sauwan

terra/sol or opus/haiku give an indication of size (the sun is bigger than the earth, opus is larger than a haiku). Carbon is smaller than argon (periodic table speaking), so that's no longer a helpful way of thinking about the model sizes...

Weird. That's definitely not my experience. On the $100 plan and I can burn my entire week's budget in a few hours easily with Astra. It's borderline unusable.

I think the point is that Yud and the rationalist community have put themselves in an awkward position where their lifestyle now depends on "p(doom) is big" being true. Which means they can't necessarily be trusted to give unbiased updates on the true level of risk we're facing as new data and information comes up.

Can "ASI will inevitably kill us all" be falsifiable before we create ASI? If not, then what separates it from religion?

Not sure how much I agree with that position, but that's just how I understood it.

I'm guessing they do this to help move people off old models they want to depreciate. It's not a price hike for old models, it's them removing a discount!

I think you're missing the counterfactual. I now have two choices: 1) sit behind my desk and work, or 2) walk and talk something out.

Before this model, the voice models were pretty dumb and annoying to work with. We'll see if this changes that.

Unless the US Gov bans inference companies from serving Chinese models to US customers...

Yeah, bummer. I was very excited for this release, but this killed it.

It's either more expensive or dumber.

>My wife and I had a really complicated spreadsheet for tracking how much we owed our babysitter – it was just complex enough to not really fit into a spreadsheet easily. I vibecoded a command line tool that's made it a lot easier.

Ok, please help me understand. Or is this more of a nanny?

As a Linux and Windows user I'm feeling unseen.

Well, the benefit here is the application of fluoride, which isn't something most people are actively thinking about maximizing.

For pure text responses, agree 100%. Gemini falls way short on tool/function calling, and it's not very token-efficient for those of us using the API. But if they can fix those two things or even just get them in the same ballpark like they did with flash and flash-lite, it would easily become my primary model.

Hah, I quit Prime and they just gave it back to me. No charge. I can't cancel it. I can't figure it out, but perhaps they realized that their margin on me before I canceled was well over the cost of Prime? I'm not sure, but I still only use it a fraction of what I used to...

Long running agents?

Yeah, Apple watches are a non-starter for me. Zero chance I'm charging my watch daily.

Does this use a local STT model, or is our data leaving our computer?

So, if you don't subscribe, do they pay you to offset the worse gas mileage your car has to carry around but never have access to?

I'm not a lawyer, but is this not a class action suit waiting to happen?

Presumably, AI has advanced in 10 years?

This is perfect timing for me - was just thinking about how to do this. But pricing is a bit steep for a startup currently looking to prove the market. Would you consider a cheaper option (e.g. 1 free inbox, or maybe $20/mo for 5 agent inboxes and a more limited storage level)? I'm building something that I might consider this for, but I don't know how long my runway is before I get sustainable client revenue, so $100/month is a deep hole being burned in my personal pocket before I can prove my MVP out.

Would be great to see how our previous months usage stacked up and when, if at all, we would have been rate limited.

I'd be pretty surprised if I were to get rate limited, but I do use it a fair amount and really have no feel for where I stand relative to other users. Am I in the top 5%? How should I know?

I liked calling him the "Ope Pope"

It's not cheaper than Deepseek V3.1, though, and Deepseek outperforms on nearly everything. And only between 1- 3x the throughput based on the openrouter metrics (near equivalent throughput if you use an FP8 quant). Wish I could be a little more excited about this one.

I'd be surprised if this was a new base model. It sounds like they just did some post-training RL tuning to make this version specifically stronger for coding, at the expense of other priorities.

I'd agree with you if that cut was 5, maybe 10%. I just don't understand how they can justify 30% outside of "you don't have any other choice".

I'm assuming this is true of all experimental models? That's not true with their models if you're on a paid tier though, correct?

Or Phi-4. Which seems to outperform it and be a smaller model from 4 months ago.

How do you stock Tamiflu? Isn't it prescription only? Or is it OTC in non-US countries?

I think this AI take is spot on.

Definitely just feels like a way to make it more painful/expensive for them to sell off their assets and become a for-profit entity.

Are you using the model to create a key-value pair to find/replace and then reverse to reanonymize, or are you using its outputs directly? If the latter, is it fast enough and reliable enough?