You can actually get just as good results for packaged food if you just allow the agent to use a web search tool & keep a growing local database of "known food items". I did the same for a while, but it was quite expensive (this was a while back with less powerful models and required a bit of orchestration to make it reliable)

Presumably it would Just Work if I wrote "1/2 bag cool ranch doritos", which is one of the virtues of this input/processing system.

Since I discovered and started reading Lesswrong a couple years ago, I keep seeing dismissive or sneering comments across the (technical) internet, usually with very little substance - like this one (what does it even mean? I find the article rather reasonable and on point). I know that yudkowski's more recent stance is controversial (I haven't started reading his work on how AGI will kill us all so I'm suspending judgement ), but is there a valid reason for the widespread smugness or is it just a bandwagon effect?

How do the economics of your statement work out? Clearly inference providers don't have a time to ROI of 10 years on their hardware costs; and that's without even taking ongoing energy costs into account. What's missing here?

Output tokens are actually kinda expensive for the provider.

The input cache hit tokens are incredibly cheap for them, (incredibly high margin too, except for deepseek).

And input tokens are in the middle. Input tokens can be processed very efficiently.

Also his math is wrong. $100k gets you 22.7B output tokens at $4.4/M which is how much GLM 5.2 costs.

At 500/s 22.7B is just 500 days. Or about 1.54 years. Which is much less then the life of the hardware.

The inference providers are running batch sizes much larger than 10

Inference providers have been getting a firehose of investor cash to keep the chips running (and are looking around very nervously as that firehose starts to sputter).

I very begrudgingly started paying for grok for this exact reason. They nailed the voice UI and it works incredibly well with android auto unlike Claude&Gemini (which don't work with android auto at all) and chatgpt (which works well but has hardcored system instructions that make it's voice mode feel like a dopamine deprived Gen Z)

Reddit absolutely has algorithmic feeds since it ipo'ed (maybe earlier but I used third party apps so I wasn't subjected to them). 90% of my home page is bullshit I didn't ask for.

Yeah there is argument that reddit wasn't social media, but currently is trying to be

What does "ai in production" even mean? Writing production code ? Depending on how it gets reviewed and the qa mechanisms in place, could be stupid or not. Read only access to production systems and data? Again it depends on safeguards, but probably not stupid, it can be very useful for debugging. Unsupervised write access to production data or infrastructure? Incredibly stupid, but I dont think anyone serious does this.

Why not?

Aside from display quality I wonder if humans are more perceptive to changes in specific colors? Towards the end of the game some examples felt impossible while others were trivial .

Thanks! You mean to be able to search for movies that occurred in a specific set of locations? That's a really cool idea, I hadn't thought about it. The data is there already; it's just not searchable in that way. I might give it a go next week. I need to think of a nice UX for it.

Thank you! I'm fixing some early feedback about the UX and I'll do just that!