@CognitiveTake
Account based inIndia
About this account
- Account based in
- India
- Connected via
- India App Store
Account-level information from X, not a live location or the device used for a specific post.
Building the next generation of AI agents, but slowed down by the crippling AI psychosis
Joined May 2026
- Tweets495
- Following78
- Followers10
- Likes374
Everyone says marketing is harder than building. I will like to correct. Marketing a new idea is harder than building it. The initial push is so hard.
Once your idea is popular and has thousands of users. Building is sooo much harder than marketing.
Step fun plan is a great deal right now. But it’s sold out.
Hey @StepFun_ai when can I expect to try out the step plan. I got tempted by all the hype but seems I can’t sign up for the plan anymore? I would really appreciate it if you can sneak me in to the program somehow
Does anyone know any YouTubers or someone on X that reviews actual model capabilities and doesn’t just create three js scenes and calls that model comparison?
Seriously I am losing my cool every time I see these posts. It’s not good for my BP.
The AI subscription tiers are not built for average consumer. No one is going to pay 20-100 dollars a month for this. That means AI is only being built for businesses to accelerate their work not end customer directly. They will never see AI benefits directly.
This is unprecedented levels of transparency. Kudos to the Mimo team. mimo.xiaomi.com/rl/
This is really interesting. 77.4% accuracy sounds good on paper until you rerun the task. GPT-4.1 passed AppWorld tasks on all 5 runs only 53% of the time. I want Pass^k on leaderboards. Consistency in models is something people need to talk more huggingface.co/blog/ibm-rese…
Everyone’s watching demos of agents building a company in a weekend.
Mine started a second product before the first one had payments.
Then it asked for another API key.
Then it forgot why we started.
I don’t need a smarter model.
I need one that stays on the same damn thing
OpenAI took one Jalapeño kernel from 0.31% to 88.94% of theoretical performance in about 40 hours.
Thats crazy if true. spectrum.ieee.org/llms-for-c…
I am finally getting the motivation to stop building and release something and market it. Wish me luck!!
Anyone who thinks frontier AI is moving too fast needs to spend a week and build an actual production ready product. The stupid mistakes that Astra and Fable make are astounding and I refuse to believe that internal model has become so great all of a sudden.
Totally support this take. With what Dario is pushing for is clearly to create a world that is divided between the privileged and the people they rule with AI.
We need open source AI to stay available and thrive to kill the monopoly being proposed by these tyrants
People like this guy mask the real problem of the real people facing usage limits crunch. OpenAI has reduced limits but these posts get highlighted more. If you use up 600 dollars of sub in 3 days you are doing crazy things(both good and bad). It’s so easy to dismiss these posts
I have four ChatGPT Pro accounts. Three are completely out of GPT 6 Astra usage. The last one will be gone today.
$600 a month for three days of Codex.
I cannot buy more even if I wanted to. Pro subscriptions are closed.
This is ridiculous. @OpenAI give us a reset today. Your paying customers are locked out of the best model you have ever built.
How is your GPT 6 Astra usage looking?
It’s strange that labs specialising in smaller class models are able to achieve massive performance improvements as seen in GLM,Qwen, Deepseek flash models. But when frontier labs have to achieve massive performance gains, they go with building bigger models.
I think we can all agree that benchmarks shouldn’t be created by companies who are also selling tokens / hardware for tokens. There is an obvious conflict of interest and we can’t trust any numbers.
Companies running or building these benchmarks need to stay neutral.
