Helping early stage founders to Fortune 500s through product design. Currently building a developer tool.
Joined September 2009
- Tweets 4,847
- Following 453
- Followers 578
- Likes 127
Claude 4.5 is pretty good at UI Design. But you need to provide guidance and initial styling. The rest Claude will figure it out on the fly. Instead of asking to make it to be more modern or beautiful which ended up with purple gradients all over the place.
The current slowness of model inference is like internet speed when it first started, it was slow at the beginning and eventually a fast internet speed would become the new standard. I think this applies the same with model inference speed right now.
I think this is one of the best YC talk i've watched so far. It teaches you how to see the world from an entrepreneurial perspective. Great talk @Jacob_Heller
piped.video/watch?v=l0h3nAW1…
Windsurf new model SWE-1.5 and Cursor new model Composer 1 should trigger innovations for big lab like @OpenAI and @AnthropicAI to deliver SOTA fast frontier models for coding.
This is interesting if we take a look from the business perspective rather than engineering perspective. The secret of successful product is not determined by the code quality, but rather by how well the product really solves the user's problems.
"A lot of engineers think that code quality is important to building a successful product. The two have nothing to do with each other."
@blocks CTO
While I respect VS @code, but it really falls behind @cursor_ai, not just features even the performance is still inefficient. I think we can't mock Cursor for just a VS Code fork, they really bring innovations on top of VS Code.
Tried cursor new model composer 1 to fix an issue, it fast but failed to solve it after few tries. GPT 5 Codex fixed it in one shot.
why is @AnthropicAI extending the retirement date for Sonnet 3.5
Each AI models have their own vibes. Claude models are charm, playful, optimistic and feels friendly while GPT 5 Codex feels like serious, less talk and to the point.
Irfan Rosandi retweeted
If you're building a new product, make sure you can name a real human being who you think would be a customer.
Not a persona, a person.
It sounds so extremely obvious but it's so easy to get lost in the details of a "good idea" and stop evaluating it against this basic test.
When writing your prompt to AI Agent, be careful on using words that might produce ambiguous meaning like the word "here" to refer something, the agent could misinterpret that to something else. It's better to be specific on what you want the agent to change.
GPT 5 Codex can run 20, 30 or even 50 minutes for a simple feature. I think this model is more suitable for a feature that you have planned well, not experiment features you want to test how it might look like, the back and forth just too long compare to other models like Sonnet.
MashaAllah for every amazing thing I see, Alhamdulillah for everything Allah has given me to this day.
LLM can have their own interpretation of your prompt, if things are not working after first try, check if they understand your intention before executing. This could save time from fixing stuff or redo things again.
This is a dropdown on Apple website to select language proficiency, I think this is smart if they know what they are doing, but also potentially could create dishonest impressions for those who are genuinely select the last option. Or they simply want to sort it alphabetically.
I think one of the way to make app development with AI to be more cost efficient, starts the foundation with a good frontier model like GPT-5 Codex, and then after the foundation is set up, you can continue with cheaper model to add features.
Trying out Droid by @FactoryAI, the free trial is quite generous (thanks for this). Let see how it performs.