Connecting people thru technology- @scale_ops ex @axoniusinc @wiz_io @microsoft @google @awscloud -F1, Corinthians, Pats fan! 🇧🇷➡️🇺🇸➡️🌎 opinions are my own
Joined May 2009
- Tweets 10,846
- Following 970
- Followers 649
- Likes 33,367
449 Photos and videos
Pinned Tweet
Jun 1
Most training content isn’t broken because people don’t care.
It’s broken because the pipeline is outdated.
Slides, LMS editors, scattered docs, duplicated decks, recorded sessions, and a ton of manual work from teams already stretched thin.
AI can create content faster, but that alone doesn’t fix the problem.
Training content needs a source of truth.
That’s the idea behind Fluxxo: AI-assisted authoring, course-as-code, and flexible delivery so teams can create training faster and actually maintain it over time.
That’s why I built Fluxxo.
Want to try it? DM me or join the waitlist. No spam. I hate cold calls as much as you do :)
fluxxo.ai
10
79,920
Software engineers solved the 'write once, deploy everywhere' problem 20 years ago. L&D is still rebuilding the same course four times.
Here is the concept they used, translated for course authoring:
• First: separate content from presentation. Your course structure lives in one file. Your output formats are just renderers.
• Second: version control everything. Branches for SME review, commits for each draft, rollback when a stakeholder changes their mind.
• Third: automate the formatting layer. The human writes the logic and pedagogy. The system handles slides, SCORM packaging, and voiceover.
This is not a futuristic workflow. It exists right now. Most L&D teams just have not made the jump yet.
2
2
16
I had not used ChatGPT (only Codex) in the past 3 months. Is it me, or is the chat dumber than before?
It behaves nothing like Codex, even if using the same model.
1
2
27
Jun 13
𝗗𝗲𝗮𝗿 𝗳𝗿𝗼𝗻𝘁𝗶𝗲𝗿 𝗹𝗮𝗯𝘀, 𝘀𝘁𝗼𝗽 𝘀𝗲𝗹𝗹𝗶𝗻𝗴 𝗹𝗼𝗻𝗴𝗲𝗿 𝗰𝗼𝗻𝘁𝗲𝘅𝘁. 𝗦𝗲𝗹𝗹 𝗴𝗼𝘃𝗲𝗿𝗻𝗲𝗱 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝗺𝗲𝗺𝗼𝗿𝘆.
Frontier AI needs an enterprise “memory/adaptation layer,” not just longer context windows or short-lived prompt caching.
The model should let companies ingest large bodies of knowledge once, convert that into a governed reusable memory/adapter, and attach it to future calls. Not full fine-tuning or RL. More like persistent, versioned, enterprise memory.
Charge for ingest, storage, refreshes, and usage. Give customers controls for source lineage, permissions, retention, rollback, audit logs, and evals.
It’s basically the cloud shared responsibility model applied to AI memory: the provider secures the model and memory infrastructure; the customer governs what knowledge gets attached.
This would be far more valuable than repeatedly charging massive token costs for the same context over and over.
1
2
39
Raph Soeiro retweeted
Jun 4
Will OpenAI reset Codex quota soon?
Inspired by @theo's prediction this week, I built a Super Scientific Methodology™ to predict when @thsottiaux will press the quota reset button next.
Prediction signals:
🔹OpenAI team vague-posting
🔹API/Codex downtime/issues
🔹Potential feature or product release soon
🔹OpenAI parties or events soon
willcodexquotareset.com
51
12
292
112,314
May 26
Hey @Google, have you thought about this? If I type google.com.br, I want to access the Brazilian site, NOT BE REDIRECTED to the US site.
93
May 25
I spent weeks trying to make fluxxo.ai more “agentic” by adding more agents.
More parallel tasks, more concurrency, more little AI workers running around. I honestly thought I was being clever.
But the output still wasn’t good enough.
Then I added one boring review pass and it got better. I added another intentional pass after that, and suddenly it clicked.
The answer had been sitting in front of me the whole time.
I didn’t need a swarm. I needed a sequence.
create → critique → improve → validate
I had confused “agentic” with “parallel.”
Turns out sometimes the “breakthrough” isn’t a bigger architecture. It’s admitting you were solving the wrong problem.
1
88
May 23
I’ve been reading lot about that we need to "use the right LLM for the right job" and honestly?
That's corporate astrology. 🔮
Here's the thing nobody wants to say out loud:
We've collectively decided that switching between 5 different LLM APIs is somehow a “feature” and not just technical debt wrapped in a buzzword bow.
"Oh GPT-5 is better for this, Claude for that, Gemini for changing its product strategy every 6 months…"
STOP. 🛑
You know what's actually better for the job?
The one you already integrated. The one your team knows. The one your budget can actually afford.
But that doesn't sound like strategic decision-making, does it?
It sounds like... picking the cheapest option. Which is literally the most strategic decision ever, but we had to pretend it was about some mystical "optimal allocation framework." ✨
The emperor has new clothes and they're a multi-modal transformer. 👑
Reality check: 90% of "LLM selection criteria" is just post-hoc rationalization of whatever was already in your tech stack.
The other 10%? Actual benchmarks that will be obsolete in 3 months.
So here's my unpopular take: Pick ONE. Make it work. Stop optimizing the optimizer.
Your delivery timeline would thank you.
What am I missing? Genuinely curious if you've actually seen "the right tool for the right job" philosophy pay off in production. 🤔
1
3
97
May 22
“I still hop into the IDE for some of my work”
“The future of coding is agentic..”
Read what you are writing. The Google arrogance is off the charts. How about listening to your audience?
This gives me the mid-2010s “you need to use GAE or GTFO” vibe while the world was still getting a grasp of VMs in the cloud.
Unbelievable….
May 22
Antigravity IDE is just now a separate app from Antigravity 2.0!
I spend most of my time in 2.0 personally and use it to review diffs view files. I still hop into the IDE for some of my work...or if I'm feeling nostalgic for autocomplete / want to handwrite some artisan code line by line 🙂
The future of coding is agentic and we wanted to give users the option to use either or both the agent-first 2.0 experience or the traditional IDE experience. Same agent, same models, different form factors.
More context: for existing AGY users, Antigravity IDE was merely renamed and is still installed on your machine. If you're a new 2.0 user and want the IDE, you download it from our website separately: antigravity.google/download
1
1
110
May 22
AI providers need to drop the 5-hour usage caps and offer a "Whole Weekend" flat-rate window. Just let me burn tokens for 52 hours straight from Friday night to Monday morning.
@sama 👀
1
40
May 21
Can @AnthropicAI and @OpenAI come to an agreement on how to display usage?
Consumed vs remaining fries my brain every time I need to check what's going on!
1
34
May 21
Hey @AnthropicAI, why is @claudeai refusing to keep working even with usage credits turned on??? Is this tactic to get more money out of me?
34