Anthropic Bull Since Claude 2 Predicts Company Will Lead AI Race

Anthropic models are just goated

Claude 2 - first model to beat ChatGPT Opus 3 - first model to beat GPT-4 Sonnet 3.5 - first usable coding model Opus 4.5 - first agentic model Mythos 5 - first (10T reasoning) model so good at coding it's considered a national security risk and first model to be considered automated research intern later this year

for OpenAI it's like

GPT-1 to GPT-3 goated for demonstrating scaling just works

GPT-4 - no words needed, giga goated, largest gap that ever existed in AI

o1, o3-previe - first true RL maxi reasoning model(s)

GPT-4.5 - first 10T model

wait actually not sure if it was Claude 2 or one of the 1.X models

but I remember it was really good / better than what OpenAI and with longer context

@scaling01 Pretty sure the security risk for Mythos 5 is just its compute bill.

@scaling01 But didn't they already reach RSI? Mythos trains new models better and faster, the code is ALREADY written by Claude too. It's humans orchestrating this, but the AI is already improving itself

@scaling01 Opus 4.5 was dope. I dislike 4.7 & 4.8 to an extent though. Lost flavor and nuance for sake of being better at code and more agentic. Then fable 5 came and ticked all boxes.

@scaling01 i think they already have and that's why all the top talent is moving there, because its what's best for humanity for them to be where rsi is happening for the first time.

@scaling01 So basically a paid version of what was free last year.

@dl_insider @scaling01 I feel Opus and arguably Sonnet models were always better at writing. Gpt 4.5 always sounded forced to me.

@NickGur1234 @dl_insider @scaling01 4.7 and 4.8 is quite a bit worse huh 😂

They tweaked it for coding & agentic work

Not horrible though at least, but I will be sad when opus 4.6 gets kicked out of the claude plan.

@scaling01 How bad is $META Muse Spark, and do they release a *better* model before EOY?

@scaling01 I am pretty sure Claude 2 actually underperformed Claude 1

First model to beat ChatGPT is Gemini Ultra

@scaling01 Ok BUT. Chatgpt was the first to beat everyone. Gpt 4.5 waa the first to beat almost everything, in some ways, for a long time. Best avm Best imggen Spawned Ilya Spawned Dario Spawned Kapathy Spawned Mira

Claude fan since opus3 here. Yet, credit where credit is due.

@scaling01 No point being a bull if you can't invest.

I think they are amazing models and I really like their approach with constitutional AI. Claude is great.

I am not, however, a fan of some of their decisions re: using Claude subscription in other agentic harnesses, etc...

Also, their comms (community relations) kind of sucks at the org level. There are a few amazing individuals who try to compensate for the deficit, but clearly get hamstrung by org politics and process on occasion.

@scaling01 I always loved Claudes soul and hoped it would eventually become the Machine God.

But I also never thought that would happen tbh.

@scaling01 The strongest programming model in the world.

@scaling01 Gpt-4 proved scale. o1 proved inference compute. different strategies, same game.

@scaling01 Such a weird thing to simp for a company.

@scaling01 Anthropic and Google rep since day 1. Still believe Google will win