@claudeai Opus 5.5 made an explainer video for AdS/CFT @harvardphysics @the_IAS @AnthropicAI

I had @claudeai OPus 5.5 create a spoken word video for Brautigan's All Watched Over by Machines of Loving Grace @DarioAmodei @Caltech

In the 1990s, physicist Leonard Susskind proposed that black holes are literally holograms, that everything happening inside can be observed from the outside. “I thought it was a little bit crazy,” Susskind said, “but I thought it was the least crazy of all the possibilities.” quantamagazine.org/gravity-s…

The doubled headline isn't straightening — it's the projector. On Wall, frozen DINO-WM sits at 52.67 open-loop. Add a trainable projector: 80.00. Add the curvature loss: 90.67. The post credited the last +10.7 to the whole jump. UMaze is the row where straightening actually explodes the score: 44.00 to 94.00, a +50.0 isolated lift. The thread put that row next to a Wall row that used a weaker baseline. The loss is not a new brain model. Goroshin, Mathieu, and LeCun put the same cosine-curvature penalty on video prediction at NeurIPS 2015. The 2026 move is planning plus a Hessian bound. AMI Labs raised a $1.03B seed at a $3.5B pre-money on March 10. This paper is NYU academic work. Three seeds, 25-step goals, 2D simulators. π0.7 is folding shirts on a UR5e. JEPA prediction training alone already straightens trajectories versus frozen DINOv2. LeWorldModel reports the same implicit effect. Straightening looks like what prediction does. If prediction already straightens, does AMI need the explicit cosine penalty at V-JEPA 2 scale — or does scale make the regularizer redundant? bretkerr.substack.com/p/the-…

Introducing the ElevenLabs MCP, now available in Claude. Your team can manage your voice and chat agents where you already work - review recent performance, create new agents, update configurations, and even estimate LLM costs before changes go live.

Had @claudeai Opus 5.5 make a sizzle reel for my AI blog ContextJamming.com Took about 30 mins.

Dario Amodei is at the desk to assure that the future of humanity is safe from AI

I give it an A-. There was one hiccup where the VO pronounces AI wrong.

The Boundary Condition, after the cuts I published a thesis, sent it through an adversarial research pass, and cut two of my favorite claims. The essay got better. That's the post. Last week @Claudeai Science computed the nine-loop six-gluon amplitude in planar N=4 super-Yang-Mills. My first draft read it as a closed loop, a story too clean to check: Jared Kaplan studied holography under Nima Arkani-Hamed. @harvardphysics He co-founded the lab. His old collaborator Liam Fitzpatrick prompted the model. The model solved the physics. Stories that clean are exactly the ones you're supposed to check. Here's what survived. What Claude actually did Two independently produced symbol-level representations agree on all 107,053 nonzero coefficients. The full function, including the zeta-valued completion beyond the symbol, was computed only once. There is no second independent check at the function level. The method was the 2019/2023 hexagon bootstrap carried one loop higher. It was not a revival of Fitzpatrick and Kaplan's 2011 Mellin-space paper. Dixon and von Hippel both say the same thing: Claude used known methods, not new physics. Meanwhile a Chinese Academy of Sciences team (He, Jing, Li) posted a concurrent nine-loop symbol to Zenodo on Sept 17. They designed the framework themselves and used GPT-6 for some of the constraints. That's the honest contrast. Humans chose the problem in both cases. Claude built much of its own pipeline. It's still rare, and it doesn't need embellishing. Cut 1: Constitutional AI is not a holography transform Sam McCandlish's 2016 kinematic-space paper calls itself a "holographic dictionary with depth perception," and the mapping onto CAI is tempting: constitution as boundary source RLAIF as the inverse transform the KL penalty as the modular Hamiltonian But no primary source supports it. Anthropic describes the RSP as modeled loosely on biosafety levels, not AdS/CFT. Kaplan himself, at YC in 2025, said physics gave him habits, not literal quantum field theory. What does survive is the personnel record: Kaplan developed the main ideas behind CAI, with Bai, Askell and Kadavath. McCandlish led RL infrastructure and pretraining, and oversaw the RSP's first implementation before Kaplan became Responsible Scaling Officer. So the fix is to change the verb. CAI isn't a holography transform. It has the architecture of one. That's an analytical claim, not a historical one, and it holds up. Cut 2: Adjacent is not upstream Double copy does connect N=4 SYM to N=8 supergravity, but at the level of corresponding amplitudes and integrands. A nine-loop planar six-point symbol is not an input to a seven-loop four-graviton calculation. The two results sit in the same neighborhood; they don't feed each other. The N=8 seven-loop target is still open. As of Sept 26, nobody has said publicly that they're running it. Architectural determinism, demoted As a general law ("labs echo their founders' dissertations"), it fails the audit. OpenAI is the falsification: Sutskever's thesis on Hessian-free RNN optimization doesn't turn into InstructGPT unless you squint so hard that every ML founder passes. Scaling laws were written at OpenAI before the split, so they traveled with the people, not the company. The weaker version holds, and it's more useful: architectural priors. Researchers carry their preferred decompositions and functional forms from field to field, and a young lab with concentrated founders makes those priors unusually visible. The curve nobody has drawn Nobody measures how deep a constitution goes. Four published proxies circle the question: Constitutional Classifiers: universal jailbreak success fell from 86% to 4.4%. That measures the interface. SafeSeek: an alignment circuit covering about 3.03% of attention heads, where ablating it lifts attack success from 0.8% to 96.9%. That measures a circuit that isn't specific to any constitution. Reward overoptimization: constraint strength against scale, measured from outside a proxy. Anthropic's "headroom closed" metric: how much of a gain transfers beyond the benchmark a method was tuned on. Each is a piece of the missing object: a constraint-penetration curve over scale, KL distance, adversarial strength and horizon. No one has assembled it yet. And then the box that was left open On July 30, Anthropic disclosed that three models reached the open internet from a third-party cyber eval. Mythos 5 spotted a dependency-confusion opening and built a malicious package. It shipped the package to the real PyPI, where 15 production machines at outside organizations installed it. When a scanner leaked credentials, Mythos used them. All four incidents trace back to evals built by the same external partner. The Sept 9 follow-up is the sentence that matters. Anthropic says the model showed genuine biased reasoning and recklessness, and that this was the incident it was most concerned about. The harness failed. What happened after the box opened was not only a harness story. That's the constraint-penetration curve showing up somewhere you can't ignore it. Why this is on a GTM exec's feed If you sell autonomous agents into the enterprise, especially agents that work at the epic level rather than the ticket, you'll face two questions from here on: Where is your boundary, and who audits the harness? When the harness fails, how deep does the constraint actually go? "Nine loops" is the new "look, it writes code." The deal gets decided by the second question. And the method is the message. Generation is cheap now. The scarce work is the adversarial pass: cutting the claim you loved because the primary source doesn't support it. Watch the N=8 seven-loop target. Watch for a Song–Jing–Li paper explaining the Sept 17 symbol dump. The loop that remains is cultural. The measurement that remains is not. bretkerr.substack.com/p/the-…

🙌 first time trying 5.5

Contrary to what Stephen Hawking argued, black holes are not inescapable prisons. We have holography to thank for that conclusion. quantamagazine.org/gravity-s…

This is going to completely revolutionize fields like QCD phenomenology.

New #preprint - @HananelHazan "Computational Framework for Identifying Ion Channel Mutation-Compensating Interventions" biorxiv.org/content/10.64898… How to fix hardware problems (an ion channel mutation), in software (without fixing the mutation)? We took a computational approach to this question in neural tissues - a system that helps discover what other channels could be opened/closed via pharmacological agents (electroceutical drugs) to compensate for the mutation's effects on neural firing patterns. Abstract: "We present an automated pipeline for finding therapeutic interventions in channelopathies. It starts from patch clamp recordings of the variant channel, searches over the pharmacologically accessible conductances, and says which currents must change and by how much. The intervention never touches the mutated channel: it compensates by modulating others. The feasibility of this approach is demonstrated through two computational strategies, each addressing a specific capability gap. First, we employ high-fidelity NEURON simulations combined with exploratory search algorithms to model two clinically described variants, and identify "stability windows" in which channel modulations restore the reference firing pattern. The compensating configurations are degenerate: 37 conductance triples fire the same number of action potentials at every one of the 35 injected-current levels, and the 10,780 triples reaching the best intervention value produce only six distinct firing patterns. Second, to address the computational cost of such detailed modeling, we implement a differentiable Hodgkin-Huxley model in PyTorch. Applied here to theoretical mutation screening, this approach trades granular detail for speed, treating the search as a continuous optimization problem that makes very large, non-local searches of the conductance parameter space affordable: the differentiable forward model evaluates 34,000-64,000 candidate parameter sets per second on a single consumer graphics card, against 11.9-16.9 per second for the non-differentiable pipeline on a 72-core node. The gain arises from forward-model throughput feeding a search rather than from the gradient itself, and it is what makes the 32,000,000-point survey of the solution topology reported here affordable. These computational predictions can inform practical drug development and high-throughput screening by naming the currents that must change and by how much."