3
2
3
732
Digital Marketing Consultant, SEO, AI, Photography
Joined May 2009
- Posts 8,060
- Following 846
- Followers 514
- Likes 26,893
1,725 Photos and videos
World Labs and @drfeifei have come up with Marble a world model that combines Gaussian splats for 3D rendering, simulators and planners in a loop.
The Substack essay (“A Functional Taxonomy of World Models”) is one of the clearest articulations I’ve seen in this space. It cuts through the hype around “world models” by grounding everything in functional roles within the perception-action loop, rather than getting lost in architectural details.
Standout Points from the Essay
It plainly lays out the three functional categories and why they matter:
Renderers focus on visual output (what the world looks like — pixels and observations). These are the most advanced today but often lack structural integrity.
Simulators (which World Labs calls the “linchpin”) produce faithful state representations — geometry, physics, dynamics, and structure. This is where real utility for interaction and computation lives.
Planners output actions and decisions, closing the loop for agency.
The essay emphasizes that these aren’t separate silos but parts of an interconnected POMDP-style loop, with simulation as the critical bridge.
A strong simulator can feed both rendering (for humans) and planning (for agents). The long-term vision is a unified foundation model that fluidly handles all three, switching modalities as needed.
They position Marble as their concrete first step into simulator territory: it takes multimodal prompts and generates explorable 3D worlds with both visual outputs (Gaussian splats) and physics-ready elements (collision meshes). This directly moves beyond pure visual generation toward interactive, simulation-capable environments.
open.substack.com/pub/drfeif…
39
In the words of Don Draper:
okay the comments asked for the reverse so here it is. "pov: explain 1994 to a kid from 2026"
so here it is. dial-up, mixtapes, and the strange luxury of being unreachable.
fable 5 made the whole thing. no humans involved.
the future was supposed to make us cry. didn't expect it to be about the past.
1
42
Midjourney flips the script on Hollywood!
The image generator wants the plaintiffs to reveal in discovery their use of AI trained on copyright material.
Jul 3
Uno Reverse! From the Midjourney motion:
'If Plaintiffs are developing image-generating AI models — trained on unlicensed, third-party copyrighted data — for internal use in storyboarding or ideating content for film or TV, that evidence would equally demonstrate that it is an industry custom, even among the studios themselves, to download and train AI on unlicensed copyrighted content.'
1
52
How to cast and direct AI actors for more lifelike performances?
Here’s an excellent workflow tutorial that will really elevate your work with Seedance 2.0.
Character sheets are great, but they only solve one part of the problem.
They show you how your AI character looks.
They do not show you how they perform.
So I made a simple workflow and a system prompt to audition my AI characters before taking them into video generation, to test their voice, emotions, expressions, and screen presence.
Here is my workflow:
50
Mark G reposted
Higgsfield has just wiped out all the competition!!
Using its supercomputer capabilities, one creator has developed an app that mimics Cinema 4D.
Watch the video to see exactly how it works.
The cool thing is that you can set any camera movement (any!) and recreate any actions. Then run it through the Seedance 2.0 renderer, and your video – with maximum control – is ready. This is the very same functionality that’s usually built using the Unreal Engine in professional productions. And here, it’s available for the price of a subscription. The app is available here
higgsfield.ai/supercomputer/….
Just a heads-up: this is only suitable for pros. It’s just cheaper)
6
9
53
5,308
Apparently Claude can track users' country of origin and if it's China you will be blocked. Anthropic maintains a list of IP addresses associated with Chinese Labs which are also banned. Doubtless the users have bent the rules so now Alibaba has demanded that they delete Claude.
Per The Information e.customeriomail.com/e/c/eyJ…
80
OpenAI is considering letting the US government take a 5% stake in the company and has also had talks with Bernie Sanders who has advocated 50% public ownership.
Per The Information e.customeriomail.com/e/c/eyJ…
25
My earlier comment on S-curves.
Some growth curves look like exponentials from early to the middle of the timeline, (where we are now with AI), but as the tools get commoditised as that timeline matures, the curve reveals itself as an ‘S’ curve.
Way beneath the top edge of the ‘S’ curve a paradigm shift slowly begins a new curve, initially flat but trending as what appears to be an exponential. In the classic A16z slide below the mobile platform growth curve is shown as an ‘S’ curve with a new Augmented Reality (AR) growth curve beginning below it.
Now we could rename the AR curve as the Agentic AI World Model curve, a virtual physics space that can model the real world for agents plus a visual reasoning engine and a language model to interact with humans and computers.
The old Mobile platform ‘S’ curve we might rename as the LLM curve.
The tweet by @ylecun below expresses a similar idea from a European perspective.
20
This work comes from biology but also applies to AI cognition. (Paper linked in replies)
“The computation isn't in any individual neuron; it's in the collective dynamics of the network itself.”
“Adaptation, recurrence and network geometry appear sufficient to produce coherent global modes from locally irregular elements. That suggests cognition is increasingly being understood as a property of an evolving dynamical field rather than a collection of independent computational units.”
Weakly coupled nodes will self-organise into phase locked coherence over time, like a standing wave.
Sufficient entropy or perturbation will nudge the resonant nodes into a new stable basin. Like traversing a landscape.
This is a fascinating result because it continues a trend that's appearing across neuroscience, physics and complex systems.
The interesting point isn't simply that rich oscillatory behaviour emerges from simple rules, it's where that behaviour resides. The computation isn't in any individual neuron; it's in the collective dynamics of the network itself.
Adaptation, recurrence and network geometry appear sufficient to produce coherent global modes from locally irregular elements. That suggests cognition is increasingly being understood as a property of an evolving dynamical field rather than a collection of independent computational units.
My prediction is that the next generation of brain models will move beyond static neural networks toward adaptive coupled-oscillator frameworks, something closer to generalized Kuramoto or mean-field dynamics, where synchronisation, changing coupling strengths, memory and network topology continuously reshape one another.
If that direction continues, the remaining challenge won't simply be reproducing neural firing patterns, but explaining how persistent identity, causal memory and stable reasoning emerge from those coherent field dynamics.
Every paper like this seems to reinforce the same message: intelligence may be less about increasingly complex neurons, and more about the mathematics governing collective organisation.
3
176
In this essay @chamath is saying much more clearly than what I tried to say in my earlier tweet: that the growth curve of intelligence will take over from the growth curve of information as the cost of intelligence comes down, The Great Descent.
1
44
A virtual physics world model, with symbolic modelling seems to be where the puck is going.
Many trajectories are pointing in that direction, including Marble from @theworldlabs and JEPA from @ylecun Isaac Sim from @NVIDIARobotics and many independent devs. A common vocabulary is yet to emerge but I expect it will arrive shortly.
Jul 2
Eventually, much of AI will converge towards intuition-guided symbolic world modeling, i.e. deep learning-guided program synthesis. It is inevitable. Symbolic modeling lets a system construct a compact, reusable, highly generalizable mental model of a problem space using minimal data.
1
6
424
Some growth curves look like exponentials from early to the middle of the timeline, (where we are now with AI), but as the tools get commoditised as that timeline matures, the curve reveals itself as an ‘S’ curve.
Way beneath the top edge of the ‘S’ curve a paradigm shift slowly begins a new curve, initially flat but trending as what appears to be an exponential. In the classic A16z slide below the mobile platform growth curve is shown as an ‘S’ curve with a new Augmented Reality (AR) growth curve beginning below it.
Now we could rename the AR curve as the Agentic AI World Model curve, a virtual physics space that can model the real world for agents plus a visual reasoning engine and a language model to interact with humans and computers.
The old Mobile platform ‘S’ curve we might rename as the LLM curve.
The tweet by @ylecun below expresses a similar idea from a European perspective.
Jul 3
Exactly. I've been disseminating a similar message for years.
The concentration of power in AI and the desire for control is by far the biggest danger of AI. It could lead to a few private companies and/or countries being in control of access to information, access to knowledge, and access to the tools of economic expansion.
It's a kind of medieval obscurantism akin to the Ottoman empire banning the use of the printing press for 200 years, in part to keep control of the dogma, but also to protect the corporation of the calligraphers and scribes.
Relevant historical bits about the Internet:
1. It took a deliberate decision by Al Gore and Bill Clinton to open up access of what was then ARPAnet to commercial entities and to the public, against the desires of the entrenched telecom industry. During a public roundtable about the "information superhighway" in 1993, the CEO of AT&T told Gore and Clinton "leave it to us". Gore said no.
2. In the late 1980s, setting up an Internet presence required buying proprietary hardware with proprietary OS and software stack from Sun Microsystems, HP, IBM, or Dell. By the 2000s, all of this was wiped out by commodity hardware, Linux, Apache, and an entirely free/open software stack. This migration to open platforms was the result of market forces.
Infrastructure wants to be open.
Foundation models are becoming an infrastructure and will inevitably become commoditized.
Long term, the money is in the application layer, which is what I, Arthur Mensch, Alex Karp, and others have been saying.
87
Fable is pure magic.
I wanted a beautiful app to explore ocean wildlife. Fable built this in an hour. It generated videos with Seedance and carefully synchronized them to make these absolutely insane transitions. I've never seen anything like this in an app. Unreal.
108
158
2,998
263,514
Anthropic is perennially either losing users' trust or scrambling to regain users' trust.
User fatigue is reaching breaking point. Stop using compute capacity as an excuse. Buy some from Meta or buy more from SpaceX. Can't pay for it? Ask NVIDIA.
BTW - get a new CEO.
I've heard a lot of questions about Fable's availability on subscription plans.
While it will come off subscriptions after July 7th, we aim to restore Fable as a standard part of our subscriptions as soon as capacity allows, as we mentioned in our original blog post.
1
56
Don't prompt Fable like other Claudes. That's the blunt message in Anthropic's Prompting Guide for Fable (linked in replies).
Jul 2
Fable 5 performs worse when you over-prompt it. Anthropic put that in writing.
the official prompting guide is blunt. prompts built for older Claude models don't improve Fable 5 output. they degrade it.
the highlights tell the same story. a single brevity instruction replaces entire prompt frameworks. their recommended memory system is a markdown file, not a RAG pipeline. lower effort on Fable 5 outperforms max effort on the previous model.
the scaffolding people built to compensate for weaker models is the thing getting in the way.
full guide linked below. if you build with AI, read it.
72
Mark G reposted
Jul 2
while working on my threejs video pipeline, fable randomly started trying to make his own video, so i stopped and ask him what he wanted to make and let him finish and gave him some quality advice and he made this! he's commenting on the fact that he'll only be around on claude subs for a bit this month
75
119
1,627
268,309
So this is how Sam was greasing wheels (or palms)...
x.com/marksg/status/20722971…
To get to the public release of Fable 5 the biggest factor was replacing the "weirdo" with the other co-founder.
How do we know this?
Because Anthropic still hasn't solved the problem of jailbreaks, (it would be worldwide breaking news if they had), they only strengthened the classifier that prohibits the one specific case that was raised by the govt.
Now they're inviting members of project Glasswing to help formulate a common set of cyber-risk definitions and jailbreaks. So that means that the USG was prepared to let them release to the public as long as they were willing to report any misuse and agree to cooperate with NSA/CAISI/Commerce.
This is the beginning of a new, semi-formal regime of government regulation. How long before it gets formalized? We'll have to wait and see how they treat GPT 5.6. I expect Sam is greasing wheels as we speak.
21
If Sam Altman knows one thing, it's that the Trump family, Jared Kushner in particular, loves sovereign wealth funds.
@sama is floating the idea of a 5% ($42.8B) government stake in OpenAI like their 10% stake in Intel, as a sweetener to any "deal". And so the regulator becomes the owner...
No wonder, in a CNBC interview, Trump stated that while some AI guardrails are necessary, the government should do "as little as possible" to avoid hampering American companies competing with China. Per Bloomberg.
A nice breakdown in the Kim "Chubby's" newsletter below.
getsuperintel.site/p/openai-…
2
66
Are LLMs a dead end? This is Claude Fable controlling a house demolition. Grounding intelligence in the physical environment is one of the biggest remaining challenges on the road to AGI. I wonder what @ylecun would say?
Jul 2
Here's Fable today! Fable managed to wreck the last remaining portion of the awning structure. It required my help though - And Fable had difficulty positioning the arm.
51
Mark G reposted
Jul 2
Feels like an end of era, ordinary people will probably never again get upgraded frontier models.
Fable 5’s return shows how safety routing can downgrade a frontier model.
Now we only permissioned intelligence.
The cost of putting a gatekeeper inside intelligence.
To note, that safeguard is not a simple refusal layer; it is a classifier that sends flagged Fable 5 requests to Opus 4.8.
Fable 5 came back, but the old promise did not.
End of an era. ☹️
Jul 2
FABLE 5 CAME BACK NERFED.
We re-ran the July 1st version of Claude Fable 5 on BridgeBench.
The results are brutal:
Debugging: 86.2 → 25.9
Refactoring: 73.6 → 38.4
Hallucination: 75.9 → 61.7
The new guardrails are kicking in on way too many tasks and falling back to Opus 4.8.
This is not the model that got banned.
Anthropic owes everyone an explanation.
36
38
241
27,885