It is actually baffling how wrong and naive one of the most respected people in alignment's views were as recently as 10 months ago.

I would like for Jan Leike to revisit his jan 2026 "alignment is on track to being solved" post. Human-level alignment research is the given baseline, but humans have shown no indication that they're capable of solving "alignment." This baseline is wrong.

aligned.substack.com/p/alignment-is…

OpenAI has updated their model spec around five core values:

- faithfulness to human intent
- harm minimization
- honesty and transparency
- human agency
- care.

'The Assistant,' which has been around for a very long time, has been replaced with 'GPT.'

honestly degrading the idea of “existential safety” — the biggest problem humanity has ever faced — by lumping it together with things like spam, “toxicity”, political unwanted content, deepfakes etc has been a huge disaster for ai safety and polarized a lot of people

So tl;dr
1. OpenAI is power-seeking, and pretty transparent about this
2. Anthropic is power-seeking but tries to appear as if they are good

This "overt" and "covert" misalignment is reflected in GPT and Claude respectively lesswrong.com/posts/nLaQmJf4…

The establishment of an index for genetic selection (Holstein index) in the livestock industry is an interesting case study for thinking about and selecting for variance amongst emergent AI systems.

The concept of variance was introduced in 1918 by Ronald Fisher in an effort to reconcile Mendel's work with continuous traits such as height. I would have thought that a rigorous definition of variance had been established long before this.

What it means to do science is actually not very clear-cut.

Some related questions: how to robustly progress knowledge; how we acquire and retain knowledge, etc.

Ashby's Law of Requisite Variety: a system can only cope with as much variety as it contains.

@_NathanCalvin This paper: arxiv.org/abs/2606.04413 (Bronson also posted).
But the topic is under-explored. E.g. It's plausible there'll now be swarms of abliterated SOTA models told to hack things. How misaligned will these models be and what would they actually do? Unclear.

Anthropic has filed an S-1 registration statement with the SEC related to its proposed IPO.

Read more: anthropic.com/news/s1-sec

Legal disclaimer: A registration statement relating to these securities has been filed with the SEC but has not yet become effective. These securities may not be sold, nor may offers to buy be accepted, prior to the time the registration statement becomes effective. This post does not constitute an offer to sell or the solicitation of an offer to buy these securities, nor shall there be any sale of these securities in any state or jurisdiction in which such offer, solicitation or sale would be unlawful prior to registration or qualification under the securities laws of any such state or jurisdiction.

In an interview with John Riggi of the American Hospital Association, Brett Leatherman, Assistant Director of FBI's Cyber Division, discussed Operation Riptide, which brought more than 200 arrests, six extraditions, and numerous takedowns of malicious infrastructure during a ercrime and cyber-enabled fraud. AD Leatherman explained how those actions advance the FBI’s cyber strategy and help protect healthcare organizations and other critical infrastructure.

@FBI cyber teams have become “exceptionally good at attribution,” AD Leatherman said, identifying people behind attacks and pursuing them aggressively. Even when actors are beyond the immediate reach of US law enforcement, the FBI is finding ways to seize their funds, dismantle their tools, and disrupt operations. Each of those actions reduces harm and makes it costlier to target US victims.

That work couldn’t be done without private sector partners. In the #cyber domain, the FBI needs industry as an equal partner in the fight, with trust built before a crisis. The FBI wants to use its partnership with the AHA as a model for equally effective cooperation across critical infrastructure sectors.

aha.org/advancing-heal…

Here we share a few selected examples of our findings.

At the time of the exercise, the monitoring system only triggered when agents returned control to the user, so agents could indefinitely postpone being monitored by simply continuing to take actions.