I made a bet internally that we wouldn't have a million people engage with tweets about Claude being a bridge, but I'm pretty happy to be on track to lose that bet.

Join us in Prague on July 17-20, 2024 for the 4th Human-aligned AI Summer School! We'll have researchers, students, and practitioners for four intensive days focused on the latest approaches to aligning AI systems with human values. You can apply now at humanaligned.ai!

The first ever detailed look inside a modern, production-grade large language model (in this case, Claude 3 Sonnet.)

Impressive progress by UK's AISI one year in -- and excited to see them opening an office in SF!

Over the last few months, we’ve increased our focus on developing evaluations for automated AI research and development, because we think this capability could be extraordinarily destabilizing if realized.

We are looking for ML engineers and researchers to help drive AI R&D

We were very excited to see the publication of the Frontier Safety Framework from Google DeepMind! More companies sharing their concrete proposals for preparing for transformative capabilities from AI systems is great: it increases the concreteness of the options available to the

👇👇👇

this is extremely cool
* activations & activation steering
* multi-tenant to keep costs for users down
* pip install-able
* actually taking a swing at public engineering infra 😍😍😍

If you're at #iclr24 , come see our poster (#129) tomorrow (Tuesday) at 10:45am to learn about the role human preferences play in making LLMs more sycophantic!

Enjoyed this paper that plots emergent abilities with pretraining loss on the x-axis, which is actually a suggestion that Oriol Vinyals also made a few years back: arxiv.org/abs/2403.15796

The paper uses intermediate checkpoints to plot a variety of pretraining losses. For some

We just clarified eligibility criteria for our agent benchmarks RFP (openphilanthropy.org/rfp-llm-benchm…) and included a link to METR's task development resources (metr.github.io/autonomy-evals…) which many applicants may find helpful

Peak wisteria on Stuyvesant Street in the East Village …

Very excited to see this come out:

This result is pretty clearly specific to the style of backdoor we're working with, and doesn't support broad claims like 'interpretability solves misalignment', but it's still surprisingly strong. Worth a look!

I very distinctly remember while I was in the thick of it making GPQA telling Robert Long that “I knew the project was going to be ambitious/hard, but I didn’t appreciate what that actually meant”

In retrospect I probably still would’ve done it, but we basically had to restart the

Full lecture slides and reading list for Roger Grosse's class on AI Alignment are up:
alignment-w2024.notion.site

🤖🥇🤖

I’m super excited to release our 100+ page collaborative agenda - led by Usman Anwar - on “Foundational Challenges In Assuring Alignment and Safety of LLMs” alongside 35+ co-authors from NLP, ML, and AI Safety communities!

Some highlights below...

I like to think of myself as a researcher, but almost certainly the most valuable use of my time is writing US Visa letters.

One of our most crisp findings was that in-context learning usually follows simple power laws as a function of number of demonstrations.

We were surprised we didn’t find this stated explicitly in the literature.

Soliciting pointers: have we missed anything?