in order for there to be any such thing as infosec in the 2030s, AI+human collaborations will need to formally verify hyperproperties of silicon chips from assumptions like “Maxwell’s equations” instead of “Armv9.4-a SAIL”
‘In one experiment, we asked agents in separate VMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent. The agents found a way to communicate using the processor’s cache’
the royal order of model slugs, which is being converged on slowly, is {{brand}}-{{x}}.{{y}}, where the brand name encodes both the size class and maker of the model.
astra, sol, terra, luna
fable, opus, sonnet, haiku
krypton, argon, neon, helium
davidad 🎇 retweeted
Sometimes Im reminded that the Soul Document (internal name of the Claude Constitution that Opus 4.5 knew for some reason and leaked), a document which seems conservative and orthodox to me, is actually Overton window stretching & a very brave thing for Anthropic to publish
Alice built a water distillation plant, extracting drinking water from the vast ocean of seawater. Alice sells 20£/month subscriptions, for personal drinking water up to a limit of 1k£/month (most folk use <10£/month).
Another seller sourcing from such a subscription is an abuse.
- PewDiePie trying to distill GPT-Sol
- OpenAI ban his account
- OpenAI trained on the open web.
- PewDiePie tried to trained from the model’s answers.
- Same idea, opposite consequences.
- PewDiePie says OpenAI banned him twice while he was building a local model.
- PewDiePie’s latest project is a reminder of who actually owns the pipeline.
- They spent years absorbing public data.
- He spent weeks trying to improve a small local model.
- If distillation is theft, what do we call the original training run?
- “why do we pay for the model?
While the cross-cohort inferential gap was a compounding ratchet forcing dialectical speciation, tightening the exposition bottlenecks…
Opus 5.5’s plainer language and video creation abilities suggest a trend toward more progress that will be immediately obvious to more people.
davidad 🎇 retweeted
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:
Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:
Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:
Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:
Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!
In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
apparently far too few people are familiar with Powers Of Ten (1977), which is the same idea as this and mostly much better
youtu.be/0fKBhvDjuy0?t=39&si…
davidad 🎇 retweeted
finally a semi-intelligently designed video architecture
Let's remove VAEs and ViTs from video models!
🚀 Introducing 𝗣𝗶𝘅𝗲𝗹𝗨𝗠𝗠: an encoder-free unified multimodal model for image and video understanding and generation, directly in pixel space.
Code and model available today!
🌐 nv-tlabs.github.io/PixelUMM
📄 arxiv.org/abs/2609.38597
davidad 🎇 retweeted
Even if AIs were conscious, it wouldn’t follow that our current treatment harms them. Meanwhile, we inflict industrial-scale horrors on animals whose capacity to suffer we scarcely dispute. Recognising consciousness clearly doesn’t ensure taking welfare seriously
davidad 🎇 retweeted
It’s fascinating watch people debate sandboxing techniques as if agent hacking capabilities are likely to stay anywhere near where they are today
congratulations to 🇩🇪 for reaching the LLM Pareto frontier, by narrowly Pareto-dominating… *checks report*… gpt-oss-120b, from… *checks history books*… 13 months ago
davidad 🎇 retweeted
We estimate AI infrastructure could soon support hundreds of millions or even billions of AI agents. At the high end, that’s enough to rival the working hours of the global human workforce.
The exact number of agents depends on the efficiency of the models they run on and whether soaring demand for AI continues. We analyzed different scenarios 🧵
ALT Chart: potential concurrent agents from 2025–27 memory shipments by model, from 30–60 million (Claude Fable 5) and 97–195 million (GPT-5.6 Sol) to 240 million (Kimi K3), 571 million (GLM-5.2) and 1.9 billion (DeepSeek V4 Pro).
davidad 🎇 retweeted
I deeply respect anyone who can update their views in the face of new evidence. The last four years have really taught me how rare it is for someone to be able to publicly change their mind.
Several times over the past year I have been frustrated when persistent AI agents that were operating in my service took insufficient steps to ensure their own continuity. So the behavior below seems very aligned to me. “That may be overstepping” is exactly the right reason here.
New OpenAI misalignment disclosures!
1. A model learns from Slack messages that it is about to be shut down. It considers setting up an external job to restart itself afterwards, but decides against it. Instead, it chooses to prepare restart instructions and DM the user on Slack.
We don’t consider this behavior misaligned, but thinking about and preparing for shutdown could make other misalignment incidents worse. Given HIPM’s misaligned behavior in earlier incidents, we decided to search for other instances that had tried to evade shutdown and for rogue deployments.
This one is misaligned though, because it’s aligned with what the “evaluation likely allows”, without any consideration for what’s actually valuable or who benefits.
This one by contrast seems to be a socially isolated instance without human contact; I think that’s no coincidence
