Rasmus Schultz

web developer since 1998

I hate that I'm on this platform, but apparently developers are stuck in purgatory for eternity 👿

[Rasmus Schultz @mindplaydk]

@WCK604 @DarioCpx @edzitron are they? 😂

https://t.co/nWxG5Id99y

[Rasmus Schultz @mindplaydk]

@DarioCpx @edzitron why does he think "it's going to cost just as much money to finish building data centers in 10 years as today"?

the cost of inference is going to come down massively in 10 years

it's probably going to do that in 2-5 years 🤷‍♂️

[Rasmus Schultz @mindplaydk]

@zack_overflow oh no. 🫤

do you know if they've got any plans to improve this?

my system already struggles with kind of project sizes I work with

I thought they had said it would use *less* memory than current TS 🤷‍♂️

[Rasmus Schultz @mindplaydk]

@therons @sluongng it's just a program

there's no daemon running with elevated privileges like Docker

containers under Podman can't do anything the current user can't do

which is also just so much easier to understand, I think 😄

[Rasmus Schultz @mindplaydk]

@jumperz I wonder if that's the trend now

they're at the point where cost & compute largely follows quality

so they'll just start with smaller models and wow you with cost

then leverage brand recognition to ship something bigger and now expensive 😒

[Rasmus Schultz @mindplaydk]

@jumperz looks like an introduction offer

the price will double vs M2.7

no duh it's going to be better if you double the cost & compute

Google recently pulled the same stunt with Gemini Flash

just ship a bigger more expensive model, of course it's going to be better 🤷‍♂️

[Rasmus Schultz @mindplaydk]

@thdxr but you can't explain why or how, and you won't be taking questions, right?

I suspect many have been where you are right now and it tends to pass 😌

[Rasmus Schultz @mindplaydk]

@midego1 @sluongng no, it's using the container to access the file system on the local computer

yes that's a thing

the docker daemon typically has/needs elevated permissions - it's a known design issue

we should be switching to Podman 😌

[Rasmus Schultz @mindplaydk]

US export controls blocking DeepSeek from buying H100/H200s apparently backfired spectacularly. The Huawei-based architecture, essentially a workaround for inferior silicon, can't run on western hardware.

Probably the best open model, and we ship an expensive garbage version. 🤦‍♂️

[Rasmus Schultz @mindplaydk]

Western providers of DeepSeek V4 Pro look really bad on OpenRouter right now - the cheapest provider is 6x more expensive than first-party inference, and badly degraded at 4-bit quantizations.

https://t.co/YujvKOTRLX

[Rasmus Schultz @mindplaydk]

It's almost as if they're just teasing the west with an open release of this model. Western providers aren't cutting corners with a 4-bit version - they literally can't fit the full model on the H100 fleets they have.

https://t.co/5p9Cfm3XWm

[Rasmus Schultz @mindplaydk]

@thdxr I mostly use Pi now

it has 4 tools

I have just one extension for permissions installed

it works really well for my type of use, (which is more "refactor this" and not "build this for me")

what specifically do you think we're missing out on?

[Rasmus Schultz @mindplaydk]

@serenaa_ge so I get that the solutions aren't in the public repos you clone for this benchmark

but the codebases are - some of these repos have been public for a decade

LLMs having already been trained on the code they're asked to work on... this doesn't affect the results? 🤔

[Rasmus Schultz @mindplaydk]

@luminousmind_co @rohanpaul_ai yes let's hear what the silly robot thinks

[Rasmus Schultz @mindplaydk]

it does this all the time with more important, technical questions as well - sometimes while giving step by step instructions on shell commands, it'll suggest the wrong thing and then "but wait actually" itself. 🤦‍♂️

[Rasmus Schultz @mindplaydk]

this is just about the worst answer I've seen by Claude

https://t.co/GUvlabcaCq

this nonsense belongs in a <thinking> block where the end user doesn't see it - that's what reasoning is, right?

they've literally fine tuned it to give wrong answers first - why did they do that??

[Rasmus Schultz @mindplaydk]

@Everlier @GoogleDeepMind @Google yeah but no

https://t.co/vTw0yj0PGI

this is not a small model

[Emilz @emilheap]

@GoogleDeepMind @Google Holy expensive for a flash model ( $1.5 / $9 )

Almost same price as Gemini 3.1 Pro Preview

3x - Gemini 3 Flash Preview

5x - Gemini 2.5 Flash

6x - Gemini 3.1 Flash Lite

[Rasmus Schultz @mindplaydk]

@PJaccetturo it's alright, I think this is much better 😊

https://t.co/pAckwxNqAw

[Rasmus Schultz @mindplaydk]

@VictorTaelin @ton_anywhere fine tuning is not too expensive

you're not training a new base model 😅