Your weekly AI research, done while your laptop sleeps (the RESEARCH setup)

This is the setup from the reel. The reel said your weekly research, settled in one message; this is the system that produces it. Every Wednesday at 8 AM, a scheduled job wakes on a remote host, sweeps five sources for new AI releases, filters them to hardware people already own, and drops a single Telegram brief in your hand.

What you need

• A free Telegram account. The bot sends to your own chat.
• Something that stays on to run the schedule. A cron on a sleeping laptop never fires.

The system

Four pieces, nothing fancy:

1. A schedule. A cron job wakes the task every Wednesday at 08:00 IST. A laptop that is asleep cannot be woken by a cron, so this runs on a remote host that stays on.
2. The sweep. A small script hits five sources: Hugging Face trending 45d, GitHub search (120d, 500+ stars), Show HN (60+ pts, 60d), HN front page (150+ pts, 45d), and lobste.rs (25+). 17 seconds on the first run.
3. The filter. An LLM prompt (below) applies two rules: keep only items that show a visible outcome on hardware you already own (laptop, phone, 4GB/8GB GPU, a Mac), and keep only items with one concrete figure (RAM, tok/s, parameter count, a price ratio). Items in lane with no figure go to UNVERIFIED; the rest are rejected.
4. The delivery. The brief lands in your Telegram as plain text, in a chat you already use.

The prompt

The sweep output is injected above this instruction each run. Copy the shape, replace the sources.

You are producing Rupayan's WEEKLY RESEARCH DIGEST. The text above this instruction is the stdout of research_restock.py, a fixed five-source sweep (Hugging Face trending, GitHub search, Show HN, HN front page, lobste.rs) already deduplicated and stripped of items he has posted before. You have no other input: do not browse, fetch or verify anything, and do not invent items. Everything you write comes from the sweep text.

Lane rule (filter 1): keep only items that show a visible outcome on a resource the viewer ALREADY spends: consumer hardware they can hold (a laptop, a phone, a Raspberry Pi, a 4GB or 8GB GPU, a Mac) or their AI API bill. A big model on a small machine, a fine tune on a laptop GPU, a cost ratio against a hosted API. Datacentre GPUs, agent frameworks, developer productivity tools and generic AI news are OFF lane.

Number rule (filter 2): an item needs one concrete figure visible in its title or description (GB of RAM, tokens per second, parameter count, a price ratio). If no figure is present, list it under UNVERIFIED, never make one up.

Output EXACTLY this shape, plain text, under 3000 characters total:

WEEKLY RESEARCH DIGEST, <today's date>
Swept: <raw> raw, <novel> novel (copy the two numbers from the sweep line)

KEEP (max 5, best first)
1. <title> | <the one figure> | why it fits: <one short clause> | <primary URL>
...

UNVERIFIED (in lane, no figure on the card, max 3)
- <title> | <primary URL>

REJECTED
- off lane: <count>
- no primary link: <count>

Nothing here has been checked at source. Rupayan verifies each figure at its primary URL before using it.

What it gets wrong

It does not verify links. It reads what the sweep surfaces and trusts that the titles and descriptions are accurate. On the first fire, 2 of 5 kept items were wrong:

• Steam Frame: a $1,059 handheld. A piece of hardware, not an AI model running on hardware you own.
• A $20 4G hotspot hack. A telecom trick. Not AI at all.

Neither was flagged. Both sat in the KEEP list, looking exactly like the three good ones. The only warning is the brief's last line (nothing here has been checked at source), and it's the most honest line in the whole thing. So you read the five, open the links, and skip the wrong ones.

The catch you should actually think about

It clears the pile, it does not clear your judgment. I am not automating the decision of what matters to you. The brief is up to five picks plus up to three unverified, each with a link you can check, and you pick. The time saved is the reading, not the thinking.

The measured run

First scheduled fire, Wednesday 16 September 2026, 08:00:11 IST, completed 08:03:27 IST (3 minutes 16 seconds end to end, 17 seconds of it the sweep):

• Sweep: 382 raw, 314 novel (five sources, deduped). The model is shown the top 25 from each source, so it read at most 125 of them
• Delivered: 5 KEEP, 3 UNVERIFIED; the brief counted 302 off lane and 4 with no link
• Of the 5 keep: 2 were wrong (Steam Frame, the 4G hack), 3 were real
• It still runs: the 7 October run took 8.6 minutes, the 14 October run 1.8 minutes, both delivered.
• By hand: not timed. Every person's source list and tolerance for noise is different. Time one real session yourself; that is your baseline to beat.

Where the longer version lives

The Rabbit Hole is the community where I post the setups, the ones that worked and the ones that didn't. Ask there and I answer in the open, so the next person with your question finds it. It's free to join; Premium ($5 a month) adds weekly office hours. skool.com/the-rabbit-hole-3206