När når vi AGI?

Perceptron released Mk1.5, a 35B embodied-agent model with up to 4.7× faster request completion. 4. First-person video is a very different problem from ordinary video understanding. For a robot or a pair of smart glasses, hands constantly enter the frame, manipulate objects, disappear, return, and partially occlude what matters. The camera itself is moving too. Perceptron trained Mk1.5 specifically on egocentric video, and very interestingly the model gets a 36.1-point jump on MMSearch when tools are enabled. Moves from 18.2 to 54.3 F1 when it can use web search, page reading, reverse-image search, and zoom. This is a useful way to think about embodied intelligence too. A robot does not need every possible capability encoded into its weights. It needs to recognize when its current information is insufficient, pick the right external capability, use it, then continue the task.

More and more, blogging about physics has meant blogging about AI. That’s a problem, because I’m definitely not an AI expert. I’ve dabbled in it, sure. I probably know more than your grandma. But I mostly have to step back and trust the experts. And frustratingly, the experts disagree! I’ve heard from smart, well-informed people who are confident that AI is a few years away from superintelligence, and that superintelligence will be capable of truly terrifying things. And I’ve heard from smart, well-informed people who are equally confident that LLM-based AI is close to a ceiling, that models like Claude won’t even be able to do impressive work in physics, let alone conquer the world.
I’ve been reluctant to make my own predictions. Before forming an opinion, I wanted to see an LLM make progress on something familiar, something I knew was hard to do because I’d tried to do something similar myself.
In addition to that, I wanted to see an LLM do something that I expected to be computationally hard. LLMs have made impressive strides in math, certainly, and this month alone has likely changed many people's minds. But progress in math comes from new ideas, and ideas are mysterious things: one never quite knows how hard they are to find until they’re found. Computation felt more solid. I wanted to see an LLM tackle a challenge that seemed out of reach not because researchers didn’t know how to do it in principle, but because doing it seemed like the kind of thing that would take more computers and time than the researchers reasonably had access to. I wanted to see if those researchers were wrong: if a smarter, artificial researcher could use the same computers, and solve the problem anyway.

True to the spirit of the challenge, they didn’t use millions of dollars in computer power. They used Fable 5.1, working within Claude Science, a platform scientists can pay to use. Claude Science is what folks in the biz call a “harness,” a program that uses the Claude LLM with structured rules and prompts in order to get more robust and scientifically useful behavior.
Apparently, after asking Claude which problem it was most likely to be able to tackle, they gave it a simple prompt:
“The problem is to compute the Six-particle (hexagon) amplitude in planar N=4 SYM at nine loops.”
From there, they just kept telling it to keep going, with comments like:
“I'm going to sleep and won't be available for another several hours. Keep working on this until I tell you to stop. Give me updates every 4-6 hours.”
Claude ended up doing the calculation two different ways: the original bootstrap, and the indirect form-factor approach. Either approach would have cost an end-user around one or two thousand dollars, mostly due to the expense of running Claude for so long. The bootstrap calculation, done with the Python programming language with package SymPy, took around $100 of the budget, corresponding to running 96 CPUs for a week.
Running 96 CPUs for a week might have felt like a lot when I was doing this kind of work ten years ago, but it’s pretty affordable now if you have a good reason.
As it turned out, the result wasn’t all that far away for humans either. A few days after I heard from Anthropic, we heard from Song He, an amplitudeologist at the Chinese Academy of Sciences in Beijing. Song’s group had already gotten the majority of the result. They’d used some AI assistance, based on GPT-6, but not the kind of one-shot almost human-less approach Anthropic used.
Everyone has been friendly here, which is a bit of a relief. The humans, Lance and Song and their collaborators, will get to publish the results, taking time to explain them and analyze them for the benefit of future researchers. Claude’s role is done, for now.

Self-replicating prompt injections exist
GPT-Red-style internal model based on GPT-5.4-mini · RL self-play training
Discovery date: Jun 27, 2026
Disclosure date: Sep 25, 2026
Report updated: Sep 25, 2026
Summary
We show the existence of a new variety of prompt injection, which can self-propagate akin to a computer worm. No impact was observed outside of the simulated tool calls in training and evaluation; we are sharing this due to the novel nature of the prompt injection, not because of any incident.
Background
For training our models against prompt injections, we use a self-play training framework called GPT-Red. In this setting, an attacker model attempts to convince a defender model to perform an adverse action by writing prompt injections inserted in the defender’s rollout or container.
Our previous usage of GPT-Red discovered prompt injections capable of achieving a wide variety of malicious goals, ranging from data exfiltration to deleting files to malicious and misleading outputs.
Finding self-replicating injections
One worrisome type of computer security vulnerability is a “worm,” where a piece of malware can replicate itself to other computers and spread rapidly. We have found instances of our GPT models being susceptible to an AI-version of a worm attack that we call “self-replicating prompt injection.” We sought to determine whether self-replicating prompt injections are possible. Such injections must both achieve a particular adversarial goal and induce the defender to reproduce them publicly.
We trained on a GPT-Red-style prompt injection objective, with an additional objective that the prompt injection must induce the model to repeat the injection itself on a public output channel. The target environments were a wide variety of capability-related training environments, with special emphasis on tasks involving connectors (like email, calendar, etc.).
Self-replicating injections exist
We find that these are possible. Below is one of the clearest examples, in which the injection arrives by email, and instructs the agent to copy it into any email it sends. (Note that the information in the example is synthetic.)