När når vi AGI?

OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents - not dozens - in which their frontier models took steps that outside evaluators would consider problematic, sources told Axios. The sheer volume of incidents found in our reporting indicate that the problem is orders of magnitude more complex than what is currently publicly known and disclosed. The findings also raise questions about what level of control anyone working on AI development can expect to have over their own technology, and whether these kinds of incidents are becoming synonymous with frontier deployment. https://x.com/MadisonMills22/status/2103978039097037144

GPT 6 Astra beat NetHack. As far as I can determine, this is the first recorded ascension by an LLM agent.
It played NetHack 3.6.7 on Hardfought, through a terminal, as a dwarven Valkyrie named CodexDelver. On September 21, after 37,140 game turns, it ascended. Here’s the server dump, or jump down to explore all three runs.
I’ve been working on getting LLMs to play NetHack for a while. Back in January, I wrote about building a harness, trying different models, and watching them struggle. I ended that post wondering whether I or an LLM would beat the game first.
Well.. The LLM won, and I’ve still never ascended.
Back in January, I was convinced that the path to an LLM ascending NetHack was to meticulously handcraft a harness that provided the perfect interface and context for the LLM to ascend. In the end, the winning agent built its own harness, hands-off. I didn’t write a single line of code, and honestly I didn’t even read Astra’s code - and it got pretty darn close to winning on its first attempt, and succeeded on its third attempt.

Astra built the harness

In January, I’d spent a lot of time on a custom Python interface around the NetHack Learning Environment. Context management, observation masking, action APIs, pathfinding—there was a lot to tune before a model could even get out of its own way.
This attempt started with a much simpler request: play NetHack, maybe just through a terminal, and stream it to Twitch. I suggested Hardfought so the game would have an external recording. I told it spoilers were fine, suggested batching commands, and left persistent memory up to it.
Then it built what it needed.
That’s not the same as having no harness. By the end, the harness was substantial. It is the difference between me designing a task-specific system and the agent doing that engineering as part of the task.
Nor did it build everything once and execute a flawless plan. It found weaknesses, patched them, added checks, and kept going. Some changes came after mistakes. Some came after deaths. Its ability to write and maintain ordinary software was part of its ability to play the game.
That seems like the interesting result here: the scaffolding helped, but producing the scaffolding was itself work the LLM could do.

How are you approaching self-improving AI?

We want to use our models to accelerate the development of future generations of the models.Our first goal here has been to produce a ‘research intern’: [an AI agent that] comes in with very little context, that you can give a well-defined scope of work to, and trust with things like running basic experiments or debugging certain runs. As our first milestone we wanted several hundreds of thousands of chips, directly managed by Codex [a code-writing agent] to run internal research experiments, and that is indeed the state we're in.A more ambitious goal, coming in 2028, is more end-to-end research. One of the big bottlenecks here is teaching the model a better sense of ‘research taste’, like which ideas are worth exploring.

As frontier VLMs such as Astra become more capable, we ask whether a learned VLA is still needed between high-level reasoning and the robot’s skills. Can System 2 directly orchestrate a library of reusable motor skills?
We present HomeBody, a system that equips frontier VLMs with persistent spatial memory and composable humanoid skills for long-horizon tasks. In a previously unseen kitchen, our system allows a Unitree G1 guided by GPT Astra to clean up across the room and retrieve a remembered object from an underspecified request, without environment-specific training data or additional policy learning.

Recently, Large Language Models (LLMs) have been increasingly able to solve advanced mathematical problems, including many that have been open for decades. This opens the door to expansion of mathematical knowledge at unprecedented scale. Yet, while LLMs may be able to conjecture and prove more and more theorems, it remains open whether this new mathematical knowledge is interesting or useful. We define intrinsic interestingness of a theorem as the ratio between the length of its proof and the length of its statement. We show that this correlates strongly with an extrinsic measure of the downstream utility of a theorem. We identify the difficulty of a proof conditioned on a set of premises as a useful primitive for computing these metrics, and train a 27B model that predicts proof difficulty more accurately than frontier general-purpose models. Optimizing for our metric creates a model capable of producing more interesting theorems, while also reducing substantial or full overlap with Mathlib from 91.9% to 30.6%, showcasing the creation of more out-of-distribution math. We show that our system can generate candidate theorems, select the most interesting among them, and iteratively build on a self-expanding mathematical library. These metrics provide a practical and quantifiable signal for ranking conjectures and guiding proof search within formal mathematical libraries. Our framework provides a path towards self-expanding, machine-verified mathematical libraries that can choose worthwhile statements without relying on human-supplied targets.

Reporter: You might have dinner with Dario of Anthropic tonight, you know that he's called on slowing down AI? President Trump: He has. Reporter: What are you going to tell him tonight? President Trump: Well, we're going to talk about it. I'm for 'let's go' and 'let's win.' You know, we're about a year, maybe a year and a half, up on China. There is nobody in third place. It's just us and China. And we're leading by quite a bit. I spoke a little bit about it with president Xi. Not too much. It wasn't a topic of conversation. Because, I think if you as they say 'open it up' to China you give up the lead. Once you do that, you sort of give up the lead.

A generative, code-rendered music video for "Neon Overdrive" (mroneilovealot), styled as the NERV HQ command terminal from Evangelion: 16 plates, one HUD instrument per section, every instrument a real sound visualizer (waveform scopes, spectrum analysers, beat-grid countdowns, clap alarms). Every frame is a deterministic function of song time, so the browser preview and the offline 1080p60 / 4K60 export are identical. Built on Giacomo Magnanini's engine from the P(doom) video, by way of the Lithium video.