I was thinking that on this, folks need a cron task to run a trivial prompt at 5-6am and get that 5hr timer running so that it the majority of the quota is available in the working day morning, and then a new 5hr block starts around lunch time. This should maximise use of included tokens by a standard work day spanning 3 blocks rather than 2
Also useful for paid APIs like DeepSeek's, where they have cheaper inference price (50%/75% off) for UTC 16:30-00:30, so being able to schedule some stuff you know would take a ton of tokens for that time period would make sense.
I'm in a similar position of never having found a use where memorising lots of facts would be useful. The main use I keep seeing is vocabulary building when learning a language. I'm sure people are using the system for learning other stuff too though?
Seeing this did make me wonder how I might be able to get better at memorising important parts of iso/iec standards at work, but I can't see how that maps to flashcards
In what context do you find yourself wanting to recall a specific part of an iso/iec standard? Distill that context into a short description and put it on the question side of the flashcard. The answer side then has the corresponding information you want to memorize.
But of course it's possible that you almost never need the same information twice, in which case committing it to memory wouldn't be particularly useful.
I've built one for the US Amateur Radio exam pools (which are public) as I'm trying to sit for both my Technician and General in the same volunteer exam session.
Mixing the questions between both pools and studying as a unit I've found has had two great benefits: 1) I'm not focusing on Technician first and then going for General as a "bonus" and 2) it helps me see the connections between the material.
I have used Anki to memorize cube numbers and roots, recipes, and music theory. If you're interested in other ideas, you can browse public decks here: https://ankiweb.net/shared/decks.
I've been following 16hr fasts by skipping breakfast and eating my first meal at around 12:00 each day. Normally have an afternoon snack, then dinner at 18:00 with my young family. Perhaps a sweet treat by 20:00 after putting the little one to bed. Honestly, it's not that tricky if you bulk up lunch a bit.
Edit: As a sibling comment says quite rightly, you do feel hungry in the late morning, but reacting to that feeling is optional
I stopped feeling hungry in the morning a long time ago. Just unhealthy amounts of coffee, without sugar or milk. If I eat even just the tiniest snack or sweet, the food processing tract will "wake up" and it's over. But if I can avoid that, I only break the fast because of convention, not because of hunger.
But it's also very contexy sensitive: currently working from a place where I usually go for high calorie throughput sports (think Tour de France climbs, but higher and a heavier rider, obviously a lot slower but the energy demand is mostly mass x elevation, almost unpacked to speed) and my body is in "eat! you will need it!" mode every day. Crazy weight gain on the working days.
I came here to say the same as the parent comment - it's an amazing achievement, but you may well have built a medical device which needs certification in order to be on the market in the territories you want to use it in.
I work (freelance) with a consultancy [1] that helps specifically with software-as-a-medical-device (SaMD). My email is in my profile if you want to chat about what might be the best way forward.
I'm working on scoping out specific problems facing "software as a medical device" (SaMD) companies. In particular issues around being able to release software at a reasonable cadence. I've been a CTO in this space for a couple of years and I am now consulting with other firms around the intersection of tech and regulatory.
It's a tight-rope walk of ensuring that all testing (software and non-software testing) and evidence is produced correctly and being able to release at a rapid pace to derisk each release. It's not uncommon for software to only be updated yearly, leading to very conservative changes and little iteration. Monthly releases are okay, but still not great.
I want to make it possible to release at least weekly and to do so safely.
If you work in this area, I'd love to chat and hear your experiences (email available via my website in bio).
Neuronostics is revolutionising diagnosis and prognosis of neurological conditions. We create algorithms and software which can be used to aid medical professionals gain more value from EEG data than ever before.
Role involves: Working within the tech team to take scientific code from our science team and get it into the hands of our customers, help to build out the rest of the tech team.
Does anyone have any useful rules-of-thumb or heuristics for balancing this trade off of upfront cost v.s. power cost? e.g. how much does an N100 cost to run for a year v.s. say a i5-2400s (the CPU for the first row on the linked site)?
I used to calculate costs of lightbulbs: 1 Watt running the whole year, at 0,28 eurocent/kWh costs 1 Euro per year. Until someone corrected me and it turned out that every 1 Watt 24/7 will be 2 Euro per year.
In the US electric power might be cheaper. And if it's running only part of the time, you should adjust the calculation.
My desktop/server runs 24/7, so I prefer having a CPU with 65W TDP over one that is 125W TDP. That might run up to 120 Euro per year difference for me (if it would be running at 100% CPU).
Real world energy use is nothing like what you see on spec sheets. And not just because manufacturers differ in how they compute TPD. And TPD is also not a good indicator for energy use at (near) idle. With underclocking/volting in the BIOS you can get a beefier CPU to outperform smaller CPUs per watt. Because CPUs get really inefficient as they use more power undervolted or capped high TPD chips might be much more power efficient in the real world than their low TPD counterparts.
My NUC13 with i3 has a nominal 15w TDP, but while idling on a KDE desktop with a browser open to reuters (1 tab) it hovers around 3 - 4w (5% CPU usage). If there's REALLY nothing going on (no desktop even) it's 1.0 - 1.3w (1% CPU usage).
Edit: I should note that there's no fan drawing power because I put it in an Akasa passively cooled case.
I tried to find this out myself. All I could find easily was the TDP of different processors. But I'm not sure if it's a good measure of how much power it will use.
I went down this rabbit hole earlier this year. Best I came up with was to calculate the TDP at max for the whole year. Full TDP is unrealistic, but it gets us a worst-case "max running cost" . Energy for me is roughly $0.12/kWh, so the yearly max running cost for a 35W TDP is $36.79, 65W is $68.33, and the 95W would be $99.86.
I ended up going with a HP EliteDesk 800 G5 Mini I5-9500T (35W) off of Ebay for $100 and it does the stuff I need it to do just fine. According to my current monthly power usage graph, it's averaged 7W which accounts for $0.61 of this month's power bill.
No, sadly the TDP tells us every little about the idle power cost, which might be where you spend most of your time depending on the workload.
Just from tweaking my laptop, I’ve noticed that when it is really idle (or I’ve intentionally put it in a low frequency mode), the big power drains are the wireless interfaces (don’t forget bluetooth) and the screen (OLED helps as long as the screen is mostly black). Gotta tweak the whole thing.
The only real way of knowing is to measure it. If you already have a system in place an energy monitoring smart plug can help you calculate the current running costs and help estimate the savings of using a lower-power machine.
When I did this I was surprised by how much - or how little - it cost to run various devices. It's quite addictive.
It's not always accurate because a lower-power machine doing the same task will often need to work at its full power more often, so the savings may be less. For example, a Raspberry Pi 5 may often be more power effecient than a Pi 4, despite drawing more power at full capacity on paper, because it spends less time at full capacity than the Pi 4 does.
On the other hand, when I upgraded my work PC I found it used less power but I also had to run my office heater more often in winter, as the new PC wasn't as efficient at heating the space.
Thanks for the heads up. There is an email address to send applications to in the posting, and I am getting the applications opened up again. Thanks again!
It's a 2 step process for me:
1. Sanity check the python (easy for CSV, less so when it's extracting a table from a pdf)
2. Ask chatGPT to do a checksum of some sort, e.g row count matching
