redox99

My math takes that into account. It's 20x more expensive but you get 60x subsidy on subs. I.e you can do more tasks on a $20 claude sub than with $20 in deepseek API.

Opus 5.5 is according to AA cost per task, 23x more expensive than Deepseek with max effort (which is dumb, you shouldn't use) and 13x more expensive on xhigh.

Claude sub has a 60x subsidy, so a $20 sub gives you $1200 in API.

Based on this math, Claude sub is a much better deal than Deepseek. And it's a stronger model of course.

I have a missing Amazon order that says delivered but it was actually never shipped, you only need to check DHL tracking to verify this (and DHL customer support confirms Amazon never sent the order)

After multiple supervisors appeals etc Amazon won't refund me

I think I need to give these agents a try.

It's really not neck and neck with Opus 5.5.

~1700KG with 600L of nitrogen gas at 10,000 PSI

If you do the math that gets you about 13 seconds of hovering.

https://www.youtube.com/watch?v=22rvpUecLts

Having to waste 15 minutes with an AI is still crazy though.

Muse wonder is still benchmaxxed.

It performs worse than Muse Spark 1.3 and 1.2. And it has an outdated attention mechanism which at 1M tokens will need over 200GB just for the KV cache.

Very disappointing release. We're lucky to have Chinese models.

I never understood why people considered specialized hardware more secure than general purpose hardware.

A hardware wallet is way way way more likely to be backdoored[1] than an off the shelve Intel or AMD CPU.

[1] by people wanting to steal your crypto (unlike the NSA or Intel)

How does someone this dumb get hired at openai?

Maybe? But it's so many orders of magnitude apart. I can easily do 1B tokens a day (output tokens is probably like 3M?).

When I see non technical people use that Google reply I'd be surprised if it's more than 1k output tokens, and with a tiny model. So a 3000x in output tokens, on a model at least 10x as expensive, I wouldn't be surprised if hardware usage is 30.000x for a dev that uses coding subs vs someone who uses the Google AI overview a couple of times a day.

And that's not counting the absurd difference in cache reads.

> I bet they actually resolve more LLM requests than any other company at this point. The lay person still uses Google exclusively.

Unique users, maybe. But a simple query that's answered in a paragraph is negligible compared to devs running agents continuously for tens of hours.

Not for coding (antigravity).

https://x.com/OfficialLoganK/status/2110038940841545827

Only available in the $100 plan, and it's worse than Opus 5.5? Why would you ever buy this?

> Everyone made fun of Zuck when [...] forced out Yann LeCun, a Turing Award winner.

Really? You'd have to be pretty dumb to keep LeCun, someone who hates LLMs and keeps making claims that end up being proven false in matter of months.

I'd consider that solved with LLMs (the capability is already there and will only improve). The only remaining hurdle I'd say is hardware attestation.

Number of employed SWEs has maintained the same, slightly increased in fact.

For now that is balanced by the fact that we build more software, things that wouldn't be worth it before.

You can migrate off deno in a single day. It's not a big deal.

Pro has a specific limit of weekly messages (50 for the $100 and 100 for $200 I believe)

The rest is basically unlimited unless you're doing something very weird.

Meanwhile I cant ask a simple question to Claude, not even to Haiku, because I used my claude code 5h limit.

The training data back then was sooooo bad. You'd open something at random in Common Crawl and it was truly unbelievable that models (like llama) were able to output coherent stuff. The architecture was truly miraculous if it could train well with terabytes of garbage.

There was so much low hanging fruit back then in terms of training.

Interesting. Since GPT4 came out I strongly believed LLMs would replace human programming. By November 2025 when opus 4.5 came out I thought it was pretty established already and most people agreed with that premise.

By human programming I mean humans typing code.

The stochastic parrots have predicted the next token once again.

> Being a doctor is one of the jobs expected to be least at risk from AI replacement

Is it? At least as a customer I'd love to replace a doctor with AI. Humans suck, you have to wait so long for an appointment, they spend like 10 minutes on your case, and often have very outdated knowledge. The main issue is that vision still sucks and lack of embodiment is a problem. It could invoke "human subagents" for that.

Yeah but they are already RL'd too. No doubt you can further improve an RL'd model by doing your own better RL on top of that. But it's not going to be the same as starting with a base model or instruction tuned model.

I think there aren't any recent base models to post train on. Labs don't release them any more.

I love local models, but the AFM advanced or whatever its called (the 20B local one) sucks. And it's not really used in MacOS outside of "fm" and very niche stuff. Siri is always cloud even if you're offline.

My problem with plugins is that I don't want to install stuff from third parties. It might be malware, become unsupported, or generally have a lower quality. I rather have a curated and polished feature set out of the box.

It's nice for having your AI create your own plugins. But I haven't really found the need yet.

Yes, although you probably want to calibrate your model if you want the probabilities to actually be meaningful.