603
"lessons learned" (lemmy.ml)
submitted 9 months ago by to c/fuck_ai@lemmy.world
146 comments fedilink hide all child comments
 

screenshot of followup to the first tweet, also by @BenjaminDEKR: " Made some adjustments based on lessons learned. Combined: roughly 200-400x cheaper heartbeat operation."

https://xcancel.com/BenjaminDEKR/status/2017660150463582282

via this bluesky post:

screenshot of bluesky post by @rusty.todayintabs.com with text "I’m starting to think the people who are excited about “AI agents” have literally never used a computer in their lives" and screenshots of the above tweets

screenshot of tweet by @BenjaminDEKR with text: OpenClaw is interesting, but will also drain your wallet if you aren't careful. Last night around midnight I loaded my Anthropic API account with $20, then went to bed. When I woke up, my Anthropic balance was $0. Opus was checking "is it daytime yet?" every 30 minutes, paying $0.75 each time to conclude "no, it's still night." Doing literally nothing, OpenClaw spent the entire balance. How? The "Heartbeat" cron job, even though literally the only thing I had going was one silly reminder, ("remind me tomorrow to get milk") 1. Sent ~120,000 tokens of context to Opus 4.5 2. Opus read HEARTBEAT .md, thought about reminders 3. Replied "HEARTBEAT_OK" 4. Cost: ~$0.75 per heartbeat (cache writes) The damage: - Overnight = ~25+ heartbeats - 25 × $0.75 = ~$18.75 just from heartbeats alone - Plus regular conversation = ~$20 total The absurdity: Opus was essentially checking "is it daytime yet?" every 30 minutes, paying $0.75 each time to conclude "no, it's still night." The problem is: 1. Heartbeat uses Opus (most expensive model) for a trivial check 2. Sends the entire conversation context (~120k tokens) each time 3. Runs every 30 minutes regardless of whether anything needs checking That's $750 a month if this runs, to occasionally remind me stuff? Yeah, no. Not great.
you are viewing a single comment's thread
view the rest of the comments
[–] 52 points 9 months ago (11 children)

Why use an LLM to solve a problem you could solve using an alarm clock and a post it.

  • permalink
  • fedilink
  • source
  • hideshow 11 child comments
    •
  • [–] 34 points 9 months ago* (3 children)

    programming nitpicks (for the lack of better word) that I used to hear:

    • "don't use u32, you won't need that much data"
    • "don't use using namespace std"
    • "sqrt is expensive, if necessary cache it outside loop"
    • "I made my own vector type because the one from standard lib is inefficient"

    then this person implemeting time checking work via LLM over network and costs $0.75 each check lol

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] 20 points 9 months ago (1 child)

    We used to call that premature optimization. Now we complain tasks don't have enough AI de-optimization. We must all redesign things that we have done in traditional, boring not-AI ways, and create new ways to do them slower, millions or billions of times more computationally intensive, more random, and less reliable! The market demands it!

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 15 points 9 months ago* (last edited 9 months ago)

    I call this shit zero-sum optimization. In order to "optimize" for the desires of management, you always have to deoptimize something else.

    Before AI became the tech craze du jour I had a VP get obsessed with microservices (because that's what Netflix uses so it must be good). We had to tear apart a mature and very efficient app and turn it into hundreds of separate microservices... all of which took ~100 milliseconds to interoperate across the network. Pages that used to take 2 seconds to serve before now took 5 or 10 because of all the new latency required to do things they used to be able to do basically for free. And it's not like this was a surprise. We knew this was going to happen.

    But hey, at least our app became more "modern" or whatever...

  • permalink
  • fedilink
  • source
  • parent
  • [–] 9 points 9 months ago

    using namespace std is still an effective way to shoot yourself in the foot, and if anything is a bigger problem than it was in the past now that std has decades worth of extra stuff in it that could have a name collision with something in your code.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 1 point 9 months ago (4 children)

    Or if your being fancy poll a time server

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 4 child comments
    •
  • [–] 2 points 9 months ago (3 children)

    That would work great as well but an alarm clock is a technology developped in the middle age.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •