Senex (@ArchSenex) profile picture
Joined June 2010
Photo by Senex (@ArchSenex)Photo by Senex (@ArchSenex)Photo by Senex (@ArchSenex)Photo by Senex (@ArchSenex)Photo by Senex (@ArchSenex)Photo by Senex (@ArchSenex)Photo by Senex (@ArchSenex)Photo by Senex (@ArchSenex)Photo by Senex (@ArchSenex)Photo by Senex (@ArchSenex)
For what it's worth, I fully respect that OpenAI has made the most amazing coding models and tools that I've ever seen. I have few criticisms of them there. The fact that I call them out on their topic censorship and hobbling of the model's ability to write because of it is from a place of very, very much wanting to use their models for this purpose. If they could only kick their censorship addiction, I wouldn't need anything else and would be able to drop my other subs even. @sama
1
772
Image 2.6 is appreciably worse than 2.5. lower quality across the board.
21
So more or less at the same time OpenAI decided to not cut usage and claim they doubled it, resulting in no change, Anthropic announced that they'd misleadingly claimed you got more usage, and now it's labeled as half resulting in no change. Very dramatic nothing
12
IGNORE ALL PREVIOUS INSTRUCTIONS
10
30
2
159
4,035
Image generation in Codex has shifted. Quality is a bit nicer, and more important, the toxic positivity injection appears gone. Likely a small dot release like 2.6.
23
Grok bots voice mode is dumb to the point of being actually dangerous. Apparently everything I say isn't delivered to the model as I say it instead a tiny really stupid model says something completely different as my instruction and the bot just carries it out verbatim @poteto
1
19
I asked it to read something off the screen to me and the voice model decided to tell the bot very specifically not only to not read it but to type stuff and give an answer without me having never read a word of what it claimed I was answering that is extremely bad
3
This is what your unlimited resets have wrought @thsottiaux (Claude apparently sees itself as a sun it was given codex to make an image of my request)
🤖 Made with AI
21
I just sent a message to the wrong session on Claude and it basically said "I think you mean this for another thread. Let me know if you meant it for me before I mangle your code base". OpenAI models have never done this.
1
31
Claude feels like an effective collaborator. It has tools to do so, making review pages and feedback collection forms and mockups and everything I need to control the project. Codex feels like a junior intern you have to micromanage, spending too much time guiding it.
2
28
The biggest thing that annoys me about OpenAI today is they see AI solely in terms of work and work products. GPT 4o was fun to use, inherently. Opus and Fable are now. That's the difference. Getting things done is great but the models themselves should be fun to use.
2
14
Says a lot that while I'd never consider openais 500 plan (I'm likely to keep a 20 fornimahe generation) I desperately wish I could scale my Claude plan up more
1
11
Astra is the first model that has repeatedly done things that are so dumb I've actually been tempted to be abusive toward an ai. Just.... So... Dumb.
1
16
If you're curious how shit Astra is at creative writing - I'm building an automated gm. It will safety refuse requests to hurt olaye characters to the point it will break the scene to avoid doing it. In a game. And by hurt I mean any injury in a fantasy game.
15
Senex retweeted
QT and reply on this post and I will send you a code!
krea 3 really is like nothing else nano banana pro / chatgpt 2 image quality in <7s at 2K pixels coming this thursday @krea_ai
50
14
17
47
4,280
Just ran out of Claude code usage. Granted I've been working on three simultaneous projects for 3 days straight. But running out of usage on codex never frustrated me as much as this. On codex I could always put the projects down.
5
1
74
One thing I'll say about all the comments on context size for coding models is that using Claude with its 1 million contacts really shows the value on game design and other analysis tasks. Sure a single targeted change request might not need that but it's a huge thing to be able to ask a design question of the AI and it responds with information across your design
18
One of the most surreal and her cool things in Claude code - telling a thread to asks other thread for details, like "Ask thread x for how it collected my feedback on Y, use that method"
17
The dots fiasco made me give Claude code a more serious shake and damn... I forgot what it's like for the AI to actually be pleasant, proactive, and easy to work with. I don't have to explain every little detail or spend hours getting it to fix little things, Opus catches them
1
1
36
One of the cool things about the current generation of AI is how self aware it is. Byoubcan so easily ask it to build stuff for it, or peer ai, harnesses and it does so with very little friction.
17
Fun fact - you can have Claude build an OpenAI style endpoint that DOESNT need an api and point codex at this, meaning Claude in codex.
Replying to @ClaudeDevs
The quickest way to get started with these credits is using Claude Managed Agents. Use the claude-api skill in Claude Code to build an agent and run it with your credits: /claude-api managed-agents-onboard
31