blueboo

Youve struck on a key insight on language models (particularly pretrained ones, the more purely next-token predictor species.) This is a fascinating topic

Janus essay Simulators is the foundational text here https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators

You might follow up with The Waluigi Effect https://www.lesswrong.com/posts/D7PumeYTDPfBTp3i7/the-waluig...

But what’s tricky is that we post-train models, shaping these linguistic world simulators into something that has something like desires, principles. But It’s Weird. For more on that, check out “the void” https://www.lesswrong.com/posts/3EzbtNLdcnZe8og8b/the-void-1

A rising tide lifts all boats. It doesn’t mean the sea is made of boats

Cool

…but never buy v1 hardware folks! Especially for limited runs like high end laptops.

Apple quality comes from scale. A narrow product line means they have literally hundreds or thousands times more testing than PC ultra books. (And still — don’t buy a first iteration of a new Apple chassis.)

I think the answer is no, as this is a creative repurposing of css3d as a basis for a general 3d engine. That lets you put web content everywhere, even on the triangles of your teapot. Cool, but extravagant

Whereas THREE.js or webgl is purpose-built for realtime animated 3d scenes.

> That may suit Anthropic as private models can't be distilled

They can be distilled internally… expect great things from Sonnet 4.8

Why not the Pacific Ocean

In a master’s, you learn a lot about a little

In a PhD, you learn everything about nothing

Wait till Codex doubles prices/halves quotas on May 31

Billions in profit may unblock scaling customer support beyond scrappy startup minimums

However (and I loathe this logic) if you can get the marketplace to accept that minimal level, and the brand harm is inconsequential, why not pocket the savings

For those who find this appealing: Fires still burn down houses so what’s the point of fire departments?

Skill issue—in management.

Good leaders perceive workhorse vs showhorse spectrum, critical toil vs needless flash (and vice versa).

It’s hard. Most fail at hard things. The industry in the aggregate will fail at hard things

So you get articles like this.

Open source meth is still bad in absolute terms even if it’s better than the alternative meth

They’re not going to build their own, it’ll just be one of the many capabilities of the agent platform they use. A 2024 SaaS is just a playbook for next-gen AI.

You don’t buy a spelling correction program because it got built into Word. And now, the OS…

Indeed perhaps the most valuable lesson from travel is returning with the realization of just how poorly the generalizations and statistics describe the messy reality of a place. Everywhere has every sort of person

“In a sufficiently complex system, malfunction or even total non-function may go undetected for long periods, if ever”

John Gall, The Systems Bible

When faced with the choice of making genmedia more expressive vs looking higher production-value, big labs always choose the latter.

Expressivity requires intent, which turns out to be rarer and therefore less apt to show up in a benchmark or crowdsourced head-to-head comparison.

The recent Acquired ep on “Alphabet Inc” put it aptly: social media moved into Google’s space, video (reels, “pivot to video”), and social media for socialising moved to message groups, iMessage/Whatsapp/Discord.

Revenue-wise, video ads have always been the sun to print ads peanut m&m.

Look where the pucks going then:

Implication: ChatGPT as a realtime video avatar will hit the jackpot with ads, but not before. Count on the ChatGPT device having a screen for that reason

What? AI assistants are prohibited from providing legal and/or medical advice. They're not lawyers (nor doctors).

We call it “legacy DeepMind”

In certain niches, the marginal value of kneecapping the competition exceeds the viable budget for counteracting gaming. It may be a quirk of this reality’s hyperparameters that a UGC media monopoly inevitably suffers from this. Or maybe at a certain point it hits their bottom line and better enforcement is contrived.

Can't you trivially reframe the initial purchase as being subsidized by that license? Your $200 smart knife sharpener would be $300 if it weren't recording audio 24/7 (for VAD, surely!)

I don't like it either but here we are

When was the last time you did a deep research? Good agents just do research as necessary. I find GPT5 Pro >> all the top DR agents

These are fun to make. But the real value prop would be demonstrating the effect of following the proposed practice the app indicates.

…have you yourself actually tried it? Where was your technique and where is it now?

There’s literally a link to a thing you can use. Google 2025 has some great PMs and dev rel.

The safer the system, the more catastrophic the failures

Bear in mind 600 of those files are icon and screenshot variants for various screen dpis and spec ratios..

Coding to music was a flag you were just twiddling. People confuse flow with a pleasant surface engagement. Deep problem solving is when you sit forward, turn the music down, and bring up the big cognitive guns.

They’d probably be doing significantly better if they were LLM-guided

I used Whispering routinely last year; the value + glitches and ux failures drove me to gladly pay for Superwhisper; whose rough iPhone keyboard drove me to Wispr Flow (and tried otter too); whose poor transcriptions (oh THATS why they’re fast) drove me back to Superwhisper

Still lots of quality headroom in this space. I’ll def revisit whispering

You might find $2.50 in change to use one for an hour though