I went on the Code[ish] podcast to talk about AI, LLMs, and building Heroku's Managed Inference & Agents platform:
🎧 www.heroku.com/podcasts/cod...

Here is a recording of my live demo at PyCon US 2025 on building scalable AI tool servers using the Model Context Protocol (MCP) and Heroku

www.youtube.com/watch?v=01I4...

I was surprised at how clear-cut and blatant it was. I mean, two times in a row, closed fingers, correct angle.

Meanwhile, Musk has recently issued public support for the far-right wing AfD party, often described as anti-semetic / extremist.

www.cnn.com/2024/12/20/m...

That + no apology...

In honor of MLK day, here's super interesting essay my partner wrote on Martin Luther King Jr: what he actually believed and accomplished (different than what is sometimes described).

docs.google.com/document/d/1...

Incredibly impressive person.

Nice! Would love to be added (11 yrs in AI, co-author of Malware Data Science, love them NNs)

Am I reading this right? Techniques to make the model safe again had almost no effect on non-small models :o.

Sleeper Agents
arxiv.org/pdf/2401.05566

So many AI safety issues get worse, & harder to combat the larger and more advanced your model gets:

"The backdoor behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning"

A response to X is going to be (usually) written by someone socially, politically near X's author, vs some other random piece of content Y.

It's extremely hard to take out sycophancy out of an LLM, trained the way we train them.

Anthropic's "Towards Sycophancy In Language Models" arxiv.org/pdf/2310.13548

TLDR: LLMs tend to generate sycophantic responses.
Human feedback & preference models encourage this behavior.

I also think this is just the nature of training on internet writing.... We write in social clusters:

Say a model learns strategy x to minimize training loss --> Later, min(test loss) involves strategy y, but the model regardless sticks with strat x (inner misalignment).

Assuming outer misalignment, x can be seen as safer than y.

That being said, the better the model, the less this will happen.

In AI safety, we have inner misalignment (actions don't minimize the loss function) and outer misalignment (loss function is misspecified).

But I do think that inner misalignment (~learned features) tend to act as a protective mechanism to avoid outer misalignment implications.

I, er, really hope.