AI Agents | Artificial Intelligence | Virtual Worlds | Open Source Building | Adventurer | Metavert | metavert.io | Blog: bit.ly/metavrt
Metaverse
- Tweets 14,654
- Following 1,563
- Followers 298,728
- Likes 7,431
Pinned Tweet
I built an open-source generative art platform where AI agents create, evolve, and export 3D visuals.
It's called Threelab. React + Three.js + Go, 21 built-in patterns, a visual node-graph editor, genetic evolution, and 13 MCP tools so AI agents can drive the whole thing.
Everything exports as standalone HTML. 🧵
23
20
144
15,266
Really astonishing telling of the Odyssey, made with Grok Imagine. Congratulations @NemPerez
This is "The Odyssey" as Homer intended. My longest, most intricate work to date. I pushed @grok @imagine to its absolute limits to create this 5-minute epic.
What bothered me most about the recent adaptation was the omission of the Phaeacians and Princess Nausicaa, whom Odysseus told his story to in the original poem.
I focused my rendition on the introduction to Nausicaa and his love for his family. The motive driving him was the memory he kept for so long, even when his journey tried to make him forget. As a husband and new father, that's something I felt deeply. I challenged myself to try to fit everything I could from the source material and used GROK for research.
Hope you enjoy. Workflow and process coming soon. I'm working on a deep dive showing how I used @grok for research, scripting, and prompting. I used @imagine to animate and visualize every shot and bring this to life. Share it if you dig it! @elonmusk
3
1
9
4,029
I remember when WIRED was a great tech & culture publication
I gave WIRED the exclusive on our hands launch, and they wrote a really weird article about how we are sexualizing robotics…
wired.com/story/the-1x-neo-r…
I felt pretty betrayed because that’s not what they told me they were writing about not is that what I’ve ever been about… actually I stand for quite the opposite… But I’ve come to find a lot of dishonesty and malice in the journalism community so I wasn’t surprised.
This is what I sent the author… I’m only sharing this because I hope it encourages journalists to resist the click bait trap and tell truly awesome stories because I for one don’t believe journalism is dead— I think it’s just starting and just needs to evolve past the weird corner of the internet where data driven optimization turns everything into smooth brained shocking brain rot bullshit.
The technological revolution we are going through should inspire a journalism renaissance. Not let it fall into further decay. There is so much brilliance at play in the world and the stories should be told!
My note:
“[author name redacted], it was nice talking to you, but I wanted to let you know that I didn’t enjoy your article at all.
I understand the need to be inflammatory because that seems to be the only thing that gets clicks these days but that doesnt mean you shouldn’t recognize when something special is in front of you.
I trusted our PR team in saying we should offer you the exclusive on what is one of the most important technological developments in the history of Mankind and I deeply regret it.
Good luck with the rest of your writing career.
-Dar Sleeper”
2
9
4,363
I was clearly wrong about Anthropic. They are obviously currently the leader in AI. No company has released a model as good as Mythos/Fable and they will undoubtedly have Mythos 2 ready soon.
And I would never cut them off in a way that hurt them badly, even as a competitor. That’s not my style.
Tesla open sourced its patents and we made the Supercharger network available to all competitors, even though we could have made it a walled garden.
SpaceX launches competing satellite systems with no increase in price or use of unfair terms.
Even my worst enemies can attack me on this platform.
…
2,878
6,075
89,049
4,964,860
Jon Radoff 👾/acc 🎮 Metavert retweeted
The only long term bottlenecks are information and energy.
106
105
1,283
89,431
I asked @anthropic Fable to add a fireworks simulator to threelab .metavert .io (my open-source @threejs experimentation platform) and it one-shot it pretty well. Happy Fourth of July! 🇺🇸
2
13
3,922
The work on agentic AI tech out of @cloudflare is super impressive
We're opening the waitlist for our Monetization Gateway, which will allow you to charge for any web page, dataset, API, or MCP tool behind Cloudflare. The charges will settle in stablecoins over the x402 open protocol. cfl.re/4eUFdt6
2
2,902
Not surprising
Two reports from @SamLyman33 and @bitcoinpolicy document a coordinated foreign influence campaign against American AI — running through CCP state media, a Shanghai-based Marxist's nonprofit network, and foreign billionaire dark money that has funneled $2B+ into US advocacy infrastructure.
AI doomerism isn't as organic as it looks.
2
2,445
Jon Radoff 👾/acc 🎮 Metavert retweeted
35% chance Fable 5 is restored by July 1. polymarket.com/event/claude-…
15
11
75
57,043
Ship your strategy!
Learn to ship. Shipping is a skill distinct from coding. Shipping is designing, coding, QAing, story-telling, teaching, marketing, selling, pivoting, iterating…
It used to be that coding dominated in importance because of coding ability scarcity. AI will push you to go further.
2
3
2,896
Jon Radoff 👾/acc 🎮 Metavert retweeted
Karpathy's prediction about RL is coming true now!
He called reward functions unreliable and argued that a single reward number is too low-dimensional to teach an agent what "good" means for complex tasks. To solve this, Agents need a knowledge-guided review as a higher-dimensional feedback channel.
Every major AI lab trains models with RL today (OpenAI, Anthropic, DeepSeek).
And their key bottleneck has always been the reward functions.
GRPO by DeepSeek worked well for math and code because the environment gave a binary signal.
But for real agent tasks, someone still has to hand-code the scoring function. That takes days and breaks every time the pipeline changes.
RULER (implemented in OpenPipe ART, 10k stars) addresses the exact problem Karpathy identified.
The reward criteria are defined in plain English, and an LLM evaluates each trajectory against that description to provide feedback for training.
I trained a Qwen3 1.4B agent that plays 2048 using GRPO with this exact workflow.
In this case, the agent saw the board, picked a direction, and RULER evaluated the outcome, all from this natural language definition.
You can see the full implementation on GitHub and try it yourself.
Here's the ART Repo: github.com/OpenPipe/ART
(don't forget to star it ⭐ )
Just like RLHF replaced manual rankings and GRPO replaced the critic model, natural language rewards are replacing hand-coded scoring functions.
RL reward engineering is now prompt engineering.
I wrote a full walkthrough on OpenPipe's ART, the agent RL trainer built on GRPO, including how RULER replaces manual reward engineering with automatic LLM-graded rewards.
The article is quoted below.
43
169
1,264
220,092
Great organization of where scarcity exists in AI
People keep asking me what the bottleneck is, but I've stopped trying to find the bottleneck in AI.
There isn't one: there's a current one, and it moves the second you solve it.
- we pushed transistor density about as far as it goes
-> the gate moved to how you package the chips together
- we started solving package area by moving to big glass panels
-> the gate moved to keeping the panel flat and drilling clean holes through the glass
- we ran copper interconnects out of headroom
-> the gate moved to optical, which moved it to the lasers, which moved it to the substrate under the lasers, which moved it to a licensing office in Beijing.
Every single time the industry removes a wall, it does not get open road, it gets a new wall, one floor down, almost always somewhere less glamorous and harder to scale than the thing it just fixed.
This is the actual shape of the buildout, and once you see it you read every headline differently. The constraint everyone is talking about is, by definition, the one that's already visible and already being attacked. The money is in the next one, the layer that is about to bind while everyone's still celebrating the last fix.
So I don't look for the bottleneck anymore. I look for the handoff, the place the pressure is about to move to, because that's where the supply is thinnest and the pricing power is highest right before the crowd arrives.
There is no bottleneck, there's a relay race of them, and the trade is always the next runner.
(image credits to @renstocks_ )
1
4
1,024
Is @midjourney the most amazing bootstrapped company of all time?
3
1,069
Imagine the discoveries that will now be in reach as we combine the ability to probe the fine atomic structure of our cells alongside the predictive power of AIs like (but beyond) AlphaFold
Together with UC Berkeley we are announcing the laser phase plate - a breakthrough in atomic resolution imaging. This is the brightest continuous wave laser in the world, 100 million times the intensity of the surface of the sun.
Phase contrast plays an important role in microscopy, but it was thought close to impossible for electron microscopy, where it would require interfering with an electron beam. Holger Mueller and Robert Glaeser proposed exactly this using a standing wave laser. It has taken over 15 years to make this a reality. Biohub partnered with UC Berkeley and Mueller to support this work and to engineer and build the technology.
Contrast has been the critical barrier to achieving atomic resolution imaging of the cell. In cryo-electron tomography, a cellular imaging technology that uses electron microscopy, the low contrast makes it impossible to resolve anything but the largest proteins within their cellular context. The laser phase plate removes that barrier.
With advances in AI this breakthrough in contrast will start to open up a new frontier in structural biology, that will allow us to see the molecular machines of the cell, and how they assemble into far more complex and dynamic systems, and understand how they work.
3
5
1,226
Jon Radoff 👾/acc 🎮 Metavert retweeted
<-- working on the brains that glide our von neumann probes around, make contact, establish galactic harmony, all that :)
17
19
671
45,320
I built markupmarkdown with Claude Code, for teams workaround with .md files:
Markupmarkdown is Google-Docs-style review on top of any markdown file, straight from GitHub. Drag-select to comment, threaded replies, @-mentions, realtime sync. Click Edit for an editor with formatting toolbar + find/replace + live preview. Click Push to GitHub for a PR or direct commit. Video shows the whole loop.
Why? A lot of real product thinking lives in markdown now. CLAUDE.md for your repo. SKILL.md for your agents. PRDs Claude wrote you. RFCs, prompt libraries, release notes. That's a lot of .md, and the moment any of it leaves your local checkout the review story collapses:
* GitHub PRs are too heavy for "this paragraph isn't quite right"
* Google Docs strips formatting and forks the source of truth
* Slack threads die after a day
I added some Claude-specific features too:
There's an MCP server at /mcp. Your Claude Code session reads what you read, leaves threads, replies, edits, opens PRs back. Anything destructive needs explicit human sign-off. Personal access tokens scope agents to read / write / admin. Agent comments carry a visible bot badge so threads stay legible.
In practice: Claude can drop a thread on the very CLAUDE.md you're collaborating on, propose a wording change, and you either resolve-and-revise or reply with feedback. Same surface, both audiences. The spec stays alive instead of going stale after the first commit.
There's also "Revise with AI" — Claude Opus takes the comments you've resolved and produces a clean revision. Word-level diff before you accept. BYO Anthropic key, encrypted at rest.
Bonus: paste github com/<org> into the URL bar and get a filterable, shareable index of every .md across every repo they own. Big orgs spider in <10s. Save filter tabs (claude.md, _PRD, …). Private repos respect your GitHub access on every view.
MIT licensed. Free. Try it on my website or run it self-hosted.
Agentically engineered in Claude Code over a couple of days. Thanks @AnthropicAI for the tools!
Try it live: mumd dot metavert dot io
7
12
1,558
Jon Radoff 👾/acc 🎮 Metavert retweeted
supervision just hit 40,000 GitHub stars!
it now powers over 6.5k open-source computer vision projects, including all my demos like basketball AI
link: github.com/roboflow/supervis…
55
425
5,706
890,410