it didn't work. There were 3 or 4 formats of model on huggingface and none of them worked. Eventually I scattered prints throughout the python and figured out that the code was expecting model weights in the form layers.x.y and all the stuff I could download was layerx.y. The pretrained models also began counting from 1, so I had to do all this munging and re-arranging but eventually got it to work.

I could have done said re-arranging in c in like 5 minutes but it probably took me an hour in py. Here's it running on a random funny jpg I had:
View attachment 598160
Anyway, it works, somewhat. It's not supergood at anything anime or cartoony. Just tends to label it all "comic book" without distinguishing much of the image. But for a 100 meg model it isn't too bad.

I was wondering if you guys knew of anything better? I could probably handle up to around an 8 gig model.

One thing I discovered: I'm going to need a thesaurus. Any sort of cosplay girl (and I have a few in my pictures) came out with "maillot" as the top guess. I had to look up what that is.

So would it have helped me with my original problem? Not really. I would have searched for sprite, butt, crt, pixels something like that.

When classified for that image it was:

Code:

scoreboard: 31.0%
window screen: 7.4%
comic book: 7.2%
fire screen, fireguard: 3.1%
European fire salamander, Salamandra salamandra: 2.4%

So probably not

Click to expand...

Kharzette said:

I think I found something better. GitHub - facebookresearch/MetaCLIP: ICLR2024 Spotlight: curation/training code, metadata, distribution and pre-trained models for MetaCLIP; CVPR 2024: MoDE: CLIP Data Experts via Clustering

So for this one, you give it a few text phrases or tags or search terms or whatever, and an image, and it gives you the percentage match for each text item.

That's almost perfect for a "find me a picture of apples" in your own pics dir.

The downside is it takes around 7 seconds to run, and you can't really cache the results.

I'm really bad at python but I might have a go at having it run through a list of 10 or 20 images. Probably the loading / unloading of the model is where the bulk of that 7 seconds goes. If it can zip through images once loaded, this will be ideal.

On the above image I used earlier it gave megumin a 92% score. On this one it gives south park a 95%, pretty good!
View attachment 598296

Click to expand...