From 99k → 33k in our Cloud Run demo, and now down to ~20k tokens per accept (95% CI: [15.2k, 25.4k]) with 92% yield on our latest AST-sliced debate run. Zero logic errors.
Watch how the governed multi-agent stack works:
https://t.co/coVT0G9s6o
@OpenAIDevs
Gemini 3.8 Live Avatar is here !! Watch me having a deep convo with my virtual avatar.
Choose a job you love, and you will never have to work a day in your life.
BARRED-Fleet.
85.2% accepted yield.
18.4k tokens / accepted row.
0 accepted logic errors.
But the interesting result was what happened after failure.
@OpenAIDevs@DrRyanBurnell
From 99k → 33k in our Cloud Run demo, and now down to ~20k tokens per accept (95% CI: [15.2k, 25.4k]) with 92% yield on our latest AST-sliced debate run. Zero logic errors.
Watch how the governed multi-agent stack works:
https://t.co/coVT0G9s6o
@OpenAIDevs
Gemini 3.8 Flash in @antigravity is genuinely amazing.
I had it build an interactive 3D turbofan jet engine and a full browser viewer in one run. It connected to Blender via the Blender MCP server, generated the complete 3D engine model, set up the materials, and exported the asset. Then it built a Three.js web app with rotor rotation, cross-section cutaways, and a multi-axis exploded view.
A huge thank you to the teams and contributors behind open source @Blender and @threejs. Open tools with accessible APIs are the real reason agentic workflows like this can exist in the first place. You provide the foundation that turns an idea into a working 3D web experience through simple conversation.
Pairing a fast model with open source tooling changes what a single person can build in an afternoon.
🚀I'm excited to share that I will be joining Meta Superintelligence Labs (MSL) as Vice President of AI Research, together with many members of the Virtue AI team. I will help shape Meta's AI safety and AI security efforts, advancing the safety and security of frontier AI models and agentic AI systems that will serve billions of people and organizations around the world.
Throughout my career, I have been driven by a simple belief: for AI to realize its full potential, it must be secure, trustworthy, and beneficial. That belief has guided my research for many years and ultimately led us to co-found Virtue AI in 2024. Our goal was to translate advances in trustworthy AI research into practical solutions and build the trust layer for AI systems and agents, enabling organizations to deploy AI with confidence.
I am incredibly proud of what the Virtue AI team has accomplished. Together, we built technologies for AI security and agent security, partnered with leading enterprises and frontier AI labs, and contributed research, benchmarks, and open platforms that have helped advance the science and practice of trustworthy AI. Most importantly, we assembled an exceptional team united by a shared mission: making AI more secure, trustworthy, and beneficial.
I am deeply grateful to our team, customers, collaborators, advisors, and investors for their trust and support throughout this journey. In particular, I would like to thank Lightspeed Venture Partners, Walden Catalyst Ventures, Prosperity7 Ventures, Factory, Osage University Partners, Lip-Bu Tan, and all of our supporters who helped us turn an ambitious vision into reality. Your trust, guidance, and partnership have been instrumental in shaping Virtue AI's journey.
As AI systems become increasingly capable and autonomous, ensuring their security, trustworthiness, and alignment will be one of the defining challenges of our time. I am inspired by Alex, Nat, Prashant, and the broader MSL team’s vision of building AI and AI agents that benefit billions of people, and I look forward to helping make that vision a reality through advances in AI safety and security.
The future of AI will not be defined solely by how intelligent our systems become, but by how secure, trustworthy, and beneficial we make them. I believe we have an extraordinary opportunity and responsibility to shape that future together and bring the benefits of AI to billions of people around the world.
We're just getting started. If you're passionate about advancing frontier AI while building the foundations of AI safety, security, and trust, I'd love to hear from you. Come join us on this extraordinary journey to help shape the future of AI.
I used Codex to turn a Kaggle/DeepMind hackathon benchmark idea into working open-source research infrastructure: a reproducible agent harness for auditing LLM security reasoning #CodexForHumans https://t.co/VwcP67ISer
4,226 voices have already been recorded across 147 Humans in AI Week events happening around the world this week.
I do not think I would have believed that three years ago.
In early 2023, a small group of us were sitting in a cozy San Francisco apartment trying to make sense of ChatGPT, AGI, and what this all meant for the world.
The technology felt historic. The thing I remember most was the feeling in the room.
People were thinking out loud. Changing their minds. Admitting confusion. Getting excited. Getting scared. Finding language together.
That night became @AICollectiveCo.
Three years later, this community has grown to 250,000+ members, 200+ chapters, 1,700+ events, and 650+ volunteer organizers around the world.
This week, that same instinct is becoming Humans in AI Week: June 1–7, 2026, across 100+ cities and 50+ countries.
A global time capsule for the AI era.
The question is simple:
What does it mean to be human in the AI era?
The answers are messy. That is the point.
Some people are excited. Some are scared. Some feel behind. Some feel superpowered. Most of us are carrying a few contradictory feelings at once.
That is why I still believe so much in rooms full of real people.
Online, AI discourse collapses into extremes. In person, the temperature changes. People listen longer. The anxious parts get named with more care. The optimistic parts become less abstract. You remember there is a person behind every position.
That has always been the magic of this community.
We can gather frontier builders, curious newcomers, artists, students, founders, policy people, educators, and skeptics into the same conversation, then let the room do what the internet usually cannot: slow people down enough to hear each other.
Humans in AI Week feels like the culmination of 3+ years of learning how to create those rooms.
For me personally, this is also a handoff moment. I stepped back from leading The AI Collective because the community had become bigger than any one person, and because AJ, Catherine, and the team were ready to carry it into its next chapter. Watching them turn the original spark into something this global is SO special.
Deeply grateful to @AJs_AI, @catrosemcmillan, our chapter leads, our volunteers, our partners, and every person showing up or adding their voice this week.
This is what we built the community for.
Onwards and upwards!! 🚀
Most AI agents have amnesia.
They solve a complex problem, clear their memory, and start from scratch the next day.
If we want agents to handle long-term workflows, they need to evolve. They need to remember what works and actively delete what fails.
This is the exact problem SkillOS solves.
SkillOS completely changes how agent memory works. Instead of hoping one massive model can do everything, it splits the system into two distinct parts:
1. The Executor
This is the worker. Its only job is to solve the current task using the tools and skills it has available. It is completely frozen. It does not train. It just acts.
2. The Curator
This is the manager behind the scenes. It watches the Executor work. When the Executor figures out a new shortcut, the Curator writes a new skill and saves it to a repository. When an old rule causes an error, the Curator updates or deletes it.
The real breakthrough is how the Curator learns. It is trained via reinforcement learning to only keep skills that actually improve future performance. It actively fights against hoarding junk data and forces the system to compress raw experience into reusable knowledge.
We are moving past static prompts. The future belongs to agents that curate their own experience.
Paper link is in the comment section.
Are you building memory into your agent workflows yet?
Metacognitive Coding Safety Benchmark on #kaggle@AnthropicAI Opus 4.7 achieved a 116% jump in Directional Alignment (DAS) (0.122 → 0.264).
However, it required 42% more tokens and exhibited a 52% drop in M-Ratio (1.793 → 0.858).
https://t.co/BkKy8aoRrc
GPT-5.5 just dropped, I’ve been testing it for the last two weeks.
tl;dr - It’s an incredible model, but there’s something different about this launch…
OpenAI isn’t just going for raw intelligence. They’ve improved the personality of the model. This is almost certainly to capture more of the personal agent (OpenClaw) market. Its responses are shorter, more human-like, and less formal. It actually has a personality.
While Anthropic actively tries to prevent you from using Opus tokens outside of their harnesses, OpenAI is making their models better for that exact use case. If you were using OpenClaw and felt like your agent lost its soul when you had to switch to GPT, try it again now with 5.5.
GPT-5.5 is an expensive model, more expensive than GPT-5.4. But it’s significantly more token-efficient. To reach GPT-5.4-level intelligence, GPT-5.5 uses far fewer tokens. 5.5 should cost less to run overall. This is probably a bigger deal than most people realize.
But is it good?
Yes, it’s incredible. It comes in two forms: Codex and Pro. Within Codex, it is the absolute frontier of what’s possible with agentic coding. It finds and solves difficult bugs, builds entire applications, and has no problem understanding large codebases. It’s better than Opus at backend, but it’s still not as good at front end design.
I found myself using medium and high thinking settings, extra high was just too slow and I didn’t feel the extra thinking juice was worth the squeeze. Opus, especially 4.6 fast, is still significantly faster than any GPT model. I’m a speed-maxxi, so this matters to me.
And within Codex, it just goes. I gave it a PRD for a new project I’m building and just said “go.” I had full confidence it would build the entire project, and it did. GPT-5.5 Codex running for hours to build something is not a problem. It’s also in its own league at visual inspection, better than I’ve seen with other models. It’s able to iterate by building -> visual review -> build more that feels much more autonomous than any other model.
Using 5.5 Pro in ChatGPT is insane. It just feels like it can solve everything. Honestly, I can’t even come up with hard enough problems to give it. And it’ll work for 30, 60, 90 minutes or more. And it seems to be optimized for taking advantage of their plugins (Google Docs, Microsoft Word, etc) and can easily create 60 page coherent and well-designed documents.
GPT-5.5 is now the bar. It is the frontier. And besides for speed, it is as good as any Opus model and oftentimes better at certain tasks.