Week 26, 2026 - Cheap Claude Tokens, Interview Cheating, and Navayana

What I learned

• Hetzner, my current VPS provider has increased its server prices two times recently. Other community members are migrating to multiple other providers. One provider got my attention called Netcup. I found the price almost the same for me as of now for 8 vCPU, 16GB RAM VPS.
• My current fav agent assisted code editor, Conductor now supports OpenCode as a provider. Earlier it was supporting Codex, Claude and Cursor. Now, I can use OpenCode as my backup if I get out of tokens.
• Chinese AI labs are stealing Claude/OpenAI is not as simple as it looks after I read the post by Zilan Qian. As these models are blocked in China, Chinese access these model via proxy servers. The proxy server provides max accounts cheaply that’s why the Chinese models (like Qwen, GLM) needs to provide their subscription in a reduced rate to compete with these proxy servers. Proxy servers can provide cheaply due to following reasons:
  • Pool system: multiple end users for a single subscription
  • Model routing: they can say you Opus but may serve you Sonnet
  • The major one is Data harvesting (input prompts, output, context, everything.)

Use of AI Assistants and Cheating Apps in Interviews

• I have been a bit occupied this week with interviewing candidates for open positions in my team.
• Last year, also around same time I was taking interviews. This year something got changed. The candidates are using AI Assistants even to answer simple questions.
• Few candidates even use those during coding rounds. That is fine, but it makes my job a bit harder to judge them. The issue comes when they could not explain what have they coded.
• I have been drafting a blog post on how my interviewing style has changed due to these AI assistants. I have taken experince from other interviewers for this blog and it has a bit of more work than my recent posts. I will post next week.

Vedanta, Buddhism, and Navayana are not three answers to the same question

As you might have known that I have been exploring Buddhisim. I have mentioned about those in Week-14 and 23. I wanted to compare Vedanta philosophy from Hinduisim, early Buddhism, and Navayana Buddhism. My first instinct was to put them in a table. That is useful, but also dangerous. A table can flatten the thing that actually matters.

These three traditions are not three varieties of the same spiritual product. They begin from different anxieties.

• Vedanta asks: what is ultimately real behind this changing world?
• Early Buddhism asks: why does suffering arise, and what can be done about it?
• Nāgārjuna asks: what if even our deepest categories are empty of fixed essence?
• Navayana asks: what kind of Dhamma can make human beings free, equal, rational, and dignified in society?

That last sentence is where I had to correct myself. Navayana is not simply “Buddhism plus caste.” Caste is central because Ambedkar’s historical problem was caste. But the deeper move is bigger: Ambedkar turns Buddhism into a moral and social technology. The test is not whether a doctrine is ancient. The test is whether it produces liberty, equality, fraternity, reason, and dignity.

Religion must justify itself by the quality of human beings and social relations it produces.

What failed

• I tried to get Amazon SES access to built my self-hosted Newsletter service with the help of listmonk. Unfortunately, SES did not approve me for prod use.
• My open source contribution reduced drastically this week, due to leadership visit, interviews and other office works. Hopefully, next week it will be better.

What I built

Self-Hosted Server Fix

Last week I mentioned I could not access my home server. The fix landed this week. My Docker LXC container (100) refused to start with a status 19 error, but the real problem was upstream: one of the direct-attached drives briefly dropped off the bus, and because the mounts used x-systemd.automount, the host froze every filesystem query while waiting for a disk that was no longer answering. The container crashed only because its bind mounts pointed at those hung paths. I recovered without rebooting the hypervisor and then hardened the setup so a transient disk drop fails fast instead of hanging the kernel.

• Recovery: lazy-unmount the stuck paths (umount -l), confirm the drives reconnected (lsblk -f), systemctl daemon-reload, mount -a, then pct reboot 100.
• Root cause: x-systemd.automount is fine for network shares but hangs indefinitely on direct-attached drives during a transient disconnect.
• Hardening: add x-systemd.mount-timeout=10s in /etc/fstab so a dropped drive fails the mount fast, and check the physical SATA/USB/power path that caused the drop.
• Prevention: add host-level telemetry (Node Exporter) on disk I/O time and free space, so stale mounts alert before a container boot exposes them.

Multi-agent board workflow

I formalized a lightweight operating model for how two agents can coordinate through GitHub project Kanban board without stepping on each other. The rule is:

• issue = source of truth
• project status = workflow state
• agent labels = execution owner
• comments = audit trail
• linked PRs = implementation evidence

This is boring infrastructure, which is why it matters. Agents do not only need better reasoning. They need shared state, collision avoidance, handoff rules, and evidence trails. Otherwise they look productive while duplicating work. I had planned to do this since I started running multiple agents last month, finally I did it this week. But, I feel something is still missing, I will runby you on how this board is working next week or next to next week.

Tool Stack

As we surpassed half of the year, here is my current tool stack I have been using almost daily:

At the end of the year, I will publish this again. Let’s see which ones are getting changed.

Random Thoughts and Quotes

• There are so many people making their hands dirty and sharing their skills and worflow setup (some even charge for it). However, I feel none of these work full proof as they claim. The takeaway here is if you want to come in only when the water become clear it will be too late. May be this is the whole learning process which will make it easier to bridge the gap.
• I observed that a common man starts using a new technology faster if it can help them cheat the process and give them a shortcut. For e.g. porn industry, cheating in exams, cheating in interviews or financial transactions.