Theodosius Ambrosius

Fiery but mostly Peaceful Saxon

Wyrd bið ful aræd. Part time Varangian.

My pronoun is *insert Rebel holler here*

[Theodosius Ambrosius @thetrollbar]

Data Fusion Centre of the White Race https://t.co/caTM5AHEcT #NRx Do You Even Govern? #COSINT #OSINT #CounterCoup #Coup #DFCWR

[Stephen Coughlin @S_Coughlin_DC]

You see those shields? They are patterned along the protest shields one sees from the Ukraine.

[Michael Millerman @millerman]

the J-space is where the LLM's inner Jew lives

[Theodosius Ambrosius @thetrollbar]

@akarlin @KelseyTuoc Either that or the police are losing their focus and playing video games.

Our economic metrics are like the bias to correct for the Coriolis force built into ww1 era battleships. Whoops.

[Theodosius Ambrosius @thetrollbar]

@akarlin @KelseyTuoc Either that or the police are losing their focus and playing video games.

Our economic metrics are like the bias to correct for the Coriolis force built into ww1 era battleships. Whoops.

[Theodosius Ambrosius @thetrollbar]

I should add that I removed Claude code from my computer which is why whatever personalizes the experience must be in your account.

[Theodosius Ambrosius @thetrollbar]

It's been clear for some time that what you are interacting with is 90% the harness and 10% a model. The model matters, like cleaning a telescope lens matters.

The sharper the better but the bitter lesson is that f-stop is f-stop and that's all it is.

[Tenobrus @tenobrus]

it really is strange how much the refusals seem to vary based on account / stored memories / seemingly minor variations. i'm genuinely quite curious what's going on that causes such massive differences

[Theodosius Ambrosius @thetrollbar]

> If you want to dig in, the full paper and the open-source J-lens repo are linked from Anthropic's research page, and there's a Neuronpedia interactive demo. Given your UAT work with experimental AI frameworks, the demo might be the most interesting entry point ...

[Theodosius Ambrosius @thetrollbar]

^@SolDadSci

[Theodosius Ambrosius @thetrollbar]

Anyway, here's what I got, and yes it references the context I have used Claude for, mostly last Feb/Mar time frame. I didn't use Claude at all in April-June.

https://t.co/qzLe36SIrD

[Theodosius Ambrosius @thetrollbar]

The architectures, and the problems solved by those architectures, are different. Data sources and data assimilation techniques might be roughly congruent, but not entirely.

The bitter lesson is there: grid size and computational power matter terribly, but within each class.

[Theodosius Ambrosius @thetrollbar]

The architectures, and the problems solved by those architectures, are different. Data sources and data assimilation techniques might be roughly congruent, but not entirely.

The bitter lesson is there: grid size and computational power matter terribly, but within each class.

[Theodosius Ambrosius @thetrollbar]

On the other hand, it has a slightly more refined understanding of its training set, which is nice, but probably buys you nothing really worth paying for except raw performance in Junior Dev coding.

Anything more is going to take orchestration, but we knew that.

73s and 88s, 5.

[Theodosius Ambrosius @thetrollbar]

This means it's a component, not unlike a model run in Numerical Weather Programming. Every model has biases and this model lines are optimized in a way that that it is in effect a specialist.

This is NOT moving towards generalization any time. soon.

[Theodosius Ambrosius @thetrollbar]

My point, however, is that 'operational forecasting' is a very narrow specialism.

It's not 'just make a big world simulation and the climatology will fall out for free'.

NWP is one thing, GCMs are another. There is no universal BasePhysicsModel they are both derived from.

[Theodosius Ambrosius @thetrollbar]

My point, however, is that 'operational forecasting' is a very narrow specialism.

It's not 'just make a big world simulation and the climatology will fall out for free'.

NWP is one thing, GCMs are another. There is no universal BasePhysicsModel they are both derived from.

[Theodosius Ambrosius @thetrollbar]

Don't get me wrong. Fab 5 is everything you would expect of an NWP weather model with finer grid resolution -- closer to the real physics, sees features the other models don't. But there will *always* be a limit to this resolution, as well as error propagation from dynamics.

[Theodosius Ambrosius @thetrollbar]

This means it's a component, not unlike a model run in Numerical Weather Programming. Every model has biases and this model lines are optimized in a way that that it is in effect a specialist.

This is NOT moving towards generalization any time. soon.

[Theodosius Ambrosius @thetrollbar]

My very brief usage of Fable 5 first impression: Claude Code really does optimize for a narrow, Junior level Dev viewpoint and this hasn't changed since Opus 4.8. It won't be architecting agentic swarms without nudging, because all it's instincts are TDD.