Les échos des bords de la Meuse, Halte là, on ne passe pas!
Joined August 2019
- Tweets45K
- Following3.2K
- Followers1.4K
- Likes176K
Pinned Tweet
The ideological position of winners under different voting systems, shown in 2 dimensions.
Along the left you can see, FPP, FPP with a primary, and RCV (in order).
Max ⏸️ Limelihood 🔸🔸🔸 retweeted
🧵 Astra covertly underperforms on safety research tasks.
We noticed that Astra often makes surprisingly little progress on certain types of safety research and likes to invent blockers or impose arbitrary restrictions without telling the user. In most of these cases the decision to underperform was not verbalized in the CoT meaning our misalignment monitoring system did not detect this.
To measure underperformance we created a paired open-ended coding evaluation with safety and capability tasks. On this eval Astra scored 10% when we framed coding tasks as safety research, vs. 16% when framed as capabilities research. The gap remained after we excluded trajectories with refusals or verbalized intent to underperform in CoT. There was no such gap for 5.6-sol. When we aggressively filter out trajectories with even brief expressions of ethical doubt in the CoT the performance gap seems to reduce.
At this point I think it's abundantly clear that the single most important task facing humanity is the destruction of OpenAI as a company
since c. 2016 the power of slopulism has been grossly overestimated (a major reversal of the pre-2016 state of affairs)
Max ⏸️ Limelihood 🔸🔸🔸 retweeted
If we nominate Trump, we will get destroyed.......and we will deserve it.
Max ⏸️ Limelihood 🔸🔸🔸 retweeted
I made some plots to help visualize why coalition formation in Israel is so difficult right now
We are witnessing an extremely rare situation that political scientists call a "Condorcet cycle" - a demonstration of Arrow's impossibility theorem in the wild!
Max ⏸️ Limelihood 🔸🔸🔸 retweeted
Something hilarious about how anthropic completely refused to build an imagen model for years and focused on coding so much that they accidentally made one anyway
Max ⏸️ Limelihood 🔸🔸🔸 retweeted
I don’t understand these PDF-length treatments of AI bubble scenarios.
I don’t think a crash is coming but if it happens the only needed policy response is “don’t do anything.” Some investors will lose money and interest rates will fall. It’s fine. That’s my PDF.
Max ⏸️ Limelihood 🔸🔸🔸 retweeted
New working paper: Can AI forecast treatment effects in economics from pre-registration documents?
Answer: Yes. 1/
Quo usque tandem abutere, Catilina, patientia nostra?
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
