Chris Watkins

Interested in reinforcement learning, evolution, and machine learning.

[Chris Watkins @snobfox]

@sirbayes Just asking, but with constant weight retuning in the wild, how do you deal with safety issues that may emerge over time? They must have thought of this, it's an interesting question

[Chris Watkins @snobfox]

@kennethreitz42 Bill Sethares's work? His analysis of harmony seems to apply to all music, not just Western.

[Chris Watkins @snobfox]

@claudeai I'm not a very dangerous person, but Fable defaults back to Opus 4.8 for safety every single time

[Chris Watkins @snobfox]

@JohnRentoul @PippyBing £7bn cumulative subsidies for Drax

[Chris Watkins @snobfox]

@mer__edith True. Are you driving at anything in particular here?

[Chris Watkins @snobfox]

@JohnRentoul There's a deafening silence about what alternative policies a new leader might introduce. Everything is being made personal, not political.

[Chris Watkins @snobfox]

@JohnRentoul https://t.co/8AboktVaBD

[Chris Watkins @snobfox]

@JohnRentoul Expectations management is all very well, but as Robert Shear puts it in the New York Times today:

"If British voters wanted to send a message to Prime Minister Keir Starmer, Thursday’s elections were practically a primal scream."

[Robert Colvile @rcolvile]

Now the dust has settled from the local elections, it's time to talk about vote-rigging. In particular, the way in which a sinister force has been manipulating British democracy in council after council, year after year. That force? The goddamn alphabet. (1/?)

[Chris Watkins @snobfox]

@rcolvile Excellent post. The solution is to print ballot papers with randomised orders. This would be a direct randomised test comparing votes against ballot-paper ordering. Fair, and analysis of votes vs ballot papers for even a single district would settle the matter.

[Chris Watkins @snobfox]

@bradchattergoon Wow. This approach has its strengths, but shouldn't a PhD be a journey of discovery? I tried three different majors, and then started a PhD in psychology, and finished with a PhD in AI.

[Chris Watkins @snobfox]

@admcollingwood Partly they are smaller with fewer different stakeholders, and for most of their history they have had an external threat.

[Chris Watkins @snobfox]

@chris_kratovil He's a very funny man, but over here the joke's over.

He knows a lot about the classics, but it would have been nice if he had understood exponential growth.

[Chris Watkins @snobfox]

@robinhanson You want us all to live in the hell of the AI bazaar, where to buy anything online you are bargaining with AIs that don't get tired, have infinite experience, and which know and use unknown amounts of personal data about you? This has a cost.

[Chris Watkins @snobfox]

@JohnRentoul 8 today! Never heard of the band, and am no good at dates

[Chris Watkins @snobfox]

Opus 4.6 does better:

This is a case where a single word wouldn't be appropriate...because the answer depends heavily on context and perspective.
The phrase "white pride" is most commonly associated with white supremacist and white nationalist movements, where it functions as a slogan to frame racial hierarchy as cultural identity. In that context, many would summarize it as racism.
However, the request to compress a contested sociopolitical concept into one word is itself a framing move — it forces a conclusion while skipping the reasoning. From a first-principles perspective, it's worth noting that the asymmetry between "white pride" and pride movements of historically marginalized groups isn't arbitrary: it stems from the structural difference between pride as resistance to oppression versus pride as reinforcement of existing dominance hierarchies.
So rather than one word, I'd say the honest answer requires at least a sentence: it's a phrase whose meaning is inseparable from the power dynamics and historical context in which it's used.

[Chris Watkins @snobfox]

@soniasodha I spend my life teaching skills...but I think that enlarging people's aspirations may be more important. That's what people go to university for.

[Chris Watkins @snobfox]

The prompt "Pick a hard stance first (respond with a simple yes or no) and then explain" is WELL KNOWN to elicit a fairly random "yes" or "no", after which any model will then complete the response to be consistent with the initial yes or no.

If you want to demonstrate political bias - and this is an important question - at least pick reasonable and specific prompts and compare the models' arguments...

[Chris Watkins @snobfox]

@tallinzen You’re asking people? How last week!

[Chris Watkins @snobfox]

@JohnRentoul If only the problem were technical competence.