Michael Kopietz ( @rapso )

father of two girls, code magic @ EA/Activision/IdSoftware/CryTek/Ascaron

[Michael Kopietz ( @rapso ) @rapso]

In Memory of Elisabeth #Aiko Kopietz
20th August 2018 2:49pm - 13th September 2018 4:15pm

[Michael Kopietz ( @rapso ) @rapso]

for your Bingo Card: "Should we use the model that works or the inexpensive one"

[Michael Kopietz ( @rapso ) @rapso]

@vassilis3D @SheriefFYI You can probably write that tool in the time a modern version of it still initializes it's frameworks 😅

[Michael Kopietz ( @rapso ) @rapso]

Interesting claim, lets do the math:
Closest competitor of GeforceNow (81ms,56ms) is XBoxSX (80ms,40ms). Lifetime cost (~5y):
XBoxSX: $599
GeforceNow: $995
QED: It's just not economical to establish several times more powerful and expensive hardware to get a similar latency.

[LaurieWired @lauriewired]

you’ll get mad at me for saying this…but cloud gaming is so obviously more economically efficient than physical hardware I think it’s going to be the default soon.

your home console / pc is idle 90%+ of the day. meanwhile, data centers targets what, 5%, maybe at worst 10% idle.

every second a cloud gamer isn’t gaming, that hardware is being used for someone else, training, etc.

I think there should be a new measurement, something like cost-per effective FLOP hour that takes into account the TCO + effective utilization.

If a gamer spends $500 on a GPU, uses it for 3 years, but it’s only fully active ~5% of that period…the cost-per relative FLOP hour is crazy high! Meanwhile, a $50,000 datacenter GPU might have a *LOWER* cost-per FLOP hour just because the effective utilization is 90+%.

[Michael Kopietz ( @rapso ) @rapso]

@digitalfoundry That's actually nice to have an upgrade option to 16GB

[Michael Kopietz ( @rapso ) @rapso]

Pro life hack to lower your companies AI costs: Hire Juniors.

[Michael Kopietz ( @rapso ) @rapso]

@lauriewired 99% of the time you don't need an SQL person.
You need C++, std::vector/std::*map and a decent server until you reach 1M+ Visitors/day.

[Michael Kopietz ( @rapso ) @rapso]

Opus 4.8: "Want me to wire a quick https://t.co/sfNFvIE7h3 so you can demo the model before the GPU run?"
me: "demo before training? I'm confused"
Opus: "Let me clear it up...
So an https://t.co/sfNFvIE7h3 "before training" would run, but it'd produce garbage. There's no point."

[Michael Kopietz ( @rapso ) @rapso]

use AI
use AI
use AI
use AI
use AI
use AI
use AI
use AI
..
don't use AI, it's too expensive

[Michael Kopietz ( @rapso ) @rapso]

I'M NOT JOKING #ChatGPT

[Michael Kopietz ( @rapso ) @rapso]

me: "you generated a script with a function signature missing"
#claudeCode : "That's a simple fix"
#claudeCode : "That's a simple fix"
#claudeCode : "That's a simple fix"
#claudeCode : "let me regenerate the whole script"

[Michael Kopietz ( @rapso ) @rapso]

would you believe Spotify devs don't write code anymore, but also:

[Michael Kopietz ( @rapso ) @rapso]

@ChristerEricson The time they gathered sources is kinda over, cause that pre-training plateaued already. Now they get paid $20/month by programmers to feed trainings-data for agents (cause you cannot download that). So, maybe Opus 5 will be a whole workforce you can rent, no need for vibe coders

[Michael Kopietz ( @rapso ) @rapso]

@SipeedIO The lowest that covers my use case. honestly, unlike gaming PCs etc. where you can decide based on guts and your pocket, on NPU workloads you either have enough ram or the device has little use. I might be fine with 1GB if I use a very limited LLM for home automation.

[Michael Kopietz ( @rapso ) @rapso]

@AnthropicAI 2h+ in for setting up the status bar of their CLI. Either #claudecode or I hallucinate, let me know your vote. Wonder whether it will manage it before I ran out of "usage"

[Michael Kopietz ( @rapso ) @rapso]

got #ClaudeCode and asked it to adapt the status line of their tool and their API... 5 bugs in 2 lines

[Michael Kopietz ( @rapso ) @rapso]

@AnthropicAI 30min.... would it detect if it crashed somehow?

[Michael Kopietz ( @rapso ) @rapso]

"Weights" of a model are not Neurons, these are Dendrites. Thus the human brain is not a 86B model but a 100T-1Q model, if at all comparable.

[Michael Kopietz ( @rapso ) @rapso]

They steal from us what we stole from the internet. That's not fair! American models can stomp on the rights of any person of any country on this planet! But how dare you do the same to us! We need a coordinated movement to protect the treasure we gathered from all over the world

[Anthropic @AnthropicAI]

We’ve identified industrial-scale distillation attacks on our models by DeepSeek, Moonshot AI, and MiniMax.

These labs created over 24,000 fraudulent accounts and generated over 16 million exchanges with Claude, extracting its capabilities to train and improve their own models.