Andreas

Working with AI from dusk till dawn and then some.
Having 7 subscriptions as of now.
Feeling 10x as productive.

[Andreas @AndreasPandrea5]

@Sachin_scorp7 @ChrisGPT With exponential you might be right, but then we both might be wrong too. I find it hard to predict, it is just all unknowns.
If 2028.. well.. I think lots of people will struggle with that change.

[Andreas @AndreasPandrea5]

@kimmonismus @chetaslua So many demos I see in awe.
Then I prompt it and after 100 re-prompts I still got medicore results at best.
Not this model, just comparing X results to my results for previous releases.

[Andreas @AndreasPandrea5]

@Im_IrushiK 1 everyone and their dogs will leave OpenAI
2 go to Anthropic
3 Anthropic will have outages
4 Anthropic has to lower limits, Opus will deteriorate
5 Anthropic <-> OpenAI swap, go to 1

At least for me, OpenAI had the crown though for 9 months. Cudos for that. Now I love Opus.

[Andreas @AndreasPandrea5]

@rand_longevity Stupidity.

Then what the others said.

[Andreas @AndreasPandrea5]

@AndrewCurran_ Kind of like in the old movies, when a robot or a computer would explode when you made it divide by 0 or give illogical instructions?
Here the Austin Powers fembots
https://t.co/gkAyqhBUcn
Spoiler: they explode at the end.

[Andreas @AndreasPandrea5]

@superalesha IF the prices ever come down, as a consumer, remember that.
There is AMD too.
At least not as greedy.

[Andreas @AndreasPandrea5]

@cpscott16 Makes me wonder if - as part of the "deals" and bromance - the Russian bots will run in US datacenters soon.

[Andreas @AndreasPandrea5]

@deredleritt3r Why always either or?
I think most want the advance, in a safe way.
(It is hard to imagine rn, but then probably that happened before with progress.)

[Andreas @AndreasPandrea5]

@jun_song I think RSI + ASI takes time. No one can really say when you have it.
Ant + OAI (and possibly others) will still compete over months or even years, also giving time to catch up.
Also there are other metrics like alignment - if you have a great model you cannot release e. g.

[Andreas @AndreasPandrea5]

@thsottiaux Also: if you are out of compute, get rid of reset.
Everyone and their dog switches to Astra fast to burn their limits.
Give us high regular limits, in a fair, usable way.
You could (and should) also include web + dots in limits, and increase the overall limit.

[Andreas @AndreasPandrea5]

@thsottiaux People have not internalized that you are going to cut limits by 50% soon enough.
Also you never said that 5h limits would be generous enough. E. g. going through weekly limit = 5x5h sessions.

[Andreas @AndreasPandrea5]

@rationalaussie It is not just the best/most productive employees only.
It is also how to well one gets along. Some more metrics.
But yes, there are so many who know all the rights and none of the duties.
Who argue instead of listen.
Who (now) live in the past.

[Andreas @AndreasPandrea5]

@markjenney The summary: 15%+ for what is offered is just cheeky.
They should have stayed with reasonable prices and not become greedy.
Seems to happen with many.
But like you say, the value add is not that great.

[Andreas @AndreasPandrea5]

@thsottiaux 5h limit is fine as long as it is very generous.
The weekly limit is mostly bad. Though.. with the 6.1 speed it still feels infinite. Bu that is the slow speed.
On Astra it lasts a good day on any effort limits right now on 20x plan. That is not good at all.

[Andreas @AndreasPandrea5]

@scottstts Despite the model and usage topics (which @OpenAI should fix soon!)
Where have you missed the context? Did you feel conversations or thoughts could not be held in context?
Coz I have not, for quite a while, with both.

[Andreas @AndreasPandrea5]

@elshayib_ @elonmusk If true, looking forward.
(Honestly though, Grok sadly has been the most complained about subscription on X, at least in my feed. So hope you speak the truth.)

[Andreas @AndreasPandrea5]

@fanofaliens It had become so quiet with Ant and all the (bad) news was about OAI, I thought they had figured things out.
(They had the impossible alignment jump with Opus 5.5 though where the only explanation is the model knows when being watched and fakes double hard.)

[Andreas @AndreasPandrea5]

@SM80wx @oliviscusAI @grok I had to read @Grok s answer 2x. it is not token/s but s/token. Nice effort, but probably (???) as fast as using plain CPU?