It is crazy mulling it over how GPT 4 would just straight up devolve to rambling at 30k tokens of context, while Opus 4.6 at a million, at worst, just becomes rather uncreative near the limit.

I was expecting it to increase faster than before with RL but not that fast, bit shocked if anything tbh.

I think there's a very large amount of tasks that, with access to models that costs pennies to run, are highly intelligent and don't need babysitting, will rapidly become common over the next years, it's hard to conceptualise because reliable thought is such a massive bottleneck in our current world

I get people have questions about the edge cases, but I really do wonder if opposing the concept in general ever represents an inability to imagine themselves in that situation where they were the ones dying in agony over months or had dementia or something without a policy like that.

That being said, I am much less sure on the out of context problem thing, whether the intelligence generalising will be able to handle modifying itself to handle problems outside of its RL or training scope entirely.

I think in a world where you can relatively easily simulate humanish intelligence, then the complexity created by competing intelligences is very solvable, or at least you can train on the capacity to manage it, same with modifying existing models for different goals.

It's quite plausible to imagine an economy where humans are still in charge, still reap the benefits, but actual human application of brainpower for organisation and data processing is functionally like using a horse to farm, pointless and expensive.

I really want to believe this, but any time i dwell on it, it just produces the conclusion that over time the coordination becomes just as automatable, heck if anything agentic orchestration is the same skillset as managing people.

I'm confident it's enough to do basically all white collar work, but very unconvinced we're getting any form of the machine god outcomes in the short-medium term.

Seems quite plausible that over the next 5 years, we get massive improvements from improved r&d and RL scaling, but it kinda peters out after that as we run into fundamental limits around the amount of intelligence you can fit into the frontier model weights' size.

Increasingly amenable to the idea of being fully automated if it means I don't have to spend another stupid evening bisecting a bloody obscure, undebuggable cuda issue.

How coding feels recently.

"Come and Take It" with the orange spikey Claude logo

I suppose maybe that's the point, existing organisational hierarchy design may provide a decent baseline for structuring the initial progress into RL in that area in the same way chain of thought prompting provided a good baseline for initial reasoning model training.

Ive really been mulling this recently that the current bottleneck is probably less intelligence, and more coordination, once agents can organise like humans can, there will be no need to become hyperintelligent to achieve significant gains, just add more agents.

I think my agents have a better social life than me at this point.

It's great how a thing for the right over the last few years has been excluding trans people from research on their own outcomes, arguing they're biased, then you look at the stuff the original "unbiased sexologists" that the right argues should be put in charge of trans people have been saying...

Finally found a use for my old Raspberry Pi 4, now to figure out what to do with it that doesn't involve giving it access to my Google account.

Give it a few hours and Elon will have altered the system prompt to suggest Alex was a domestic terrorist in response to every query again.

And more importantly does it even matter, i used to strongly believe it did, but it's hard to see the current status quo as any sort of end point.
I suppose the future is more in the process and architecture.

Moving to ML, I mostly learnt python, given that LLMs at the time were mostly usable for boilerplate and telling it *explicitly* what to change in a script. Now I watch Claude Code smash through problems that even I struggle with, and I wonder if I'd have learnt what I did if I'd moved recently.

Genuinely perplexing that they'd had the Google search integration working for like a year before Gmail got anything resembling a "turn this email into a calendar event" button.

I am now beginning to understand how people in the space industry feel watching Elon tarnish their entire industry’s reputation for his own short-term benefit.

Not that long at this rate from the company Tesla, becoming a historical footnote in the Wikipedia page about its own stock.

Are they going to make a pro war crimes gemini now?

At minimum it was a roe v wade style change to existing law that everyone is pretending was procedural because our stupid courts are apparently sacrosanct.

And if you argue about it, they'll tell you people were "lying" to trans people about their rights, despite the fact that several court cases had passed under the pretext that trans people did indeed have those rights. en.wikipedia.org/wiki/Taylor_...
It's complete gaslighting.

The whole line here being used is that this and the universal bathroom ban have been in force since the Equality Act 2010, a piece of legislation that literally everyone at the time recognised as a piece of legislation intended to provide additional rights to trans people. It's so absurd.

I lost a solid year of time a few years ago that I probably could have spent on more productive hobby stuff, stuck in a permanent loop between Factorio, Satisfactory, and Dyson Sphere Program. Together, they're all-consuming.”

Strong memories of working at a McDonald's on the tills where the music was exactly loud enough to be annoying, but exactly too quiet to be able to comprehend the beat or what was playing. On top of the usual lifelong hatred of Christmas music.