Do some crop circles too, while you're at it.

I feel like there's so much agentic overhang. They seem to be way too over-trained for knowledge than for execution.

On one hand, this is "good" - they're more like the Oracles/CAIs. But the gap is so glaring and there's a sizeable demand for agentic AIs, so this imbalance basically forces the AI labs to go and adjust.

We would not like a "dumber" model in general and chopping the knowledge training is detrimental for grounding. This is demonstrated by the oX-mini series of models - they IMO suck to the point of being unusable and uninteresting.

So, I expect the upcoming models to not lose more smarts (like we've seen with 4.5->5 which got OpenAI into crosshairs) while gaining more in agency. GPT-5 is a visible step-up for me in this aspect. Even coming from o3 it is clearly more in control and more prepared for unagumented tasks.

Love the game setups. The less augmented the merrier. Unlike benchmarks and CTFs they really take the learned skills for a walk. "X plays pokemon" made the other leaderboards obsolete for me.

I was preparing to do a "Claude plays Universal Paperclips" stream of my own and found some of the problematic points too.

Cookie Clicker, where you accrue currency primarily by sitting and waiting, and then spend your currency on upgrades that get you more of it. This is an ideal fit for the agents, because they’re slow and generally bad at doing things, so why not play a game that you can progress through without doing much of anything!

It is nowhere near the "ideal"! Despite the name, idlers require tons of micro-management to perform well and regularly halting to a grind.

That one Progress Knight is one of the faster-paced. Even with the auto-promote/auto-learn the agent has to switch between the tasks rapidly and the game punishes severely for sitting idle. Try to play optimally and you'd have to glue yourself to the screen or face the setup collapse.

In effect the "idlers" burn through tokens like there's no tomorrow if you want the performance that's more interesting than watching paint dry.

So I expect that only 2026 LLMs trained with agentic RLVR will give a first reasonable glimpse of what this method gets us, the shape of its limitations, and only in 2027 we'll get a picture overdetermined by essential capabilities of the method

I'm at least 50% sure that this timeline would happen ~2x faster. Conditional on training for agency yielding positive results the rest would be overdetermined by EoY 2025 / early 2026. Otherwise, 2026 will be a slog and the 2027 wouldn't happen in time (i.e. longer timelines).

Could be handy for our next "overthrow the government" day celebration /s

I appreciate that AI 2027 named their model Safer-1, rather than Safe-1

That's because they can read its thoughts like an open book.

I don't think they're blocked by an inability to run autonomously. They're blocked by lacking an eye for novelty/interestingness. You can make the slop factory to run 24/7 for a year and still not get any closer to solving alignment.

says little about the intelligence of Claude

It says that it lacks intelligence to play zero shot and someone has to compensate the intelligence deficit with an exocortex.

It's like we can track progress by measuring "performance per exocortex complexity" where the complexity drops from "here's a bunch of buttons to press in sequence to win" to "".

AIs (probably scaffolded LLMs or similar)

That was a good start, but then you appear to hyper-focus on the "LLM" part of a "blogging system". In a strict sense the titular question is like asking "when will cerebellums become human-level athletes?".

Likewise, one could arguably frame this as a problem about insufficient "agency,"

Indeed. In a way, the real question here is "how can we orchestrate a bunch of LLMs and other stuff to have enough executive function?".
And, perhaps, whether it is at all possible to reduce other functions to language processing with extra steps.

but it is mysterious to me where the needed "agency" is supposed to come from

Bruh, from the Agancé region of France of course, otherwise it's a sparkling while loop.