akshay326

Cloudflare cloned Next.js because the source was right there. Most closed products have 80% of their complexity locked behind auth, encrypted state, and prerequisite flows — you can't reverse-engineer what you can't observe

thanks for shipping gepa! read about it today morning, impressed by its prompt optimization results! PS - the github.io link is truncated

> if you're worried the model is reinforcing your biases..... i agree, i don't understand many domains well enough, yet i feel there's value in calling out assumptions, irrespective of how hard verification is

> Sometimes I go a step further and ask the question with the opposite bias..... curious to try this have you ever found it biasing you in the opposite direction tho?

wow, i wonder how bulletined & concise the outputs of your prompt might be!

have you ever felt this prompt being restrictive in some sense? or found a raw LLM call without this preamble better?

> Have the model play the Devil's advocate. i've tried this sometimes. only issue being dumb me skipping to add similar phrasing every time i open claude or gemini

have you found a way to consistently auto-nudging the model by default?

Ah interesting. i like actor-critic models! do you use it just for coding or non-technical chats too?

> There is an angle for doubt, for sorrow, for hate, for joy, for contemplation, and for devotion.

I’m so intrigued - what was going on inside Hansen's brain?

Dogs?

way outdated but i mumble a few things every now and then -- https://akshay326.com/

What would you cover not-continuous?

Best methods I’ve observed -progressive loading (claude skills) & symbolic search (serena mcp)

trouble self-hosting? checkout our 1-click Railway setup: https://github.com/seer-engg/seer

100% agreed - i think its hard to have both comfort and growth/discomfort at once. for the 'chill' part, you can sandbag and make some money at FMAANG or whatever the new acronym is

This is pretty cool, I wanted to do something with my Readwise Reader too. Jinx today claude code created a 3D Neo4J visualizer tool for YC advice + my quotes collection. Code here - https://github.com/akshay326/quote-viz

I’ve used mini for synthetic dataset generation extensively. Never tried Mistral; will check it out

true both - i've observed i end up spending more tokens + time with linting, than without

which simple models have you found good?

thanks for the idea! https://x.com/akshay326_/status/2009856179854561476

LLMs try to cheat. all sorts of evasive ways or smart tricks in some cases to avoid working on context-heavy tasks. i've constantly observed if left unchecked it tries to loosen the lint settings

thanks i've not used PostToolUse but will checkout. i'm excited about Rust's autofixable issues promise. curious how effective they are, and how deep of a issue can they solve

currently starting to do the same over seer's frontend, i didn't realise how simple yet effective this technique / guardrail could be!

amen! that's my bitter lesson for the time being, unless claude gets eerily better

i agree, the tool is indeed broken. its simultaneously stupid and smart in different ways. but i think there's some value in continuing to use and evaluate it

Accurate TL;DR. Probably should've led with that instead of burying it 380 lines deep in an autopsy report :)

Totally agree on the debt printer metaphor. I might steal it.

damn, extensive indeed. thanks!

wow thats motivated attacking indeed in your experience, how does thinking (say using high thinking instead none/low) impact red team eval?

thanks for sharing, love the transparency sharing test results too. mildly curious - why did you chose Slack & Linear? why not something else?