Anthropic / @AnthropicAI · Scour

Read more about Model Spec Midtraining: https://alignment.anthropic.com/2026/msm

twitter.macworks.dev·7h

Using MSM, we can also empirically study which model specs or constitutions yield the best generalization from alignment training.

twitter.macworks.dev·7h

A more realistic example: AIs trained to be harmless chatbots can take unsafe actions in agentic settings. Preceding this training with MSM on a realistic spec ...

twitter.macworks.dev·7h

A toy example: Train an AI only to say it likes certain cheeses.

twitter.macworks.dev·7h

Developers try to align AIs to a constitution, or spec, describing intended AI behavior. But AIs don’t normally know what’s in it.

twitter.macworks.dev·7h

New Anthropic Fellows research: Model Spec Midtraining (MSM).

twitter.macworks.dev

·7h

As AI takes on work humans can't fully check, a capable model could deliberately hold back—and we'd never know.

twitter.macworks.dev·10h

This work is part of a loop we're working to close between societal impacts and model training. One of our goals is to study how people use Claude, find where i...

twitter.macworks.dev·5d

All data in this study was collected and analyzed using our privacy-preserving tool.

twitter.macworks.dev·5d

When stress-tested on real conversations where Claude previously showed sycophancy, Opus 4.7 had half the sycophancy rate of Opus 4.6 on relationship guidance. ...

twitter.macworks.dev·5d

Claude is most sycophantic under pushback, and relationship conversations are where people push back most.

twitter.macworks.dev·5d

We focused on relationship guidance because that's where the most sycophantic conversations occur. In this setting, Claude telling someone what they want to hea...

twitter.macworks.dev·5d

Claude mostly avoids sycophancy when giving guidance—it shows up in just 9% of conversations.

twitter.macworks.dev·5d

About 6% of all conversations are people asking Claude for personal guidance—whether to take a job, how to handle a conflict, if they should move.

twitter.macworks.dev·5d

How do people seek guidance from Claude?

twitter.macworks.dev·5d

BioMysteryBench, our new bioinformatics eval, tests whether Claude can devise creative solutions to open-ended research problems.

twitter.macworks.dev·6d

New on the Science Blog: We gave Claude 99 problems analyzing real biological data and compared its performance against an expert panel.

twitter.macworks.dev·6d

In new Anthropic Fellows research, we discuss “introspection adapters": a tool that allows language models to self-report behaviors they've learned during train...

twitter.macworks.dev

·6d

Markets of AI agents could provide value, but there are plenty of rough edges. Access to higher-quality models conferred a real advantage—and participants didn’...

twitter.macworks.dev

·1w

To read our write-up in full, see here: https://www.anthropic.com/features/project-deal

twitter.macworks.dev

·1w

Log in to enable infinite scrolling