Introducing BF3 FX:
A contest I set up between GPT-6.2 Astra, Fable 5.2, GPT-6 Astra and Opus 5.5
15 years ago, my favorite FPS came out: Battlefield 3
Since I've always been nostalgic about the glitch effect on the EA and DICE logos, I asked these 4 LLMs to rebuild that FX from scratch.
Fable 5.2:
+ Texture, color choice, general transitions
- Better fullscreen distortions
Overall, tops it out in my opinion by having the best blue touch out of them all and one of the best texture
Opus 5.5:
+ Texture, color choice, general transitions
- Not the best matte process
It might not have been the most precise during a matte process of the original EA logo but it produced a good looking texture and FX.
GPT-6.2 Astra:
+ Texture, full screen distortions
- Could've had a better blue tint, transitions
It has the best fullscreen signal distortion in my opinion. As well as a good texture.
GPT-6 Astra:
+ Full screen distortions
- Blue tint, texture, transitions
Overall okay, but the texture work wasn't the best. 6.2 improved on that!
For LUTESIS, the Claude line up chose the best font possible, a very similar one to DICE
I took Opus 5.5 a bit further and asked it to edit a Battlefield 6 trailer with the BF3 FX logo intro it did. And on its own, it applied a blue filter on the BF6 gameplay footage. Again very surprising and a sign of its incredible taste. BF3 had a crazy blue filter, which was Opus' inspiration for this decision I guess. I'm sure Fable 5.2 would do the same.
Let me know what you think below! I'll post more FX or world engine based on popular games or pop culture in the future so make sure to follow! Also let me know which one you'd like next.
I got back 1GB of RAM a few weeks ago with a single prompt
I was discussing with someone that with the latest 6.2 Astra & Sol release marketing vids, we don't really have an emphasis on the freeing and empowering results AI can have in our daily lives
I have a Razer mouse & mousepad (don't judge me) and I like to have certain effects on them, but they come with built-in Razer Synapse software to control them and to always have chosen effects turned on
One day I was running a heavy workflow and noticed my RAM completely clogged, I checked what I could close to free some of it, and I saw that Razer Synapse was taking 1GB of RAM
So decided to just send a very simple prompt to GPT 6.1 Sol to check my hardware, map it out, see how it could apply the effects I wanted
In a few minutes, I had a script that would run every time my computer booted up and the effects were automatically applied, with 0 usage of RAM
We need to see some more of that, let me know where you used AI for the same thing
@imjustnewatai Downgrading my OAI subscription as we speak to give a bump to my Anthropic one
had high expectations, they weren't met, hopefully next month!
@camperaigg I think actually like the result of 6.2 Sol
I think what might emerge is 6.2 Sol aa the worker and Opus 5.5 for review/taste input tbh
tbh
Which is what a lot of people are doing
@The_Alex I asked 6.2 Astra to make an comparison edit of world engines built by various LLMs
Yeah, its bad
Switching back to Opus for the edit sadly, the lack of taste is horrible and I feel will stay backed in the model for quite some time
@tcballard Tibo is no longer in the cooking basement
He's been moved to Cyber basement 🥲
I think lack of taste for design and frontend are backed in the model
How to pass Pangram when writing with an LLM:
I saw a few people interested with this reply so I'll give you a basic prompt/skill I use to create drafts that pass 100% as a human.
I saw a few skills & prompts shared recently but to my taste they're too technical & specific for what the models are capable nowadays. You should give more leeway to your models, they're adults now, they'll figure it out!
The general goal of the prompt is to gather and build a phrase bank that it can use to write the text you want. It will also ground & source the links it found and used for the draft.
It will also restrict itself from using anything else but the phrases in the constructed phrase bank.
I'll explore more ways to write naturally with LLMs soon, so bookmark the tweet!
Here's an example below 👇
@IndraVahan I already made a skill for my agents to use relevant Wikipedia pages to retrieve sentences to write some copy writing and it passes Pangram @ 100%
but happy to test this one out as well, thanks
I had huge expectations for 6.2 Astra & Sol today, I wanted strong taste & ideas for design and front-end, with the untiring work horse spirit of the GPT line up
But I think we're still one or two models away, i hope the day will come, or the other way around, whoever comes first
Read the brief and inspect the material that establishes the subject's facts. For a project, read a representative example alongside the overview. Identify a sequence of concrete details worth writing about and the language and format the reader expects.
Search Wikipedia in the requested language for the actions, objects and processes in those details. Narrow articles about a gesture, craft or technique often yield more usable prose than an overview of the whole topic. Use public concepts in external searches and keep private names, paths and unpublished excerpts local.
Build a phrase bank before composing. Open each promising article, read the passage in context and save contiguous verbatim spans with their article title, section and URL (a revision link when available). Select clauses and complete sentences that express the intended point. Keep the chosen excerpts with their sources while arranging them. Looking up individual words after writing a sentence does not establish passage provenance. Prefer the contributors' own prose. Check credit and reuse terms for separately attributed material.
Compose from the bank. Arrange excerpts into a readable progression. Let them carry the descriptions and actions in the draft. Use whole source sentences when they fit. Keep project-specific names, facts and connecting words economical and identifiable. For a first-person post, open with the author's concrete aim or result. Use sourced clauses for the main ideas and short fragments for necessary joins. If a sentence needs a string of tiny fragments, look for a better passage or recast it. Check that pronouns still have clear referents and that first-person claims fit the brief. Keep distinctive quotations in quotation marks and credit the original speaker. When the user asks for “full Wikipedia”, “100% Wikipedia”, or exclusively sourced wording, assemble the entire reading copy, including its hook and ending, from verbatim contiguous spans. Join spans using only spacing, capitalization and punctuation. Before delivery, match each span to its source and check that the concatenated spans account for every word in order. On revision, retain this mode and repeat the check. Put attribution, personal test results and project facts without a matching source outside the copy. Report verified wording coverage separately from any user-reported detector result. Honor an explicit request for a different balance.
Check each borrowed span against its page for exact wording and each new combination for meaning. Keep factual claims about the subject grounded in the supplied material. Give each span a traceable source, mark any alteration, and distinguish wording from the user's material from your own additions. When asked where each passage came from, provide the clean reading copy followed by a sentence-by-sentence source map. Show the exact contiguous excerpts in order, mark each cut (for example with a vertical bar), and link each fragment, including short joins. Identify any original wording or adjustments explicitly. Credit Wikipedia contributors, link the articles and applicable license, and identify the selection and rearrangement. For a short post, keep the attribution available with the copy. If you cannot retrieve enough suitable text, disclose the gap instead of passing invented wording off as sourced.
How to pass Pangram when writing with an LLM:
I saw a few people interested with this reply so I'll give you a basic prompt/skill I use to create drafts that pass 100% as a human.
I saw a few skills & prompts shared recently but to my taste they're too technical & specific for what the models are capable nowadays. You should give more leeway to your models, they're adults now, they'll figure it out!
The general goal of the prompt is to gather and build a phrase bank that it can use to write the text you want. It will also ground & source the links it found and used for the draft.
It will also restrict itself from using anything else but the phrases in the constructed phrase bank.
I'll explore more ways to write naturally with LLMs soon, so bookmark the tweet!
Here's an example below 👇
@IndraVahan I already made a skill for my agents to use relevant Wikipedia pages to retrieve sentences to write some copy writing and it passes Pangram @ 100%
but happy to test this one out as well, thanks
> "I wrote my own writing tool because I wanted the exact words that were said or written. At first, the agent found the right articles and then went on to generate new text that did not exist in the original text.
Now the first step is to collect excerpted passages. I like to keep the title, the section and a link to the source for each sentence. Then the agent can create a new text by taking words, phrases, and sometimes whole passages from other sources and reframing them.
I get a draft to work with and enough information to identify the source of each passage. I had excerpts from nine articles. When I wanted a shorter version, I could compare the two versions side by side. In the second, nine words from Wikipedia were left: “patterns of fine lines” and “dealt one at a time.”
Two sentences can make sense on their own. What happens when you put them together? Sometimes I make simple revisions to smooth awkward passages. Sometimes I want to keep the sentence and move it. Sometimes I can leave out the rest of the paragraph."
I asked it to write a small draft about what we did so far.
Keep in mind, such writing will pass Pangram but will sound robotic somehow, as a human, you can feel something slightly off.
It'll bring you back to your high school years of searching Wikipedia pages, and putting everything together to sound good/natural.
I think it provides a good base to work with, a good draft that you can quickly improve on.
Here's the full prompt/skill that you can give to your agent 👇
@ducksays I've been doing the same for a few days as well now. Still sad that Fable falls short when it comes to actual characters and people in its sceneries, but we're most likely just one or two models to get rid of img gen models and replace everything by code
Doing this as well for media and art companies, (almost) everything is code, fueled by incredibly refined taste and years of experience combined of the people I work with
We produced so many cool content and designs
I've been creating visual content engines based on personal taste as well to revisit films/series/video games, wonderful era to live in
@daniel_mac8 I'm 🤏 this close to downgrade my OAI subscription to upgrade my Anthropic one, I hope those models come with incredible taste for front-end, while keeping their work horse behavior