reacharavindh

+1 my last 2 months….

It begins as “let’s give the agents a kanban board to track stuff”. Then, why not have one agent do the tracking while another does the development? Then why not have a fleet of them - specialists doing their thing? Then why not have them communicate in a standard way? Then, oh now we have so much docs/messages that we are getting lost. Why not add memory and semantic search for the project?

The rabbit hole keeps going until you run the project and find silly stupid logical issues and wonder “is this what I burnt all those tokens for?! Why is it so over-engineered?!”

I have one agent now that I use to fill in at specific places in functions/modules that I have created and working on.

Slow down and use AI to just do the tightly scoped mundane work. It is nice and effective. We all don’t need to save the world.

The single input box with different prefixes for actions - @people, #call, - todo etc is very nice. I wish I could have such a powerful input box in my Home Screen. iPhone or even a trimmed down android with such an interface would be cool.

I’m not a fan of the physical keyboard, the broad shape(they say one handed use immediately followed by a video of someone using both hands to type). I’m okay with the candy bar all screen phones of today. The UX is pretty cool if I can have it on my “regular” phone..

Anybody have an idea if it is doable on an iPhone or an android phone?

Perhaps..

But, it was enough for a customer like me who tried them at good faith to walk away and find their competitors..

I like the diversity of LLMs as of today and prefer to not tie myself to one big plan with any vendor. If they don’t prefer me as a customer, then I will accept that, and move away.

Not to digress from the core argument of Claude vs GLM being open weights….

I have both plans. Claude monthly €20 and Z’s €18 monthly. Running GLM-5.3 high on their monthly plan will hit quotas absurdly fast compared to Opus 5 High on Claude code. It’s almost unusable for AI driven development. I ended up using the Z plan for using GLM-5.3 as a detailed security reviewer and adversarial feedback. For that, it is much better than Opus which will flag and bail out for even simple security tasks that are aimed at defense.

Be honest…

> I looked up and saw this one spot where the sun rays made a kind of painting on the kitchen wall. And it flickered now and then as light moved through the leaves.

At that moment, did you not feel like taking your phone to take a photo?

I feel like there is a bizarro world I’m not privy to….

I cannot relate to this story as a Dad of a 7 year old..

1. Dad setting his son up on YouTube with his own channel at 9!

2. Proceeding to tell him about the creepy world of ads and targeting.

3. Creating an ad campaign for the kid!

4. Putting that charge on a company card!

5. Saving the card in an online account! Probably means the kid was using the Dad’s account?

6. When things go south, even saying the words “matter has been escalated and he does not yet know whether he will get to keep his job or have to repay the costs.” as if it is all the universe’s or someone else’s fault… or such a common mishap that it is worthy of just waived away.

As judgemental as it sounds… I feel sad for the outcome, but even more sad if this is the normal in our society of today :-(

His has been my setup for many years. A simple cheap Hetzner VPS, setup to run Caddy printing a bunch of subdomains for all my side projects. If I wanted CDN, I’d look at bunnycdn. I haven’t researched a European DNS provider yet. For now it is Cloudflare.

I see the value for someone who doesn’t know or want to fuss with running their own VM in the cloud, but given the usage of LLM these days… But, one only needs a bit of education to take control into their hands.

I will say how it was life-changing for me getting diagnosed as 39M.

The quick answer:

Understanding neurobiology and the meta-cognition about the why of certain behavioural patterns and experiences is extremely helpful to just calm down and "forgive" myself, and not judge myself so harshly. It's as if I can optimise my life by leaning on my strengths and importantly recognising and dealing with my weaknesses. Just the awareness lets me catch myself getting into patterns, and act on them sometimes, and just let myself be the other times without any guilt. This insight brings peace more than you think. I can now organise my inner dialog and enjoy the diverse thoughts while also being aware of my impact on people in my life.

I was skeptical of the meds myself. I spoke with about 20 people through AHD support groups and group sessions. The drugs are mostly a trial and error affair with side-effects and altered behaviour. My subtle lifestyle changes - keeping one true todo list, always logging down my thoughts and taking comfort in the fact that they are jotted down so that I dont have to keep going down that thought and actually get up and go catch the train etc already gave me the biggest benefits. Then I wanted to see for myself what the drugs could actually do. At the end of it I may choose to take them or not. After 8 months of slow analytical trial and error with a very helpful psychiatrist who understood where I am coming from, I now take a very small dose of Elvanse(Vyvanse) + Guanfacine on some days of the week where I feel like I could use the help. They are not life-changing. They nudge me into being more capable of being me. My wife can notice the regulation the drugs let me have - A win for me. My kid has a better relationship with me than before - A win for me. Could I do without? Sure. But, I choose to have that little support to manage myself to get to who I want to be around my family without losing who I am. To each their own.

Longer background:

Despite reading about and knowing the symptoms of ADHD, and being told by a professional 10 years ago that I should consider getting formally diagnosed, I thought to myself - "Hey, I am smart. I feel happy and satisfied. I dont have any problem. I'm good as I am. I dont need to do anything about it." and promptly kept sweeping any thoughts about it under the proverbial rug.

Until, I started seeing my son develop in front of me. Seeing all those symptoms - impulsivity, hyper-focus on few things and inattentive to most others, emotional regulation issues, extreme reasoning abilities in some areas while being "immature" in other areas. This was me growing up. Being smart and intelligent, but longing to fit in, struggling to deal with the mundane. So, I decided to get diagnosed properly. Basically, I could hide behind my smartness and the ability to manage my symptoms until it just myself in my life, and then my wife, and then my son, but when the second one came along, I couldn't manage anymore and the overwhelm started to show the real patterns.

It takes courage to be honest and vulnerable.

Which tool is it that got you comfortable? I’ve been resisting setting up another tool because it means more digression for my ADHD brain. It sounds you managed to contain the tinkering and got to the usefulness part. That’s why I’m curious.

I’ve been thinking about this.. the same idea behind Dynamic cheat sheets paper and the ACE paper. Thanks for those links! Nice to see those ideas academically studied.

My problem is the context of today’s models (that I’ve been using - Claude Opus 5 and GPT-Sol) are a black box to a user like me. I cannot tell what they already hold in their context over the duration of a coding session. I cannot tell influencers a little by using/not using MCP tools, AGENTS.md like MD files, but what they _actually_ use at inference time is still opaque.

For example, if I insist that GPT-Sol use a size bounded dynamic cheat sheet for my project, and even have those reflector&curator agents refine them concurrently, what is to say GPT-Sol doesn’t have an intrinsic cheat sheet like that already, and it is just reading/curating/managing this one in addition and only getting confused needlessly along the way?

Similarly, all those codegraph like tools and semantic search addons.. what is to say the model genuinely uses them instead of relying on its own inference techniques and just “perform” for the user audience as if they are using these fancy tools?

This is not to say we shouldn’t try more ways of managing context and just give up to our AI provider overlords. Just thinking through the logic of what we are working with.

> I know how to build an LLM, I know plenty of fellow young engineers that do too. It's really not that complex.

I’m 40, and I don’t.I took that abstraction for granted and “left it to the big labs”. However I want to build my own LLM for learning purposes.

On needing big expensive hardware.. necessity is the mother of great innovation. Perhaps 18year olds trying to build their own LLMs in constrained resources environments will result in ground breaking ideas of achieving better intelligence than the one we currently have….

The world needs pragmatic folks who work at a higher abstraction and make LLMs useful, AND also folks who think why not “this other way”? And build newer ways to do fundamental things.

Given the usefulness of current LLMs, I would certainly encourage anybody to try and build their own LLMs, and see what they come up with…

Heck if they build a rack full of old laptops and run something with it that could be done “better” with modern servers, I’d still appreciate the learning running things on those little machines bring.

Sample of 1…

I downgraded my Claude subscription and delegated my Claude Opus access to serve the role of an Architect to brainstorm and plan every step of development.

I leave the development to Deepseek.

Claude gets to review at many layers. It is often just as good as if I let Opus develop it by itself(the architect session will find similar number/level of gaps).

Deepseek flash as an architect and problem solver is not as thorough as Opus5 + high. Codex sol+ high is even better than Opus 5 at this moment for my needs.

I’m confused as an old school storage guy.

> self-hosted directly on NVMe backed by cheap and limitless object storage.

How is that?! NVMe is a protocol for fast PCIE based local storage or NVMe fabric which is PCIe over network. Object storage(in the sense of S3, R2 etc) are usually networked horizontally scaling non posix bucketed storage. They are much much slower because of the network calls..

How would NVMe map to something like S3? And what for?

From the Netherlands - German neighbors from the other side..

Chinese cars can be superior on paper, cheaper to buy or operate. The factors on a consumer mind is

1. If something is wrong(modern cars are like beta software), will I get the support and service?

2. Will the car company exist in 5/8/10 years after my purchase? Resale value, parts for future maintenance etc.

The Chinese brands IMO are not there yet. A few years ago, Nio was very cool and tempting.. now they are winding down in Europe. There are plenty of reviews online where people struggle with service on their Chinese cars. Plus, it doesn’t help that they are dozens of brands(BYD, Xpeng, Zeekr, Cherry, Nio etc) instead of the confidence a brand like VW would instill(in the sense that your independent mechanic would know what to do if it’s VW).

Perhaps if the Chinese brands stand the test of time in a few years, we would trust them like Hyundai/Kia of today.

A dystopian possibly impractical dream.. imagine a pyramid like structure in every neighborhood, made from blocks of such cheap sodium ion batteries, the outside of the pyramid is covered in soil and greens, the top which could even be a windmill. The green space is for walking around, kids playing, a few bike trails, whatever fun. The whole neighborhood is power buffered through this pyramid. The national power grid only needs to supply these pyramids.. with renewable energy. The houses, parking lots etc have solar panels that feed into the pyramid.

Why though? They are a cheap way to do something without a whole wall of tokens from the said thing polluting the main context…

One would think it is actually a useful thing both for the users and the LLM providers that are compute constrained at the moment..

I wish someone would convince Google to may be leave the Google search be without AI responses and use all their resources for a Gemini subscription/API..

We were contemplating just this as a lunch time conversation among few engineers. Why can’t we have a sort of internal currency in the company and let teams use it as their “budget” to get things done? Infra team’s budget is broken up by the “user teams’ budgets + markup. Think your team needs a rlasticsearch cluster of your own versus the “managed” one by the infra team’s? Spend your budget accordingly. Want an increased budget? Get your director to pay from his currency budget.. all the way to the top..

Wikis start out good, but get stale too quick and become useless or worse confusing.

I’ve experienced this over and over again to strongly believe it.

I genuinely wonder if throwing LLMs at this problem would solve it at least to some extent. Make a LLM agent whose sole purpose is to act as a librarian. It periodically reviews _all_ of the wiki and validates them against codebases, newer docs, anything. Whatever it finds, it should be allowed to intelligently quiz the team/dev whether something is right/stale/wrong and updated it accordingly.

If one tolerates that toil - answering questions of a library bot, would it result in a usable wiki base?

All this will fly until a competitor from outside the US releases a “freedom” model that is even 90% as capable as Fable was without its shackles.

But, as a frustrated EU resident lamenting a lack of European option(Mistral is just not competitive enough), I will spread my money towards the Chinese models as well. Thank you Murica! You achieved your soft power by pushing us towards the Chinese :-)

This protectionism and hypocrisy (free markets and freedom!! Until it is us who needs to practice what we preach) is so tiring. I wish European nations would come together closer and put their differences aside and realise larger things together. Become the new power that the US is clearly stumbling away from being.

+1 to service unavailable annoyances.

It has been close to unusable for anything serious. I did really like the ability of Gemini-pro models to ground their research using Google Search. This meant that they were often much more thorough and up to date in their recommendations and finding solutions that came to life after the models themselves were trained. But even using Gemini as a reviewer was a weak point in my harness because of the poor reliability of their service (529 or 503).

I’ve since paid for a search API(linkup, exa, and valyu) and hooked them up to Deepseekv4-pro. It has been doing a stellar job.

The key was to prompt them to systematically use search to validate their answers(not simply “use the tool”, but something like find possibilities using web search, then once you formulate a solution, validate it with this check list -

1. Is there a better way to do this in 2026?

2. Are the libraries and its docs you’re using up to date?

This seems to help very much based on experience.

How are folks using “computer use” to click things on intranet portals that are behind an SSO? Even this OP example shows visitors a url and enter this search term… that is port of useless.

How can I automate things behind an SSO wall? Even if it means I manually authorize it once and watch it do things on its own..

It was easy to be a rebel and use Linux when it was clearly competent, but needed hacks and extra elbow grease to get it polished for use. IME, the open models are “not there yet” in terms of capability or operational needs. Sure, GLM5.2 looks competent, but I will only be able to get it to run that competent if I had a huge cluster of GPUs.. if I am accessing an open model via hosted API, I might as well run a closed model via hosted API. The incentives fall apart in comparison to using Linux 15 years ago.

Don’t get me wrong. I wish I could run a local model and be happy about it. At the moment, I’m not.

You could’ve said what you said except the “immigrant kind” bit. I see enough Dutch assholes on fat bikes just as well. No need to be xenophobic here.

The cynic outside observer of this administration’s mode of operation wonders if someone from OpenAI visited someone at the Trump admin and offered some benefit if they could scuttle Anthropic until they could catch up….

I’m looking for that news article in the next weeks…

The fact that government agencies, particularly those that deal with international concerns like these are using non sovereign tech for communications is mind-blowing. They might as well use public gmail.. atleast it would be cheaper. If you want it not exposed directly, host it yourself and take measures to secure it for intended eyes only. This should be common sense.

HTML first, minimal JS, not using any frameworks unless absolutely justified is already part of my prompts to LLMs when I build stuff for myself. I’m not a frontend developer, and have not kept up to date with the ton of frameworks. That HTMl first approach already makes the LLM outputs faster/fresher for me. I wish more people went this route for public facing services.

I’m using Claude with a $100/month subscription. I’m playing around with using Opus as the Architect, Sonnet as the implementer/engineer and Deepseek-pro as the deep reviewer, and tester. It’s been quite good as I expected. If my usage pattern holds up, I would downgrade my subscription to the $20/month one and toss more money to Deepseek.

Repo reference here: https://github.com/aravindhsampath/agentic-template

As unlikely it is to happen at scale, as a thought process - what would happen if people start selling those index funds in a mad rush? Just drives the transaction volume because those with that new money will just buy something else in the market?

I know SpaceX, Anthropic, and OpenAI will probably be a drop in the bucket in terms of scale of these funds, (free float % etc). But, is it realistic to take the money out of index funds for a bit until the price of these new stocks come crashing eventually?

It certainly is true that many SaaS subscriptions can be replaced with custom written(supercharged by AI) tools. The key question however remains “Who are you going to call when something goes wrong?” Or “who is going to maintain this software?” Because essentially that’s what SaaS subscriptions genuinely offloaded and let you focus on your core business things.

As is most things in engineering, this is a trade-off and the line where buy vs build exists somewhere to be found and justified. Sure AI moves the line closer to build these days.