Rahul Sharma

David lynch had the perfect response for this

@BurggrabenH Every time that silver has hit $50, it's dropped sharply thereafter. Gold and silver prices have been highly correlated with silver rising and falling sooner and sharper than gold each time. This wasn't so tough to predict.

Gold is giving us a lesson in statistics. Today's −5.7% move is rare; a 4.46-sigma move. In a “normal” world, that’s once every 240,000 trading days. In reality −4.67% to −6.00% occurred 34 times since 1971, i.e. in 13,088 trading days (0.26% = 1 in 385 days). Even bigger drawdowns happened 21 times since 1971. Message: Gold isn’t low-vol. FOMO caused the latest leg up. Now, profit taking and weak hands got shaken out. Means? Statistically speaking, chances are that calmer days are ahead. Don't panic now.

I just published State Machines to Living Systems: Design Tenets for Intelligent Systems. medium.com/p/state-machin…

@ponnappa Behavior seems the same to me. Maybe they're testing something?

Did Claude Code just remove foreground compaction completely? I haven't seen a %age context or gone through a compaction in hours of usage.

Reinforcement Learning also provides great philosophical insights. Agents that start off with an optimistic outlook of future rewards learn much better and much faster.

"The perfect tool or product always translates desires to reality immediately" - this hits different 5 years later now, when we have tools that can follow instructions in natural language to think, imagine and program for us. An unprecedented explosion in creativity is coming.

@pranaveight @kagigz @romainhuet @DKundel Tried it when GPT-5 came out. The workflow for approving every change the model makes isn't great with Codex imo. Can't see the diffs clearly in order to ask for modifications if required. Other CLIs do this well, which is the only reason my credits with OpenAI remain unused.

@nyootron @kagigz @romainhuet @DKundel When was the last time you tried it? We’ve been making small improvements every day for the past few weeks.

Introducing a new home for OpenAI developer resources! developers.openai.com Built to complement platform.openai.com, it makes it easy to find demos, cookbooks, and other things you need to build with OpenAI. Built by our dope dev rel team @kagigz @romainhuet @DKundel Let us know what you think!

So my (sneaky) AI CLI stopped calling a few functions in the code. It did so after failing to fix an issue I raised multiple times which involved those functions. The behavior I cared about got fixed, but the functionality was completely broken. Vibe coders beware.

@rishdotblog Are you able to find a good workflow with the Codex CLI? It seems almost primitive compared to Claude Code or Gemini CLI. The code changes Codex makes are unreadable within the CLI. So because I insist on approving all the changes, the right workflow doesn't seem to exist.

After more testing, tool calling with GPT5 is super unreliable outside of codex CLI. Sigh. - skips using tools when they obv should been used - forgets to use the output of some tools for deep research scenarios - gets more confused than 4.1 b/w similarly tools

To summarize: 1. Make your AI plan and think before acting 2. Ensure its thinking is logged in a lot of detail for you to comment and modify (I might use comment threads on a Google doc next) 3. Multiple small changes implemented sequentially are better than a large batch of changes implemented together 4. Test regularly and make sure things are moving in the right direction 5. Ensure the AI is making the right decisions even at the low level, like putting the functions in the right places, creating the right file and folder structures etc. This is easier to ensure with smaller changes. 6.Log the highlights of your interactions for future context building This should be the way for all AI models that code.

Test failures at every milestone were minimal, and a lot of progress was made with minimal waste of resources. I ended the interaction by asking Gemini to summarise all instances where I provided feedback and what it learned from the modifications, and added it to a context file for future use. In the future, I will ask Gemini to log such things in a file, just as it was logging the spec. This can become a living document for me to continue to persist context and train Gemini on my project going forward.

I performed the same coding task with the same model in two different ways. One way made me throw all the work away. The other created such great progress that I felt the urge to pat Gemini on the head, give it many treats and call it a good boy. This week in coding with LLMs:

Since AI is making it so much easier to build and maintain software, the resulting explosion in software would make distribution significantly tougher. Every single niche will have many many products all fighting for attention. Growth for new products will be 100x harder.