Boozed_Duck

[Boozed_Duck @DuckBoozed]

@Steve_Yegge Everyday I find myself repeating the same thing to Fable: "I don't know wtf you're asking me to rule on. Have Opus prepare an HTML report explaining the issues in PLAIN ENGLISH (cut the jargon), then I'll get back to you once we're on the same page"

[Boozed_Duck @DuckBoozed]

@johnmyleswhite I had a similar interaction recently. I asked for A and Fable proposed a "hybrid alternative" which was mostly B doing lip service to A, which is how "B is off the goddamn table" became a design constraint that Fable now has to work with.

[Boozed_Duck @DuckBoozed]

@doodlestein One of the most important aspects of this process, which I had long overlooked, is declaring your aims and verifying whether the plan and later implementation remain in alignment with your aims. A high-powered rifle, slightly off center, hits the wrong thing with great precision.

[Boozed_Duck @DuckBoozed]

@doodlestein Basically, agent output is systematically audited by other agents. Plans run through the same process before approval. It's a slow process and burns a lot of tokens. Even Fable makes mistakes.

[Boozed_Duck @DuckBoozed]

@c1nnamonbunny One disturbing theory is that Covid damages the same parts of the brain associated with frontotemporal dementia i.e., part of this might be virus induced brain damage at scale. Theoretically, we could expect repeat infections to make this issue worse and more prevalent over time.

[Boozed_Duck @DuckBoozed]

@janekm @QuentinCody @doodlestein I have a hunch that that's where the issue lies; it "fixes" things that aren't broken which breaks processes it doesn't understand. I usually keep the effort level for Opus set to medium, which has worked well enough so far

[Boozed_Duck @DuckBoozed]

@simonw Opus 5 can be somewhere between Opus 4.8 and Fable; same cost and intelligence as 4.8, but with jugdment and ambition approaching that of Fable (so it will still feel like you're talking to Fable). Keep 4.8 for... whoever the target users of 4.8 are.

[Boozed_Duck @DuckBoozed]

@babayagatwt Probably Google

[Boozed_Duck @DuckBoozed]

@DarkFootage_0_o Pretty sure it's a rabbit, or similar creature, and it would probably appreciate not having a flashlight shone in its eyes...

[Boozed_Duck @DuckBoozed]

@michaeljburry This. You don't just tell an agent to build something and let it rip. You need to develop the system that agent will operate in to ensure that the plan stays on track at every stage. That’s where token burn spikes and where you want to incorporate local models.

[Boozed_Duck @DuckBoozed]

@doodlestein https://t.co/8gXzjkPVHu

I probably could have worded this better, but if you want to be maximally ambitious, aim towards building simulation software that sets the benchmark for engineering applications. How will the structure hold under earthquakes, tsunamis, etc.?

[Boozed_Duck @DuckBoozed]

@doodlestein Goal: build an open-source competitor to Autodesk modeling and simulation software which offers better performance across every benchmark

[Boozed_Duck @DuckBoozed]

@0xSero Personally, I'd be happy if they released an Opus with Fable's judgment, ambition, and initiative. Opus can come across as an underachieving moron sometimes—think high intelligence, low wisdom—which can make it outright frustrating to talk to.

[Boozed_Duck @DuckBoozed]

@doodlestein Goal: build an open-source competitor to Autodesk modeling and simulation software which offers better performance across every benchmark

[Boozed_Duck @DuckBoozed]

@reed_barnes "The load-bearing realization: packages should be audited for CVEs *before* they're installed, not after."

- Claude, probably

[Moody @MoodyWriter13]

I’m not sure what the gravest consequence will ultimately be. There are many possible outcomes.

Anthropic could gain European sovereignty out of this, and from an investor’s standpoint it may even have to, just to keep the momentum alive. We know, after all, how the stock market values companies in which the government has a say and can pull the plug at any time. And there is a difference between a government that could do this and one that has just proven it will The current valuation cannot be sustained under those circumstances.

How it could work has long been sketched out: either through a legally independent European operating entity with its own operational control, or through licensing the model weights to a European operator under European governance. The blueprints already exist.

And the consequence for the US? It loses influence.

Once again, a decision by this administration meant to project power and toughness in the short term, but achieving the exact opposite in the long run. It gives away influence voluntarily. This administration’s path so far has been paved with moves like these. It mistakes the exercise of power for the preservation of power, and won’t notice the difference until both are gone.

The same playbook is already visible in the military sphere with Europe. For decades the United States acted as the continent’s security guarantor and benefited enormously from that role. This administration chose to exploit the dependency instead, and the consequence is a Europe now rearming on a massive scale, largely without US companies at the table. The US is losing political and economic power and influence.

The export restrictions imposed on NVIDIA’s sales to China are another example. The result is that China is actively reducing its dependence on NVIDIA GPUs and increasingly turning to homegrown solutions.

„Whoever tries to hold on to the world will lose it.” Laozi

[Boozed_Duck @DuckBoozed]

@seangeng I've been scratching my head on when/why I would want to use /loop, and this made me realize I already have a system that basically does this. Plan gets drafted, reviewed by 3 agents, fixed, reviewed again, and so on until the agents only come back with nit-picks.

[Boozed_Duck @DuckBoozed]

@MeidasClips @MeidasTouch Katie Phang : "Stop. We're not fucking stupid."

MAGA: "Who's 'we'?"

[Boozed_Duck @DuckBoozed]

@usedgov (Guy pointing to coworker who's also building inexplicable bullshit) "Ha ha look what Hunter do, he weld toilet pipe to boss's car exhaust"

[Boozed_Duck @DuckBoozed]

WELL... THIS CERTAINLY EXPLAINS A LOT.