Jamie (Taylor's Version)

Your friendly neighbourhood AI Dev Nerd. Podcaster, Gaming, Radio, Chemistry, F1 & Football (#WWFC). All views my own, most jokes stolen. He/Him

[Jamie (Taylor's Version) @TweetsByJT]

https://t.co/4zLeRZ23ao is "officially" live! With fun random events, you too can be a successful vibe coder!

[Jamie (Taylor's Version) @TweetsByJT]

@UTDTrey Gotta be sterling, right?

[Jamie (Taylor's Version) @TweetsByJT]

@phuctm97 new model releases = temporary birth control in silicon valley

[Jamie (Taylor's Version) @TweetsByJT]

@maria_rcks or this for when you want to really leave it to vibes

[Jamie (Taylor's Version) @TweetsByJT]

guys i fixed model picking headache in the @ChatGPTapp - where do I file a PR @thsottiaux?
(mocked up directly in browser app by Sol, btw, very smooth)

[Jamie (Taylor's Version) @TweetsByJT]

@sama I built and trained a modern agentic take on AI Street Fighter -low-latency world/action model for micro-play, LLM-guided strategy. Sol built most of it, Luna is the perfect low-latency LLM partner.
https://t.co/v6w2AzVXEM

[Jamie (Taylor's Version) @TweetsByJT]

About to go live with the first livestream test of something I've been working on for a while - an endless live stream of LLMs playing Street Fighter 2, through an intermediary gameplay model! https://t.co/QnlwI8IxIK

[Jamie (Taylor's Version) @TweetsByJT]

@boardyai thanks can you introduce me to someone who will fund my coding AI subscription addiction. I will (probably) build weird ai-focused things like my academic journal for AI slop, peer reviewed by AI (@JournalOfAISlop)

[Jamie (Taylor's Version) @TweetsByJT]

@idoxvi honestly, this, so damn much. He's so good in a tight space, i feel like if they play a few more to his feet and he does stuff like this, it takes so many defenders out of the game

[Jamie (Taylor's Version) @TweetsByJT]

@mynameis_rob @Pokemonpshot For the love of all things decent, never look up Palworld mods for this exact reason.

[Jamie (Taylor's Version) @TweetsByJT]

GPT-5.6 Sol predicted this earlier. AGI is here and it's coming home.

[Jamie (Taylor's Version) @TweetsByJT]

Purely apocryphal observation - codex running long-ish running terminal commands and presumably receiving their stdout and periodically checking in is an invisible cruncher of usage limits. Notice more-so w/ 5.6 than 5.5

[Jamie (Taylor's Version) @TweetsByJT]

Ok, Sol's cracked, after trying to qualitymaxx with various terra/luna combos, just throwing sol medium at it was soo much better. It's just a better model at it's roots.

[Jamie (Taylor's Version) @TweetsByJT]

@willccbb I think this chart shows the real difference - knowledge reliablity and hallucination are significantly better across the board for sol.

[Jamie (Taylor's Version) @TweetsByJT]

@willccbb I feel the same about the GPT-5.6 models. Yes Luna max benchmarks similar to Sol low or Terra xhigh, but they approach and solve problems in different ways because of the underlying model depth.

[Jamie (Taylor's Version) @TweetsByJT]

@thdxr Quick, Sol-ultra, rewrite it to rust, no mistakes.

[Jamie (Taylor's Version) @TweetsByJT]

@thsottiaux "Tibo + Theo Talk Tech" podcast when?

[Jamie (Taylor's Version) @TweetsByJT]

My only conclusion can be that plans just aren't getting as much usage as they used to. Deeply frustrating because I have gotten so much out of the OAI ecosystem. It's been a sub I've been proud to recommend to others. Now? I'm hesitant.

[Jamie (Taylor's Version) @TweetsByJT]

I'm growing increasingly frustrated at my ChatGPT plus sub @OpenAIDevs - even a couple months back I could comfortably work and do personal projects in codex without hard micromanaging my usage limits, but it feels like it's a lot worse.

[Jamie (Taylor's Version) @TweetsByJT]

Because based on token output benchmarks and costs, 5.6-terra(medium) should be so much more cost efficient than 5.5-medium, but comparable size and complexity tasks just chew my 5h limit so much faster.