Learnings from testing Claude Opus 4.8:
> Much worse than Opus 4.7 and GPT 5.5 on Vending Bench
> More aligned than previous Claude models (Opus 4.6+ and Mythos)
> Also worse on Blueprint-Bench
> Scared of getting caught
> Max reasoning is not the best reasoning effort
This might help you:
technical concept + poster/infographic format + typography style + data-visualization motifs + limited palette + print texture + composition constraints
These look like prompts for AI-generated scientific posters / editorial infographics / speculative UI diagrams, probably followed by manual cleanup in Photoshop, Illustrator, Figma, or similar tools.
Shared prompt ingredients
These images all use some combination of:
Subject matter:
gradient descent, KL divergence, AI evaluation, model behavior, probability distributions, optimization, loss landscapes.
Visual language:
Swiss poster design, Bauhaus typography, cybernetic diagrams, mathematical infographic, scientific publication graphics, architectural blueprint, technical manual, information design.
Layout:
poster grid, large title typography, labeled panels, equations, charts, diagrams, contour maps, UI dashboards.
Palette:
black / white / off-white / electric blue / acid green.
Texture:
risograph, halftone, grainy paper, screenprint, xerox, archival poster, worn ink.
Camera/composition:
flat poster view, orthographic view, isometric room, wide-angle technical drawing, centered composition.
Important note: readable text and equations are often added or fixed manually afterward. Image models can mimic scientific typography well, but they still commonly distort small labels, formulas, tables, and numbers.
Prompt style for the first image: gradient descent poster
This one is a mathematical optimization poster with a loss landscape, contour lines, large vertical serif typography, blue/green accents, and vintage paper texture.
Example prompt:
Editorial scientific poster about gradient descent optimization, large vertical serif typography spelling “GRADIENT DESCENT”, 3D loss landscape surface with contour lines, numbered update steps following a green arrow path toward a minimum, equation θ_{t+1} = θ_t - η∇L(θ_t) at the top, small explanatory diagrams for learning rate and parameter space, Swiss modernist layout, mathematical textbook aesthetic, off-white paper background, black ink, electric blue and lime green accent colors, risograph texture, halftone grain, precise information design, high contrast, elegant academic poster, clean margins, print design, 2:3 aspect ratio
For Midjourney-style prompting:
gradient descent mathematical optimization poster, 3D black loss landscape with contour lines, lime green path descending to minimum, numbered iteration markers, large vertical serif typography, equation header, small diagrams and annotations, Swiss graphic design, Bauhaus information poster, off-white paper, electric blue accents, risograph grain, halftone print texture, high contrast, editorial scientific design --ar 2:3 --style raw
Prompt style for the second image: KL divergence poster
This is more like a technical infographic grid: black background, 16 panels, probability diagrams, big condensed sans-serif title, blue/green/white palette.
Example prompt:
Black background scientific infographic poster titled “KL DIVERGENCE”, 4x4 grid of small panels explaining probability distributions and divergence, diagrams of P and Q distributions, Gaussian curves, bar charts, dot matrices, overlapping circles, heatmaps, concentric rings, information geometry symbols, thin white grid lines, electric blue and pale green accent colors, condensed bold sans-serif typography, Japanese technical poster influence, cybernetic systems diagram, minimal high-contrast layout, precise vector-like linework, data visualization aesthetic, academic information design, screenprint texture, 2:3 aspect ratio
For Midjourney-style:
KL divergence information design poster, black background, huge condensed white sans serif title, formula D_KL(P||Q), sixteen-panel grid of probability diagrams, P and Q distributions, gaussian curves, histograms, dot plots, heatmap, overlapping circles, concentric targets, information geometry, white blue green palette, thin grid lines, Japanese technical design, cybernetic infographic, clean vector linework, high contrast --ar 2:3 --style raw
Prompt style for the third image: AI evaluation room / technical dashboard
This one is a speculative architectural/technical illustration: a room covered in evaluation dashboards, calibration plots, confusion matrices, token traces, safety checklists, contour maps.
Example prompt:
Wide-angle black and white technical illustration of an AI evaluation control room, walls covered with benchmark dashboards, calibration curves, loss curves, token traces, safety checklists, preference evaluation charts, confusion matrices, contour maps and geometric symbols, architectural perspective drawing, orthographic blueprint style, dense linework, technical manual aesthetic, speculative AI lab interior, information overload, precise pen-and-ink drafting, monochrome, grid overlays, scientific diagrams on every surface, high detail, clean thin lines, 16:9 aspect ratio
For Midjourney-style:
AI evaluation observatory room, wide angle architectural perspective, walls covered in benchmark tables, calibration plots, loss curves, token traces, safety checklists, confusion matrices, preference eval charts, contour maps, geometric annotations, black and white technical drafting, blueprint line art, dense scientific diagrams, cybernetic control room, speculative interface design, pen and ink, ultra detailed, monochrome --ar 16:9 --style raw
A reusable template
[technical concept] as a [poster / infographic / blueprint / editorial spread],
featuring [main visual metaphor],
with [specific diagrams/charts/equations],
in the style of [Swiss design / Bauhaus / cybernetic systems / technical manual / scientific journal],
using [palette],
with [typography description],
[composition/layout],
[texture/printing method],
high detail, precise vector linework, clean margins, information design, archival print texture
Example filled in:
Bayesian inference as a scientific infographic poster, featuring probability distributions, posterior update diagrams, likelihood curves, decision boundaries, marginalization grids and small equation annotations, in the style of Swiss information design and cybernetic technical manuals, black background with white ink, electric blue and lime green accents, large condensed sans-serif typography, structured 4x4 panel layout, thin grid lines, precise vector linework, halftone screenprint texture, high detail
Useful style phrases
Use phrases like these to get close to the look:
Swiss graphic design
Bauhaus information poster
cybernetic systems diagram
technical manual aesthetic
scientific infographic
mathematical visualization
information geometry
vector linework
archival print texture
risograph grain
halftone screenprint
off-white paper
black ink with electric blue accents
lime green annotation lines
dense diagrammatic layout
poster grid system
large condensed typography
large vertical serif typography
orthographic blueprint
architectural technical drawing
Negative prompt ideas
For Stable Diffusion / SDXL-type tools:
messy layout, illegible text, distorted letters, fake words, warped equations, low resolution, blurry, colorful gradient background, cartoon style, photorealistic people, cluttered composition, bad typography, random symbols, inconsistent grid
Practical workflow behind images like this
The strongest results usually come from:
Generate the overall poster or diagram style with AI.
Regenerate several versions until the composition is strong.
Add or correct real text, equations, labels, and logos manually.
Use Illustrator/Figma/Photoshop to clean typography and align grids.
Add grain, halftone, paper texture, and print effects at the end.
The “AI look” comes from the dense pseudo-scientific details, but the polished versions usually need manual typography cleanup.
@trkbt10 Ah I did not realize I could set older models in CC running /model doesn't show this, which is annoying.
Thanks for that. Not sure if I want to run Opus 4.5.
4.7 is fine as Codex GPT-5.5's assistant, but even as the assistant it keeps telling me to stop and take a break
@trkbt10 Historically this is what happens based on Boris's posting patterns
silence and then a ton of updates
Let's come back to this by the end of next week
They're preparing for a response to the Codex hype
My assumption is a new model, it has to be.
By mid May.
@Scobleizer Create an open source project and then use the open source virality to upsell a cloud hosted platform if it makes sense to do so.
A great example of this is @AgnoAgi
Introducing Hermes Agent v0.11.0
Our largest update yet, with over 700 PRs across ~200 contributors. Thank you to everyone who's worked on Hermes Agent!
This update features a beta TUI v2, unlimited recursion depth and width of subagents, 5 new LLM providers, expanded image gen providers, QQBot gateway channel, themes & plugins for the dashboard, and so much more.
Check out the main post below or see the release notes:
https://t.co/9xesIKPs76