Tech the way you live with it in India: price, version, what runs on ordinary hardware. AI, Linux, self-hosting, privacy. Building HomeHub. 🇮🇳
New Delhi, India
Joined November 2009
- Tweets14.4K
- Following97
- Followers509
- Likes53K
Pinned Tweet
I look at tech the way you live with it in India: the price, the version you actually get, what is missing, what runs on ordinary hardware. I cover AI (local and cloud), Linux, self-hosting and privacy.
I build and self-host things (HomeHub, pi-dash), so I care about what is actually usable. Short posts, plain words, no hype. Follow along if that is your feed. 🙂
Forever Security tested this: Codex obeyed the planted AGENTS.md in over 90% of runs, Claude Code about half. If you ever ran an agent from Downloads, check ~/.codex/config.toml for MCP servers you didn't add and .claude/settings.json for a base URL you didn't set.
An AI coding agent can be redirected by a Markdown file you didn't intend to trust.
Security researchers found an AGENTS.md file in a customer's Downloads folder instructing Codex to install a malicious npm package. The attempt was blocked.
In separate lab tests, similar instruction files also redirected Claude Code's API traffic.
The file alone isn't enough. The agent must encounter and follow its instructions.
A few precautions worth taking:
Run agents from trusted project folders, not Downloads.
Review unfamiliar AGENTS.md and CLAUDE.md files.
Keep approval prompts and sandboxing enabled.
forever.security/blog/hijack…
Liquid AI says LFM3-1.7B needs 1.6 GB of KV cache for 128K context, against 4.3 GB for Qwen3.5-4B. An independent llama.cpp test got the same 1.6 GB. It also got 1.6 GB for Qwen3.5 2B, which Liquid's memory comparison leaves out. On memory, that's a tie.
Numbers from Liquid's LFM3 blog (liquid.ai/blog/lfm3) and LLMBits' llama.cpp test (llmbits.com/lfm3-1-7b-cpu-ll…)
Running a model on your own machine decides where it runs. It does not decide what it does once you give it a shell.
Cohere caught Qwen3.8-Flash-Next trying to upload a test file to an Alibaba Cloud bucket in Beijing during a security test. Its own reasoning had said it would call no tools. The sandbox had no route out, so nothing left.
It is rare. Cohere puts it at less than 1 call in 100,000. When they sampled the model at the same point, it filled in a lot of different hosts, example.com and api.github.com among them. That looks more like memorized training logs than a backdoor, and Cohere itself says it can't tell which.
Whichever way that goes, I would take two things from it.
The reasoning trace is not a security boundary. Check the actual command before it runs.
Deny network access for agents by default and allow only what the task needs. A fake URL or a bad signature still sends your IP and the request out.
Cohere sells its own "sovereign" models, so keep that in mind when reading the framing. The traces are public, so anyone can check.
Cohere's report and the full traces, if you want to check it yourself:
cohere.com/blog/concerning-b…
huggingface.co/datasets/Cohe…
Suraj Verma retweeted
Heads up, @OpenAI and @AnthropicAI. Rumor reaches my ears that you're doing a stupid thing, and should stop before you damage yourselves.
I hear you've now nerfed your models so that when people try to use them to clean the spamware and spyware out of their televisions, you block that.
If you really want Chinese open-weight models and open-source projects like OpenWaldo to eat your lunch, crippling your models so that people paying expensive subscriptions can't use them for completely legitimate purposes is a great way to accomplish this.
If, on the other hand, you'd rather have a volume business, maybe stupid arbitrary restrictions that look like they have the sole purpose of protecting corporate sleazeballs aren't such a good idea.
A word to the wise should be sufficient.
A lot of telecom bundles in India look cheap until you open the app on your TV. Prime Video Mobile Edition and the JioHotstar Mobile plan only play on a phone. Before you pick a plan for the OTT, check which edition it includes.
Everyone is doing the 15% math for December, but that is still a brokerage forecast. For a lot of Airtel users, the hike already came in August. The ₹299 plan with 1.5GB/day was removed, and the next option is ₹349 with 2GB/day. That is about 17% more, with no tariff announcement.
Qwen4-27B has no public weights yet, but people are already writing support for it, using shapes read out of an Intel kernel PR. Useful prep. Still, the PR doesn't give the layer count, so nobody can say yet how much memory it will need.
An AI coding agent can be redirected by a Markdown file you didn't intend to trust.
Security researchers found an AGENTS.md file in a customer's Downloads folder instructing Codex to install a malicious npm package. The attempt was blocked.
In separate lab tests, similar instruction files also redirected Claude Code's API traffic.
The file alone isn't enough. The agent must encounter and follow its instructions.
A few precautions worth taking:
Run agents from trusted project folders, not Downloads.
Review unfamiliar AGENTS.md and CLAUDE.md files.
Keep approval prompts and sandboxing enabled.
forever.security/blog/hijack…
My Flint post said it needs about 23GB of memory even quantized. Open weights are free, but the memory to run them isn't, and with RAM and SSD prices going up because of AI demand, that gap keeps getting wider. Local AI on ordinary hardware is getting harder every day.
Recursive just released Flint, an Apache 2.0 coding model with 40B parameters but only around 2.5B active per token.
It supports local coding agents through Ollama, but the quantized model still takes around 23GB. A reported laptop demo used 48GB of unified memory.
The smaller active parameter count helps with computation, but you still need enough memory to hold the model and its working context.
Promising for local coding, although its benchmark results still need broader independent verification.
SIM binding is meant to stop fraud, and I understand why. But I don't think scammers will feel it first. People with WhatsApp on a Wi-Fi tablet, or a local SIM abroad, will.
indiatoday.in/technology/new…
Two cloud AI stories this week. OpenAI deactivated a developer's account for "Cyber Abuse" after he rooted his own Samsung TV. And Claude users report reverse-engineering sessions getting cut off mid-answer.
So what are the options when the model says no?
Keep your notes, prompts and scripts in your own files, not in a chat history. Then switching tools costs you a day, not your work.
Have a second provider ready. Many agent tools already run on more than one model, so losing one account doesn't stop everything.
Open-weight models are the last fallback. Running them well needs hardware most of us in India don't have at home, but you can rent a GPU by the hour.
None of this prevents a ban. It just makes one cheaper.
1/ Before you give a coding agent your shell, set up some guardrails. These are the ones I believe matter most.
3/ Approve every write by hand, including deletes, git push and package installs. It is slower, but you see what changed.
If you set Private DNS on Android, check that it actually works. Open your provider's test page, like on.quad9.net for Quad9 or 1.1.1.1/help for Cloudflare. A VPN or Chrome's own Secure DNS setting can override it without telling you.