TekJumble

A geek and a tinkerer.

[TekJumble @TekJumble]

@AutismCapital Sure it’s got root in an LXC container with only open port being 443 and no access to lan

[TekJumble @TekJumble]

@nevenvorkapic Nudim uslugu download ram-a za ove ostale nesrecnike !

[TekJumble @TekJumble]

@plotarmordev It really does, even 5.6 luna wipes the floor with any and all openweight models its just the truth

[TekJumble @TekJumble]

@sudoingX Always

[TekJumble @TekJumble]

@amritwt Shit I mostly get grglgrll from mine … how do you prompt it ?

[TekJumble @TekJumble]

What could this be for ?

[TekJumble @TekJumble]

@mgrandic @jedan_magarac Mozes da ucis do sutra o promptovima, prompt sto se tice AI je samo predlog nakon par desitina hiljada tokena kada kontekst izadje iz glavnog "attention" mehanizma umesto da prati prompt rec za rec AI ce parafrazirati i ne uvek dobro i ne uvek onako kako si misio.

[TekJumble @TekJumble]

Te kurseve koje pravis mozes da "prodas" mahom samo najvecim bolidima, AI treba koristiti iskljucivo za tehnicke stvari koje se mogu mehanicki proveriti i dokazati, napisati kurs koji je neproveren i bez citata je apsolutno ludilo.

Ono kad ga pitas o necemu sto je u tvojoj sveri znanja i dobijes odgovor koji je izlupan na max i cudis se kako je upste dosao do zakljucka koji prezentuje, a onda pitas nesto o cemu ne znas nista i cini ti se da je odgovor magija i odmah verujes je u stvari vrlo dobro poznata zamka sa AI uglavnom.

[TekJumble @TekJumble]

@miloshjokic Da generalno, medium je neka zlatna sredina isto i kao sa fable vrlo su slicni sada.

[TekJumble @TekJumble]

@0xSero I can't wait for you to post some number i am really curious on how they perform especially in larger deployments like 4+, tp4 ppXX type of thing i know software is not there yet but i would like to see some numbers

[TekJumble @TekJumble]

@MiaAI_lab Yeah same with the token usage my dude, its CRAZY the overthinking it does on xHigh, on MAX its just unbearable to watch !

[TekJumble @TekJumble]

@jun_song I mean sure just set to a lower effort but does it retain the same level of intelligence what i would like to see is same intelligence on 4.0 medium vs max on 3.8 now that would be something

[TekJumble @TekJumble]

@PawelHuryn What about high or medium ? I find that with MAX it tends to overthink ALOT !

[TekJumble @TekJumble]

@MiaAI_lab @ArtificialAnlys It has been thinking a lot more with equivalent setting compared to Opus 5 at least that is my preliminary finding

[TekJumble @TekJumble]

@donnfelker Thank you !

[TekJumble @TekJumble]

I think they are still not there in terms of intelligence, but what they could do is lean heavily on ngrams for general knowledge and fill the base model with as much intelligence as possible

[Ivan Fioravanti @ivanfioravanti]

We have reached the good enough intelligence level, I hope everyone will now focus on make models faster more than smarter.

[TekJumble @TekJumble]

@ItsmeAjayKV Better but still not good

[TekJumble @TekJumble]

Quant, context length, number of streams ? Saying just this is equivalent to saying my car goes fast … when pushed of a cliff

[Max Weinbach @mweinbach]

We are now at 3740 tok/s prefill, 149 tok/s decode batched on M5 Ultra for Qwen 3.8 Flash Next

This is nearly double what it was yesterday