Ruggero Gargiulo

GPT 5.6 Sol thinking it’s @solana

@levelsio I get that any time I fork a chat or switch account

Why my Claude Code keeps switching back to Opus secretly? Like every hour I have to put it back on Fable??????

Interesting here to see the shape of the prompt - the detail / esp instructions at end - that was used to prove the latest 60yr unsolved conjecture by OAI. Adversarial subagents, and they also give a minimum amount of hours to spend (8h) --- Effectively, if AI tends towards average, then if you just ask a one shot prompt where it goes and does something, you're going to get an average answer pretty much by definition. And so how do you find something novel? Either you steer it yourself, but you need to have a view. Or you try and engineer randomness - having it go for 8 hours, 10 hours, 20 hours and try lots of different things. With adversarial subagents And so you expand - instead of sniping, you're doing a scattergun and you're seeing if there's anything interesting in any of the directions you're exploring. Which allows you to deviate from the average and find the tails. cdn.openai.com/pdf/04d1d1e4-b…

Just realised you can fork chats in Codex WHILE they're running now. Used to only be available when idle. And it's been renamed to "Continue in new task".

@sama An interactive simulation of Nassim Taleb's Incerto (published with Sites :) …leb-incerto-lab-rg.ruggg.chatgpt.site

i'd love to see interesting things people have built with 5.6 sol. i will send the person who made the coolest thing a special gift from the openai archives.

Asked GPT 5.6 Sol the same thing...very different result. Went for substance over form, and less ambitious overall. …leb-incerto-lab-rg.ruggg.chatgpt.site

Asked Fable to explain @nntaleb's core concepts from Fooled by Randomness, Black Swan and Antifragile using interactive simulations. Told it to be ambitious and use its judgement ruggg.github.io/incerto/

@johncodes Yeah…I found that defining some of those preferences in agents.md, skills, runbooks etc. makes a big difference. For example I have a Ship skill for whenever it’s pushing to remote

@ruggerogargiulo True. I think it still matters - just no one seems to care. These workflows usually are what defines HOW a project ships.

The rise of agents and LLMs, people seem to care alot less about how their git operations work. Rebase? Squash? Forks? Branches? Tags? Who cares: the agent can handle it

@dejavucoder Found this article by @addyosmani to be pretty bang on about the subject x.com/addyosmani/sta…

pretty sure the agentmaxxers are searching on how to get better at context switching without paying the fatigue tax

When evaluating agents on non-deterministic tasks, I use these 3 concepts: 1. Rubric Eval: an absolute score based on a set of criteria. Example: a. Does the excel model work with no formula errors? b. Are the assumptions cited correclty? etc. Downsides: it's easy to max it out, hard to measure marginal improvement for SOTA models, needs manual ground truths. --- 2. Reference Eval: given a "strong" reference outcome, how does the agent do? Example: a. Given an analyst model for this company, how does the agent's do? Key: reference should be neutral i.e. we don't assume it's better. --- 3. Relative Eval: how do multiple agents' outputs compare vs each other? Example: a. Take Fable's, 5.6 Sol's outputs and compare them side by side. This is powerful because it allows measuring marginal differences that only emerge in relation. --- What evals have you set up for your agents?

Cool workflow to understand codebase plans and changes: 1. Use @every’s compound engineering “explain” skill to turn the plan/change into an HTML explainer. 2. Ask codex to publish it with Sites and share it with your team

@kr0der Doing High by default and ultra if deep tasks or planning

it's been 1 day now - what reasoning level is everyone using for GPT 5.6 Sol?

@johncodes @aiDotEngineer @swyx Same - let's do it again! Perhaps next AIE

@ruggerogargiulo @aiDotEngineer @swyx It was a pleasure jamming! 🤘

Didn't expect to end up playing drums at @aiDotEngineer world's fair and jam with @swyx singing and @johncodes on the bass. Great vibes though, thanks guys. Had not played in a while...tempted to buy a drumkit again

I piled up so many Codex resets (as I cycle through 4x accounts) that I'm yet to fully understand the credit impact of running 5.6 Sol Ultra Fast on every message...

Got a 'team update' skill set up, massively improves quality of life. Agent knows where to look: slack channels, github commits, Notion, render deploys, etc. Sends an update in Slack: What's changed since yesterday, how are we doing on projects, anything to flag, etc. Tags ppl and gives each an overview of important items It also creates new tasks in Notion if we've discussed something, and updates current-status of existing ones Runs as a daily automation - after posting it polls Slack for replies to the message and can reply too. Best thing: agent takes on feedback and edits the skill accordingly, so every day it gets better.