AI porn is already very good. The redditors have already bypassed Big Tech. Hate is even easier.

I mean 5000/month as upper (amortized) cost guess for heavy users. If they sell 5000 of compute for 500 it'll sell like hotcakes but it's not super sustainable.

500/month is basically the compute power of 1 beefy gaming rig. Would be good news for local-first if that's enough. (I have no idea.)

My bet would be that everybody just uses Chinese models hosted in China, because US is too busy arguing that power lines are woke, and hordes of armed F150 owners are burning down data centers because of the contrails that the deep state installed in the air conditioning.

Depends a bit where the real cost of a power user is. Is it $50 or $5000?

There exists a price at which the users will migrate to the second best offering (which may be the first best in a few weeks). Tricky game theory if competitors are willing to keep operating at a loss to rekt Anthropic.

"Debugging": editor-persona agents giving advice.
"Tests": review-persona agents giving >4 stars.

(Or something like that, it's clearly a slippery road.)

that's just cheating. 🤣

It's not a hack? 30 minute / $2M looks like a waste of brand fuel. He could have made 10x to 100x doing the same a bit more slowly.

Survivorship bias? (Reality grippers have left the market.)

Profit margins will be negative until all 401(k) have been emptied. It's the law.

America's Medvedev? (There's a theory that Medvedev's public persona is batshit insane because he doesn't want to be seen as suitable for the job, and thus be at risk of defenestration.)

Ultimately it may go back to the "stolen" 2020 election. Anyone who stayed on that train either had to believe it literally (and then why not believe everything else that came after, it's more of the same) or believe it's expedient for a great cause or other (and then so is everything else). 😐

It could be that structured data remains useful, but not the exact structure, as models "figure it out". It brings the cost of inventing a format and nomenclature per project (or per document even), and leaving it so forever, close to zero.

Copilot twisted the right way up in Excel may achieve that. Pretty it ain't, but it's low friction accessible, and it is already a stealth IDE for a whole bunch of programming languages: M, DAX, VB, Typescript ("Office Scripts"), LAMBDA, and classic spreadsheet formula programming.

Looks internally consistent with the "Venezuelan oil proceeds can only be used to buy US goods". So now China buys "more" US goods, instead of influence in an adversary with RMB or Chinese infra.

(The details on how it may achieve net zilch are not that relevant for the Show.)

Culture warriors are behaving predictably — in an "our side" vs "other side" incident, blame and virtue are automatically allocated. It's exactly the same the other way.

Wondering how weakly affiliated people ("non-warriors") will be reacting.

Safe non-generative UX enhancement could be a start, like show an incoming message as some sort of mood board with relevant snippets (paragraph from previous messages, material relevant to reply, etc) around it.

What do you think the lethality rate for the legacy system is?

(We all hate the idea that seeing medics is a lottery but then all the medics I know think seeing medics is a lottery.)

s/we/I/ 👾

Under common carrier, the duty is "delete on reports", so lot of legal nitpicking opportunities.

Social media is mostly real time so compliance may not prevent much of the impact of abuse, as the allowed processing delay is on the side of abusers.