Ron

Building: Try That LLM.

Using LLMs via API?
Wondering if you should change models?
Try That LLM can help.

[Ron @ronzilla]

I'm beyond excited to announce Try That LLM: a service for people using LLMs via API. Bulk-test your project's prompts against dozens of LLMs, automatically test them against every new LLM as it ships, and set up LLM judges to score the outputs.

Try That LLM is designed for scenarios like these:

"Our product uses 20 or so basic prompts...but wow those LLMs are pricey. If we switched some prompts to a cheaper LLM, would quality suffer?"

"How do I test my prompts against every new LLM that appears? I don't want to be copy & pasting every time something new ships."

"My manager wants to know if we should use LLM XYZ, they read about it on Hacker News. I guess I'm spending my day figuring that out."

"I just want something to score the responses and tell me the best one"

"Which prompts are costing us the most?"

If you get a chance to try it out, I'd love your feedback/comments.
https://t.co/QKJRlOAuZM

[Ron @ronzilla]

@maxescu These are great!

Do you find you have to iterate on your prompts much?
i.e. you have a vision for the eventual chart, and it only takes one or two prompt iterations to get it right?

Or does it take many iterations, tweaking as you go?

[Ron @ronzilla]

@lkr It seems like "regency bias" should indeed already be a thing: it's whatever the King & Queen like...

[Ron @ronzilla]

@AIWarper Did the video get cut off? It's only 48 seconds for me.

So I assume the prompt to Gemini included basically "use this output format" with an example, and Gemini filled in all the shot start/stop times, with shot notes?

Cool!

[Ron @ronzilla]

@maxescu Super cool, thanks for posting this, with real-world math.

I don't know if you've tried LTX 2.3 for dialogue yet, but the performances are shockingly good. Intersperse some character/performance notes in between every sentence of quoted dialogue, and the results are top-notch.

[Ron @ronzilla]

@AIWarper Eleven Labs is shockingly good.

You might enjoy the streaming version of their TTS API...

[Ron @ronzilla]

@maxescu @invideoOfficial Wavespeed and Fal let you toggle web search on/off too. They don't have a friendly workflow UX, but if you're using like n8n you should be in business.

[Ron @ronzilla]

@eyishazyer This looks amazing--what tool is it from?

[Ron @ronzilla]

@fofrAI Those drink cans look amazing!

[Ron @ronzilla]

Do you hand-edit/compose your json?

I tried converting that json to toon format (which uses spaces), and got basically identical images...feels a little more readable...fwiw!

---

meta:
quality: ultra photorealistic
resolution: 8k
camera: iPhone 15 Pro
lens: 24mm wide
aspect_ratio: "9:16"
style: "raw iPhone photo, natural apartment lighting, zero post-processing, authentic GRWM"
character_lock:
age: mid-late 20s (clearly adult)
ethnicity: Mediterranean / Middle Eastern / Southern European (olive skin tone)
hair:
color: "dark brown, almost black"
style: "sleek pulled-back low bun, center part, very smooth and polished"
eyes: "dark brown, almond-shaped, striking"
face:
shape: oval with refined features
cheekbones: high and defined
jawline: sharp and elegant
skin: "smooth olive complexion, natural texture"
accessories:
glasses: "oversized tortoiseshell frames, intellectual chic style"
body:
type: "slim athletic, tall model proportions"
legs: "long and lean, fully visible"
scene:
location: modern high-rise apartment living room
time: golden hour / late afternoon
atmosphere: "bright, airy, natural luxury"
camera_perspective:
pov: "iPhone on tripod, self-timer shot"
camera_mode: standard back camera - NO portrait mode
distance: 3-4 meters back to capture FULL BODY head to toe including feet
angle: straight-on at chest/waist level
framing: "vertical 9:16, full body centered from head to floor, NO CROPPING OF LEGS OR FEET"
focus: everything sharp - her AND entire apartment background clear
floor_visible: bottom of frame shows floor beneath her feet
subject:
action: posing confidently for GRWM outfit photo
pose:
position: "standing center of living room, floor-to-ceiling windows behind"
stance: "confident - weight on one leg, hip popped, feet visible"
legs: "full legs visible from hips to feet, natural stance"
hand_on_hip: "one hand placed on hip/waist, elbow out"
other_arm: relaxed at side or adjusting cardigan
posture: "tall, shoulders back, elegant"
head:
tilt: straight or very slight tilt
gaze: direct confident eye contact with camera
expression: "soft natural smile, comfortable and confident"
outfit:
cardigan:
type: oversized knit cardigan
color: forest green / emerald green
fit: "slouchy, worn open"
length: "hip length, ends at top of shorts"
sleeves: "long, slightly pushed up or hanging natural"
texture: chunky ribbed knit visible
top:
type: silk or satin camisole
color: cream/ivory white
neckline: V-neck
fit: "fitted, tucked into shorts"
straps: thin delicate straps barely visible under cardigan
shorts:
type: high-waisted tailored linen shorts
color: cream/beige/natural linen
fit: "relaxed tailored, pleated front"
length: mid-thigh
waist: sits at natural waist
details: "clean lines, quality linen texture"
belt:
type: wide leather belt
color: tan/cognac brown leather
buckle: square or rectangular metal buckle
placement: cinched at waist over shorts
shoes:
type: heeled ankle boots
color: brown/cognac leather
style: "pointed or rounded toe, stacked heel"
height: ankle height
VISIBILITY: FULLY VISIBLE - both boots in frame
accessories[3]: large tortoiseshell eyeglasses,small gold hoop earrings,delicate gold necklace (optional)
hair_makeup:
hair: "sleek low bun, pulled back tight, center part, polished"
makeup: "natural fresh - glowing skin, neutral tones, nude lip"
environment:
living_room_layout[5]: floor-to-ceiling windows behind her showing SHARP city view,modern high-rise apartment (15-25 floors up),"buildings and sky clearly visible outside, NO BLUR",light neutral walls (white/cream),light wood or gray flooring fully visible
furniture_left_side[3]: modern sectional sofa in neutral tone (beige/taupe),"textured throw pillows (olive green, cream tones)",sofa faces camera angle
furniture_center[3]: large rectangular marble coffee table with terrazzo/speckled pattern,styled with coffee table books stacked,decorative objects on table
furniture_right_side[3]: modern console table or media unit,light wood finish,minimal styling on top
flooring[2]: light wood planks or neutral tile,area rug under coffee table (neutral textured)
lighting:
type: natural window light flooding entire space
time: golden hour warm afternoon
effect: "soft backlight from windows, warm glow on everything"
shadows: natural soft shadows on floor
city_view:
clarity: 100% sharp and detailed
visible: "apartment buildings, sky, urban landscape all clear"
vibe:
mood: "effortless chic, apartment lifestyle content"
energy: "confident creator, showing off styled outfit"
content: classic GRWM full outfit post
photography_rules:
critical_framing:
full_body_requirement: MUST show complete body from head to feet
feet_visibility: both feet and shoes fully visible in frame
floor_space: show floor beneath and around feet
no_cropping: "do NOT crop legs, ankles, or feet"
headroom: small amount of space above head
camera_distance: far enough back to capture everything
camera_specs:
mode: standard iPhone photo - NOT portrait mode
focus: everything sharp - no background blur
depth: natural perspective only
exposure: balanced for room with backlight from windows
authenticity:
no_artificial_blur: true
no_bokeh: true
background_sharp: true
natural_iPhone_capture: true
zero_post_processing: true
realism:
skin: "natural texture, real pores"
lighting: true natural light behavior
colors: iPhone natural warm tones
sharpness: crisp but natural

[Ron @ronzilla]

A deeper dive:
https://t.co/mFBsrn12ks

[Ron @ronzilla]

After trying a few approaches for translating text from Ukrainian with 10 LLMs:

"Translate this into English" vs. "What does this phrase mean": Asking for the meaning generated more thorough responses than just asking for the translation.

"Without preamble, commentary, or quotation marks, translate this into English": On the other end of the spectrum, this worked to get just the translated text itself, without the LLM's commentary

[Ron @ronzilla]

All the prompts and responses in a table on Try That LLM:
https://t.co/fol9EKpCr3

[Ron @ronzilla]

Feb. 24 makes it four years since Ukraine was invaded by Russia. I thought I'd test out 10 LLMs and see how they did with Ukraine-related topics and translations. Ukrainian is spoken by 30-40 million people, around eighth-most in Europe, so it's an interesting test. I had a native Ukraine speaker friend look at most of the outputs to weigh in on their accuracy.

I'll post snippets here over the next few days; if you'd like the full breakdown and the actual prompts + responses, links are in the comments.

[Ron @ronzilla]

A post diving into more detail:
https://t.co/mFBsrn12ks

[Ron @ronzilla]

@AIWarper There are some creators who are great at highlighting open source/weights stuff:

https://t.co/DRipBKgGnL

https://t.co/vqsf78pbkG
(AI voice is terrible, but content is decent)

@mickmumpitz

[Ron @ronzilla]

@iannuttall Molmo is made for this, try https://t.co/qsTWczzPut . Open weights, can run it on their playground, or on Wavespeed for ~6 cents per footage-minute.

Qwen has a set of VL models too, but you might need to do your own frame extraction.

[Ron @ronzilla]

@d4m1n Nice tip, thank you sir. And thank you for including the prompt!

FWIW, if the window is narrow, the left-right text-image alignment causes the text to be overwritten.

But I think you don't care, because you basically made a mini-site just to generate the movie, right?

[Ron @ronzilla]

@yayalexisgay @TessTregellas Nice camera work!

(also nice writing and performances)

[Ron @ronzilla]

@kleneway I'm curious to hear about how this turns out--I wonder if ~5s screenshots would be enough to get the context of what a user is trying to do...

[Ron @ronzilla]

@d4m1n Also add "Make no mistakes!!!1!"