I find Anthropic's latest gen models to be the easiest to understand. They did a complete 180 from previous gen and responded to feedback. Great job!

Software is like a heart that pumps blood. It has to be made to pump, always. Else it is dead. I used to think, why keep building/ maintaining software always, why not build once, set it and forget it. I have seen that this fact is true for both B2B & consumer apps. New features, issues, independent or driven by change in the course of the external world (trends, new tech, cybersecurity incidents etc). Other than that, you need constant touch points with your users. Software buyers need to be explained that this is a life-long commitment, just like a car has life-long upkeep & maintenance. This pain is worth it because of the value it offers vs. the alternatives. For this reason, you cannot or should not have one-man dependent software. This key man risk is real and boutique consultants should work around this, lessen it as much as possible. Small team 2-3 guys is fine. Just not one person.

I saw a post from @trq212 a few weeks ago, that they are planning to remove plan mode. In this post, I emphasise the need of Plan Mode, while copiloting coding agents. The way I work, I call it 'cooking the context window'. I cook the context window by asking questions, making hypothesis/assumptions, let the agent deploy sub-agents to read documentation, source code, forum discussions, search the internet etc. Plan mode is just a handy way to make sure the right context goes to my fresh LLM. I could implement from the same agent window, but I do not know what context is still there in its window. It may have fetched/grepped some stuff, that may have been relevant for exploring, but not for implementation. The more 'context-full' for LLM is, the worse it gets. This may also allow some previous 'short-cuts' and anti-patterns to seep in the agent. Plan mode has always been very underrated, and the 80/20 feature of agentic coding. I also use it just to brain dump in voice-mode. This allows me to have a head start of all features or issues I have to work on. I can just dump my worries on it, one agent at a time, just before leaving my home, or closing my day, then deal with it when I start my work session. You need to understand that agents have a tendency to write short-cuts or anti-patterns, because they are not aware of their coding skills. LLMs are trained on data outputted by humans, and it would make sense for humans to go with a half-a-day solution, a quick-fix, and band-aid, vs the grunt work solution of the core problem. Rules and skills could get 'lost in the middle', especially after the chat session being cooked or overcooked. You cannot let this shortcut behaviour seep in unmonitored. It happens with the best LLM agents with the best environments. However, it comes down to human judgement, when to spend time on this extra step, vs just go for it. Small things, obviously may not need this much effort. Rather Plan Moding first thing, with a hard to understand model (even after easy language skills), may create an overwhelming thesis, too hard for the human to read and understand. At the end, I would like to say, use plan mode but also not use it. Thank you.

Don’t automate what’s not been done by hand.

I finally watched @poteto 's pinned talk on agentic coding. What I understood in one line: "Make every environment behave like Rust.

Interesting to see Anthony Morris & Thariq, both leave Anthropic just before IPO. Why not stay some weeks longer and take a ‘proper exit’? For me, witnessing such an IPO from the inside, and the aftermath is too rare to let go.

Okay, gunpowder smelling, burning sensation air pollution is here in Delhi. Take care. 😷 What AI & Mars bro, solve these problems first.

I never used an AI browser. Did it make those who use it super rich or super happy?

I am going to do this for my neighbourhood and be cool in-front of all uncles. If you have no idea, it’d be like navier-stokes for uncles.

If you ask the right questions, I am telling you, you are already halfway there.

Sad to see Railway go down tbh, I liked their vibe.

Grok 4.7 is the least sycophantic, had a feeling, BUT, it is also the hardest to understand. I find myself going back to 4.6 for answers, cause it is direct. I believe the way gradient descent works, pretty much very difficult not to overindex on something.

revised_list2_final.pdf 8. GitLab 9. Railway

Revised list of cybersecurity issues I saw in one night. 1. eMudhra 2. PHP 3. LiteLLM supply chain on Jfrog 4. State Bank of India data leak 5. Rajasthan gov citizen data up for sale 6. PingCAP / TiDB Cloud 7. ⁠iPhone 11 to 16 remote access

Having an optimised X algo is so valuable. I got to be cool to my semi tech clients telling them about the latest cybersecurity flaws before anyone else.

6. PingCAP / TiDB Cloud

List of all cybersecurity flaws I saw today as an Indian. 1. eMudhra 2. PHP 3. LiteLLM supply chain on Jfrog 4. State Bank of India data leak 5. Rajasthan gov citizen data up for sale