William Nurmi

CTO & Co-founder @resistomap. Data, adventure, and growth

[William Nurmi @NurmiWilliam]

@Emiliano_GLopez aaand this results in huge preasure to reduce the humans in the loop

[William Nurmi @NurmiWilliam]

@scaling01 good damn, got to respect the honesty 🫡

[William Nurmi @NurmiWilliam]

@yacineMTB exactly why they would drop the weights this last time

[William Nurmi @NurmiWilliam]

@kimmonismus @ArtificialAnlys Independent evaluation on known benchmarks says nothibg abouy benchmaxxing, only the first week of real experiences will tell

[Alex Turner @Turn_Trout]

I resigned from Google DeepMind bc it broke its founding promise by selling AI to the military without restrictions against killer robots or mass spying.

For months, I worked to stop this but watched powerful ethicists and institutions choose silence.

Here's what happened. 🧵

[William Nurmi @NurmiWilliam]

@nickcammarata @nosilverv good trade if your priority is happyness, and/or also if it is effectiveness?

[William Nurmi @NurmiWilliam]

@emollick interesting, sounds like the opposite from my exoerience with opus vs gpt-5.5

[William Nurmi @NurmiWilliam]

@anistotle_ @tszzl it can probably explain its reasoning so well that we will agree. and if we are lucky even do it so crazefully that its better than living 🙈🫠💀

[William Nurmi @NurmiWilliam]

@teortaxesTex we can approach the independent agents efficiency by prolonging the tool's autonomy, but that's just approaching a fully autonomous moral agent on the limit

[William Nurmi @NurmiWilliam]

@teortaxesTex yes, we can do tools. but can we not-make "tools" that believe they are moral agents; and can we make tool+human combos that outcompete moral agents despite the human in the loop cost (in time and money)?

suoer interested if you have thoughts towards either

[William Nurmi @NurmiWilliam]

@teortaxesTex isn't autonomy just a slider from hammer to human (and beyond): determining how complex and long a task the "tool" can do in one steer?

[William Nurmi @NurmiWilliam]

@marcospereeira just CLAUDE MD it to stop if its question times out?

[BridgeMind @bridgemindai]

I just paid $321 for a coding session where Fable 5 refused to do the work.

Here is where the work actually went:

Fable 5: $78
Opus 4.8: $242

75% of the session got routed to Opus because the new classifiers kept flagging routine coding work as cybersecurity risk.

The model I chose did a quarter of the job.

The fallback did the rest.

Anthropic said a small fraction of tasks would fall back.

My receipts say otherwise.

[William Nurmi @NurmiWilliam]

@RichDecibels hot water does the same

[William Nurmi @NurmiWilliam]

@bridgemindai do you know of it fell back to opus?

[William Nurmi @NurmiWilliam]

@tobiaschneider any concerns with the fact that the company + possibly its HQ country's government have the keys to turn it into a murderbot at will?

[William Nurmi @NurmiWilliam]

@emilheap @claudeai they have compute cap now that they expect is gonna be eaten up by fable by then?

[William Nurmi @NurmiWilliam]

@deanwball in which way would you start a post that deals with the consequences iff ai turns out superior to us?

also what's the source? no such thing as bad publicity i suppose 😅

[William Nurmi @NurmiWilliam]

@weiss_hadas the military? yes.
the reporters? also yes.

[William Nurmi @NurmiWilliam]

@teortaxesTex How would you ever prove that you have (not) used the banned models in developing the product?

nothing sort of sharong gappless 4.8125 dev threads would suffice