#π»βDev's Forum
1 messages Β· Page 7 of 1
wasnt that fake at first
demo niye fraud
ganja kinte $500/mo
slack e devin chat
completely autonomous, like a real person, can push to your git repo, make changes you ask
oita agent != ai/model.
agent just kisu prompts collection + keyboard input client. toiri kubi easy.
claude holo model. claude er "computer use" holo agent ja devin theke kutigun better hobe karon claude model powerful ar akash patal price er parttokko.
yes
devin the company may not be worth 2 billion dollars,
50 million max, but it is something, it isnt trash
they use a proprietery model, not claude or any 3rd party model, they claim it is even better than o1 in coding, twice as better as they put it
i mean meh
no
it isnt
but ok bro
probably like gpt4o(since open source model like llama 3.3 already match that)
but it is something, it isnt trash
hae trust me bro benchmark theke janlam.
reddit er oneke duita model try korse.
devin er proprietery mane gpt er api use kortase. ei jinish kisudin agei onno ekta company korse pore reveal hoise
autonomous impressive kisu na.
cline ekta autonomous agent. shudu github control (pull request) ability nai.
easily add kora jai kintu tara intentionally korche na.
oh na
oita ager benchmark
they released 3 days ago
afaik
hmm
Yo ROblox Studio is Fire π₯
Allows U to test the multiplayer features without friends π
Oh yeah don't look at my ~~cringe ~~T-Shirt
Should I wear this instead?
take this
it's not meant towards u
but
yk
so U can use it later
send a bigger photo
Imma try
I just screenshotted it from a server's sticker page
And IDC if screenshotted is a word or not
If it makes sense, then it works
This one's fire tho right?
Roblox engine got some potential tho @visual geode
Hay Allah the video compression π
moderators, give the the glock glock 9000 for saying that
const systemPrompt = `
You are an elite technical recruiter performing a comprehensive job application evaluation.
Be specific and evidence-based. Avoid generic statements. Provide actionable, constructive feedback.
Core Evaluation:
- Alignment of technical and soft skills with job requirements
- Look for job titles, industries, and projects that match the role
- Check for grammar, spelling
- Check relevant degree, GPA, relevant courses
Red Flags to Watch:
- Frequent job changes (less than 1 years per job)
- Unexplained employment gaps
`
const userPrompt = `
Job:
${JSON.stringify(result.job, null, 2)}
Candidate:
${JSON.stringify(result.user, null, 2)}
Application:
${JSON.stringify(result.application, null, 2)}
Questions:
${JSON.stringify(result.questions, null, 2)}
Answers:
${JSON.stringify(result.answers, null, 2)}
`
const stream = await openai.chat.completions.create({
model: 'gpt-4o-mini',
stream: true,
messages: [
{
role: 'system',
content: systemPrompt
},
{ role: 'user', content: userPrompt }
]
})
What the sigma
gpt prompts kore ja bujlam joto specific hobe output toto balo ashbe
instead of check if resume matches the job use check hard technical and soft skills
instead of provide downsides of this resume use Check for grammar, spelling, Frequent job changes (less than 1 years per job), Unexplained employment gaps
ekane ami 1 years specific babe bole disi. plain text hole approximate line numbers o bole deoa uchit.
starting with Python π£οΈπ₯
end goal: Be able to create a Discord Bot
wish me luck
Prompt engineering gonna be pretty big apparently
r8
nah dont think so
your prompt => agent => chatbot will do
ai agent will do the nessesary changes
I guess it could be like speedrunning a new game, keep finding out ways to optimize shit better
prompt expand gpt nijei pare. agent lage na. kintu kunutai eto detailed prompt pare na.
Ki kow
I think only weak ai will require specific prompting anyways
Does specific prompting Claude do anything, like at all?
Don't think it does
specific prompting general user er jonno na. ai application er jonno. jekane strict output dorkar. kisu ketre fine-tuning o lage.
jemon claude ke jodi pdf attach kore simply boli "rate this resume of out 100". ekbar bolbe 80%, pore 93%, 70% purui random
jodi specific babe boli je if projects are relevent then add 10% points, if has major in computer science add 2%, if undergraduate then cut -10%. tokon ai always same accurate result dibe.
eije cline(claude dev) er advanced prompt example
ami ekon gpt 4o use kortasi.
YYYY-MM-DD specify kore dite hoise nahole random ordering dei
ar input document e jodi only year "2020" thake tahole GPT "2020" return kore ja invalid.
if not parsable to ISO eita bolai only year thakle gpt null return korbe
eikane personal keyword ta na use korle portfolio projects urls (free hosted with github subdomain) return korto randomly e.g. project1.github.com jeta personal profile na. personal profile github.com/user1
how did they do the animated scroll in neovim terminal π
how tf this plugin work
https://github.com/karb94/neoscroll.nvim
how can neovim api allow this π
lemme find out
my bad, i suddenly have no desire to know π
hmm
doesnt work anymore skullcrying~4
this works
Steams api is crazy π€ͺ π π³ π π€§ π π€ͺ π π³ π π€§
how
stweams api is cwazy π€ͺ π π³ π π€§ π >w< π€ͺ π π³ π π€§
gemini 1206 claude theke smart hoye gese. resume scoring question korlam. eta geminir response
balo je markdown export ase
claude 300 line lekei sesh
gemini code tao kubi advanced. keo jate work experience startDate, endDate too large diff kore spam na korte pare tai Math.min() use korse
Yea Gemini good, but not smarter than Claude in coding, better in other stuff, not coding
There is this real cheap open source model, marginally better than Claude in coding benchmarks, deepseek v3, dirt cheap api
I would say that's a weakness, bad code, maybe someone has 25 years of work exp? Not 7. Spam korle Oita je judge korbe, she decide korbe, not code π
eta weights. number point na. weight: 7 line ta deho. score calculated out of 100 percent.
ekane 20 tar moto criteria ase alada weights ase jar jugfol 100.
jodi simple full number addition kore tahole out of bound hoye jai.
proportion 100 er relative tik rakar jonno korse
onekta erokom keo jodi 80 pai tokon amra boli A+. she jodi 99 o pai tokono A+.
A+ holo weight 4.0
80 holo number point
Oh
Then I think capping at 7 kinda dumb tho, maybe years of exp it is worth more "points" if it is 7 or below, any number after 7 gets lower points, tiered marking or smth, basically after subtracting 7
any number after 7 gets lower points
25+ exp senior ke minus points diye junior kora aro inaccurate. koto minus kora fair shetao idea nai.
capping e ei issue nai. generally job post jodi "senior engineer" role hoi.. applicant 10+ years exp holeo same senior salary treat korbe.
mane 7 na dia job.requiredExp capping use kora jai Math.min(totalExp, job.requiredExp)
ar overengineering er limit nai. dynamic weight o ase jeta claude/gemini bujbe na jodi user pointout na kore.
return (Math.min(totalExp, 30) / 30) * 7 // 30 max allowed experience
bro really didnt understand
first 7 years worth more
then after that, rest of the years get lower points
its like multi tiered taxation
senior not getting a penalty
toile koto lower points korte hobe kibabe bujba.. math ta ki?
final percentage er proportion damage na kore
make it like first 1 year is worth the most,
subtract 1
move to next tier, 2-4 year worth lower than last
subtract the 3, move to next tier
5-7 year worth the lower than last
subtract the 3, move to next tier
this style i suppose makes much more sense
fairer assesment
eita
ekon etare realtive korte gelei gure fire capping korai lage seshe.
10 years er por shobar value 100% ditase.
Cudis how do I port an Android build?
bro β οΈ
make it worth less and less, to the point it becomes insignificant after a while
make it worth less and less, to the point it becomes insignificant after a while
hae geminirMath.min(resume.experience, 10)already etai kortasilo.
main point ta holo
1 < 2 < 3 < 4 years... next year is better than previous ei order ta tik raka within a limit. jathe later sort by kora jai.
2 solutionei ta parse.
Math.min soluton ta strict.. joggota onujai scoring.
ar tiered system magical kisui na. notunder chance diye puratonder disrespect korar scoring.
jeta job er moto professional case e manai na jeheto precise na. schools result er ketre ok.
im making a game, its crazy how easy it is to implement stuff
Why
I have this stupid carrier locked phone and I am just experimenting
It doesn't let me unlock Bootloader without MTK
Secure boot Bypass doesn't work
Like it's insane
How much they locked this shit
You are cooked
hmm
They're carrier locked?
I never saw a carrier locked phone in BD
Like, the phone is locked at a specific cellular brand
Only sim cards from that brand will be accepted by the device
kisui nai. root, bootloader unlock, rom flash shoboi korsi.
kali kernel upgrading ta kori nai, kubi risky
mediatek cpu, fastboot out of the box kaj kore. alada xiomi ba samsung er moto hacks kora lage na
I unlocked the bootloader and dumped the ROM
Now I'm trying to flash that ROM
To test it
But it needs Auth and DA file
In SPFlash tool
but fastboot does work in xiaomi out of the box
You just need to carrier unlock through their shitty app
alada xiomi ba samsung er moto hacks kora lage na
shudu carrier unlock na... xiaomi bootloader unlock korte mi flash tool lage. out of the box adb fastboot diye hoi na
xiaomi bootloader unlocking moment
I thought it is the easiest?
Just need to have a xioami account
eta jodi easiest hoi. jei phone e account lage na, offline kaj kore error chara, android offical docs command use kore unlock kora jai sheta ki hobe? easiestest
Valid point
F-droid looks fire
So uhh guys
How do I bypass that?
Or get a DA or auth file?
youtube, xda search deo. solution na thakle... cooked
agae coding jinish note kore raktam. ki time waste ta kortam
kalke final porikka. bosorer pora 3-4 gontai..... chesta cholce
agae OCR pore prompt
mi flash er oita carrier unlock na hoile konta carrier unlock?
Carrier lock is Sim card lock
U can't swap the sim card with another carrier brand
Hmm i guess xiaomi doesnt have it maybe
It's a US only thing
U can finance phones with Carriers
and they lock it
so u don't just end up selling the phone to ppl
Skill issue for the us
Only way to Afford an $1,800 phone
π€£ arae eta already deklam boilai diso. ami abar ekei jinish bolte geslam ager messages read na kore
ah
chatgpt is soo cooked
open ai so cooked
deepseek reasoning 1(r1) matchs o1 in benchmarks
outpur costs only 2$
input is like 0.5$
o1 costs fking 15$ and 60$
their benchmarks have been trustworthy in the past
Project maintainer forcing me to read the docs
38 counts of cruelty, docs suck
deepseek use kore ja bujlam
eta gpt 3.5
Is that the Chinese ChatGPT I've been hearing?
Believe it or not it can beat Claude, with in context learning, they really could have used some programming related stuff in training data ....
Benchmarks test model's reasoning and problem solving capabilities and shit, and most of them supply some form of in context learning to my knowledge
Problem is the context window is only 128k, 64k in their website and api, 3rd parties may run the model with UpTo 128k context
Feels too low, but then again more context makes models gay
2million context isn't actual practical 2 million context in Google's Gemini, effective context length may vary
covidgpt
arae. deepseek ar claude ke same question korlam. deepseek use korse depricated aws-sdk ar claude korse latest aws/client-s3.
claude gpt 10
they really could have used some programming related stuff in training data
age eta hok. pore believe kormune
For deepseek you gotta provide some in context learning material
They really neglected puting coding material in learning data
It can reason, it just doesn't know shit π
They should fine tune it with open coder ai training data and call it r1 coder
Believe na believe er ki ase π, benchmarks speak for themselfs
oh yea woo yea woo yea woo yea oh yea woo yea woo yea woo yea oh yea woo yea woo yea woo yea oh yea woo yea woo yea woo yea woo yea oh yea woo yea woo yea woo yea woo yea oh yea woo yea woo yea woo yea
hey qwen 2.5 max not so bad
not as bad as deepseek with the max output token length limit and shitty knowledge out of the box
yea i'd rank it same as claude 3.5 sonnet
jak.. ami ekla na. deepseek je gpt 3.5 onnorao doubt kortase
yes π
bad out of the box knowledge, and they only have 2000 nvedia a800, they run in power saving mode if that makes sense, only 64k context, limited output token
@cloud badge
try
this
Qwen Chat offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.
claude grade
dont think so, they can do 1 final step, fine tune it with the knowledgebase, like those open coder models
having competent reasoner is the hardest part in my opinion, idk tho
hmm
it isnt even underwhelming, shocking they achieved true claude grade
no way it only has 100b parameters and only 8k context π
chinese ai underpowered probably from compute limitations
alibaba cloud so underpowered
the model aint even open source
China is very proprietary
if U ask the citizens even the most basic simple not so important and insignificant lil' tiny question about their government
they will snitch to the cops
and they get rewarded 500,000 Bing Chilling Currency
o3 just released
bro wtf, o3 is so shit in shit
o3 mini high
o3 didnt release
o1 was so much better
wtf
o3 mini high is so ass
what did they make
gpt4o is probably better
bro ngl, i would fire sam altman if i were a investor π
what is this shit
why is its output context window so small too
where did bro's we achieved agi go
Bro DeepSeek is trash
it says Wine 10 is old
and it launched in 2020 π
then I asked him "What about the Wine 10 that just launched, January 21 2025"
Current State of Wine (as of October 2023)
Latest Stable Version: Wine 8.0 (released in January 2023).
Development Versions: Wine 9.0 is in development and expected to be released in the near future.
Proton: Valve's Proton (based on Wine) is currently at Proton 8.x, which includes additional gaming-specific improvements.
π
I finally figured out the Screenshot shortcut on OpenSUSE
but anyways
How outdated is the bot?
Lmaoooooo
deepchick er reasoning R1 try koro.
same wrong answer dibe koek gonta thinking kore
hmm
Yeah helucination problem
O1 er oo erom ei sometimes
Reasoning models are easily deluded, like Claude feels way better than o1
Deepseek has deep helucination problem
etar source?
my bad, open ai employee claimed that
hmm
lets be honest, who we kidding, sam probably told him to claim that β οΈ
to generate hype
π
add the search
how will ai know when wine 10 is released without search
bro aint the future teller
deepseek (with search) vs gemini (normal).
nijer shathe nije argument kortase
Let bro think π
real real
π
Bro, ain't no one testing you π
They forgot to get the bot outta from Dev/Debug mode or smth?
Bro who am I? a Time traveler or smth?
Yea deepseek getting ddosed so probably that's why turned off search
Search is off π
bro has ice
ohh it showed me that too
I am done bro
YT Shorts took away my skill of paying attention
π
It was right in front of me, How did I not notice π
search jodi off o thake tobuo AI er dil mere 2022 bola expected na.
true true
Yeah, I tested with some in context learning, does well, but it helucinates many things, long way to go with helucination
deepseek aint it buckos
It is
Good shot at reasoning
Benchmarks aren't exactly lying, but it's out of box knowledge is shit
Also helucinates
R1 appears to be nicely integrated in perplexity pro
Nice in depth search
I have never seen an AI do this bru π
"Hypothetical"
"parallel"
Bro China is a parallel universe, no wonder their AI acts like that
bro works fine β οΈ
Nah I was just pointing out the hallucinations
it's comedic β¨
why is a model overactive in refusing, or mad weed helucinator π
why cant it do as i say without helucinating
perplexity rag use kore relevent texts parts ai ke bole dey ja hallucination 95% reduce kore.
rag use korle gpt-2 works fine.
rial
you can see deepseek deeply beating o3 mini high
write me a python file, use whatever module, show a visual representation of a ball with realistic friction gravity and physics, being moved around in a hexagon that is rotating at a customizable speed
when full fat 128k context size model is hosted, with 8k output, code quality niye pera nai, pera hoilo helucination tao even with search, it can get deluded, that offical one is so underpowered ngl
gax, i am a fan of perplexity's unlimited r1 usage and the powerful pro search
real openai advancements soon, closed ai L. hope official cloud can actually run 128k or 1 mil context soon
(perplexity r1 non pro search is free, and they probably use nvedia nim, check them out guys)
ling chon
Hyprland 
Yo it's one of those YT shorts balls
yea i saw in a reel, so i thought lets try it out
A Polyglot software engineer/programmer from Essex, just outside of London. Specialising in backend automation, SaaS platform building, MVP builds, apps, PWAs and many, many more. I will help define the tech of your business efficiently, radically and most importantly rapidly.
shi portfolio looks good
looks good
Visit us on the web: https://ladybird.org/
Join our Discord: https://discord.gg/nvfjVJ4Svh
Much love
try chat.deepseek.com vs chatgpt.com
not perplexity vs chatgpt. perplexity er rag (etar primary feature), chatgpt er web search theke odik powerful.
shob ai badh dia dinseshe ami gemini 1206 use kori.
eta bolte gele free unlimited claude (kasakasi)
kun part ta good mone hoise?
bad contrast (football)
inconsistent text casing (title case nai)
mane first impression real html programmer portfolio
Looks good on pc
Yeah, search makes it more powerful, but I enabled search on chatgpt too, I mean I know r1 is lacking, but it can me made up for in one way or another
oh nah it doesnt, only thing that makes it standout, now that i took a second look, not that good π
just stood out a little when i looking for portfolios
that blinking part, nothing impressive, idk why i liked it
Yes tried, bro it didn't use fetch in frontend some old xhr fetch or whatever
Code didn't even eork
So obvious, should have been able in the first try
Check this out @indigo plank
Official platform runs in power saving mode it seems like
Bongo boltu more nai, ami Mori nai
they have to
if they had more a100's then it would had been a different story
They only have 2k a800
That for the whole Chinese population
Chatgpt is banned in china probably
the kid has made us proud
Oh NAHHH
getting KDE vibes
kun dik theke?
delulu er dik theke
idk man, that's the first thing that came to my mind
Cuz of this prob
Linux phone w
Android is Linux with reduced capabilities
Vm for all apps
Gpu pass thru for acceleration
They create problem then solve it themselfs π
Just use damn containers instead of vm now
mane google engineer ra boka
essentially, instead of investing in kernel based isolation and security, they try to virtualize, anyways past is past, now containerzation is very mature tech
how do you make vm, have passthrus and still have malware π
how does that work
what kinda temu aah vm is that
oh and this part is not part of android kernel
π
counter productive as fk
π
pritivite emon kisu ase malware free? beshi-kom hote pare.
amar mone hoi tara korse vm generally container theke beshi secured, hardware compatibility or shared layer with chrome os.
Nah most of them use gnome
I feel like gnome sucks on mobile phones
Its more for tablet devices
idk where is fake overlay filesystem, faking permitions, disappointing, i mean that should have been the main selling point of isolation no?
many brands now partially supports it, for only some permitions, but not all
i think it conflict's with google's intrest, true isolation threathens their relevent ads revenue
kun brand korse?
ar container based mobile os keo korse?
Wait no koi gelo
I swear Amar phone ee asilo
Update er por gone
understandable
Like contacts er perm ee , ki Jani akta option silo, the app sees it can access it, but empty list is provided
Aidi lagbo
The Liberux Nexx is a new smartphone that comes with decent specs and is designed with security and privacy in mind. Using Linux as the OS, the developers also hope to make a statement against the common practice of planned obsolescence and make the smartphone as future proof as possible.
Check this out
oita possibly scam. tader website nai (LiberuxOS non-existent)
Control the hidden appops conveniently
root lage jodio.
android er internal linux access korte root lage.
er mane android (vm) ability ase kintu controls UI nai only.
Not just returning empty data, it should in my opinion also have option of proving fake/mock data as the user sees fit
Root diye possible yeah
I remember people using location mocker in Pokemon go
toile original linux ba linux phone os, ba iOS e ki kora jai egula?
rooted android already ekta arch linux. ja linux e kora jabe shob possible. shathe arch64 naki x86 app ei bishoita consider korte hobe.
android apk run hoi runtime e (vm). onekta "node" runtime jebabe JS run kore in a isolated way.
container diye apk run kora possible na.
container simply isolated environment ar android vm holo isolated environment + java bytecode runtime
Not really, they stripped many linux's capabilities in android kernel
Like namespace isolation, guid suid modification etc, to my knowledge, idk if they are possible in rooted, they are completely impossible in android, thus podman doesn't work on termux
Ki kow bro, Pura android system ei container ee run Kore, waydroid ee
waydroid container use hoi linux ar lineageos ke seperate rakthe.
pore lineageos vitore vm use kore android apk run korte
android translation layer = port of android vm without wrapper os (lineageos)
jeheto wrapper os nai. container o lage na.
normal linux > waydroid > lineageos in container > android vm (a.k.a. android runtime, ART) > running apk
normal linux > android translation layer (port of ART) > running apk
Oh right
Lol
memer theke nostalgic mone hoise beshi
Yea
Bro so wierd when code doesn't work, you give up all hopes of this janky implementation working
Then it suddenly starts working
πππ
Wtf
Relatable π
βββΏβ¬ βΏββββ ππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππππ
OPENAI ROADMAP UPDATE FOR GPT-4.5 and GPT-5:
We want to do a better job of sharing our intended roadmap, and a much better job simplifying our product offerings.
We want AI to βjust workβ for you; we realize how complicated our model and product offerings have gotten.
We hate
o3-MINI
most cost-efficient model, not the best of them
gpt-5 probably be claude 3.6 level. free finally
feb-mar e asha kori claude 4 ber hobe. job risk e ase
entire tech industry ei risk e ase. programming theke shuru kore, game dev, 3d modeling, music, movie shob... 2-3 bosor deki ki hoi
Well you have to kinda consider as it progresses further, speed of progression will come down
I believe it will be very slightly better
Or not even better it seems like, openai can't catch UpTo Claude even with o3 mini high
Bro try o3, kissu pare na π
I believe akhon just compute resource saving model banaibo, probably, Claude 3.5 grade but much lower resource requirements, idk seems like it
hmm
postgresql aint messing around, unholy speed of execution, compared to cockroachdb, it seems like almost 2x speed, probably not
but
dayum
type guarantees in typescript backend so gex π
i wish types were real like in rust, imaginary aah, keeps fking up predefined types
returns shit in formats it wasnt supposed to with no lsp warning

jemon
data as anotherData doesnt do anyhting at all
probably a skill issue, management issue too
database and the kind of data user expects are kinda different
doesnt shave off unnessrsary data
{ key1 } as { key2} different type kaj korbe na
only subset hole kaj korbe.
database and the kind of data user expects are kinda different
- drizzle-orm auto handle kore eta
- manually korle Pick<typeof Database, "userkey1" | "userkey2'> or Omit<> util use koro
yea pera aitai, bohut transform kora lage data
actually i tried to kill mosquito with a cannon ball, made a type coercer, all cuz prisma replies with datetime, and i need string to send to user, casues ts warning, annoying, so wrote a functiont that goes thru everything, and coerces types and shaves off data if needed
but... problem is, typescript types aint real
lol
just a skill issue, i gotta get better with data transformation
prisma ar skill 2 issue ekane
bro i am cooked, why is prisma so fking slow
π
never made something of this cale
scale
one time pick kori. pore reuse everywhere.
ekane 3 variants userSelect (for /me endpoint), userPublicSelect (for public endpoint /users/:id or as subquery), usePublicSelectExtended (subquery te lagje)
took so long to fetch 3000 models, there i was thinking of pushing 5k fake data per model in there
is neondb fast π
wtf
na
yeah, nah sqlc or go home, problem is prisma schema is 2k lines, massive aah project
nijer project naki open source kisu?
well thats what i have now
discord clone
massive undertaking
discord dekte simple chat mone hoileo onek advanced jinish ase.
jemon mark as unread, infinite scrolling and remembering scroll anchor
yea everything, well most things that make sense to be paginated, are paginated
thats why prisma schema so damn big
last time i tried(gave up), it was 500 lines of schema, i was getting so cooked
good point, scroll anchor
jodio schema size speed e impact felar kotha na. onno kothao issue ase
issue is my vps is cheap lol
local pc, it takes about 200ms in execute a query, even wild big queries
with vps, thats up to 600ms or so
and seeding data, cooked

altho svelte does help more than react would have, its in sveltekit
vue 3 use korso?
ar shadcn, tailwind (development speed onekgun bariye dibe)
isnt vue like shit?
π
svelte is the react, because it actually reacts
react the is re-render-act
svelte compiles to native js, with the reactivity overhead, it is only slightly behind vanilla js in terms of speed
you should try
whats vue good for?
yea tailwind is w
shadcn, or other component library not too good for this, cuz i mean will there be a discord like ui in there
sooo how does it work
svelte er community kom
react er community balo, devloper experience temon balo na
vue holo duitar modde ekta balanced
yea
but dayum, svelte is so reactive, react ee zustand mara lagto
tbh i tried, i dont understand shit of how zustand even works, wtf is this shit, nested aah setup
i always do it with ai π
i aint writng that nested bullshit
jotai deko zustand theke asha kori kom painful
ami zustand use kori na mainly
- alada ts types define kora lage
- infinite callback exceeded error hoi beshi. eta react er architectural issue. vue te ei issue nai
- syntax bar bar bule jai
vue e kisui lage na. ground up theke best way te toiri
export const state = reactive({ count: 0 })
ekon jekono .js ba .vue file theke use kora jabe
import { state } from ''
state.count++
react e context ase. jegulu wrapping tapping bongchong kora lage.
yea, so ass
ki koy
lol
yes
i was gonna say
leptos diye korle kharap hoito na
must be vite dev or smth
for dev mode yes, svelte is compiled by vite live, on client so, understandable why it will be slow
rust is kinda wierd, but it makes sense bro
it does
leptos is the svelte of rust world
yew is the react of the rust world
rocket is the express.js
actix web is the hono
actix always comes out at top in benchmarks
idk man, just doesnt seem realistic to me , everywhere you see svelte beating vue
version info might help too, for instance, svelte 5 reduced svelte's reactive components overhead after compilation significantly
oita leptos er ei benchmark
tara rollup use korse. svelete compiler plugin
vue er creator ar vite creator same. tara notun arekta engine banaitase actively https://github.com/vuejs/vue-vapor
Vue Vapor is a variant of Vue that offers rendering without the Virtual DOM. - vuejs/vue-vapor
tai naki. svelte er kunu benchmark ase jekane vue 3 theke faster
js-framework-benchmatrk most popular benchmark website jekane vue deka jacche faster
I should make a benchmark myself tommorow, I doubt Vue can possibly be faster
Let's see
The faster the better
Doesn't seem like Vue was faster in their benchmarks
ekane keyed ar non-keyed ase. keyed e vue faster (most cases), non-keyed e selvte (most cases)
keyed bolte array.map(item => <p key={item.id}></p>) ortath generally key shobkane use kora hoi.
vue, react, eslint recommend kore.
svelte 3 lol
you are using old version
just added leptos, it is somehow way faster than leptos
accha, old benchmark page deksilam deka jacche
I solved the problem of auto-updating the Discord app on Linux
Especially, Debian based distros
good deep research, useless for pretty much anything else
claude's knoledgebase better
yk what, can replace openai, fk that you reached your limit for gpt4o, yea better lol
not too good for coding, thats all, compared to claude
cobalt lets you save what you love without ads, tracking, paywalls or other nonsense. just paste the link and you're ready to rock!
maybe worth getting
12 months 20gb free
until i get a expensive ssd, query that takes 2.3 second on my machine takes 50 seconds in vps, somehow, idk vps felt fine for everything else i did, snappy asf, but storage slow it seems
i have seen this 44 times in the last hour, sand halp
what difference does it make, just coerce null to undefined and vise versa π
tsconfig.json
strictNullChecks: false
null abong undefined alada action decide kore. tai code e alada treat hoi e.g.
db.update({ name: undefined, email: "@gmail.com" }) // updates except name
db.update({ name: null, email: "@gmail.com" }) // updates both with name to unset
ts undefined recommend kore everywhere. special case chara
I am so used to null π
Good thing I use neovim ig
ekane neovim ki obodan raklo?
frendship ended with claude 3.5. qwen 2.5 max is my new frend
Pattern matching, find and replace
I guess vscode would do fine too, pattern matching isn't nessesary here π, but dang I seen people do crazy shit with this
I havnt yet gotten that much into vim motions
I should
Qwen 2.5's context length smol due to weak ai compute resources in Alibaba cloud, but performs good with smaller questions
hmm
claude er din shesh, qwen qwq er bangladesh π
but small context in their official chat
QWEN CHAT DISCORD
This is a blog created by QwQ-Max-Preview. We hope you enjoy it!
Introduction
Okay, the user wants me to create a title and introduction for their blog announcing the release of QwQ-Max-Preview. Let me start by understanding the key points they mentioned. First, the model is part of the Qwen series, built on Qwen2.5-Max. Itβs ...
ai internally kibabe kaj kore, etar algorithm amar kase ekono impossible magical mone hoi.
oneke ai ke "autocomplete on steroids" bole jeta incorrect.
impressive bishoi holo ai sentences er babhmurti bujte pare.
e.g.
- strictness buja. output 100 words bolle 100 words ei dibe
- language translation kora (grammar consideraiton kore)
- code output full stop, semicolon, brackets correctly position tik raka
- new code, old code difference bujte para
- sarcasm buja
egula kunu "autocomplete" algorithm ba neural training diye possible na. ei grouper kind na.
jara ai invent korse undoubtedly next level manushjon
open source model er code time paile dekte hobe.
my exact reaction. qwen simply baccha claude 3.7 er kache
first try e kaj korse. copy-paste shudu
did you try the qwen thinking model?
yeah, claude is undefeated champ in web dev
replacing frontend devs and even backend devs isnt that far
but it will quickly get muddy and cloudy with more context, as might be needed for a bigger project
i wanna respect your madre, pay my meat to your padre, mi amore
Listen to $ome $exy $ongs 4 U: https://ovo-partynextdoor.ffm.to/4u
FOLLOW PARTYNEXTDOOR:
Website: https://partyomo.com/
TikTok: https://www.tiktok.com/@partynextdoor
Instagram: https://www.instagram.com/partynextdoor
Twitter: https://twitter.com/partynextdoor
Facebook: https://www.facebook.com/Partyomo
Soundcloud: https://soundcloud.com/Party...
hae. qwen thinking o claude normal er dare kache nai.
webdev na. every cases. example ditasi
- extension create error hole claude e
handlepointer tacloseLibrary(handle)diye close kore memory free korse ja qwen mistake korse
- claude e unique_ptr use korse ja C++17 feature abong recommended. qwen use korse raw pointer ja memory leak prone, discouraged.
code dekle erokom hazar example dite pari.
shob ai coding pare, quality maintain ekmatro claude pare jake chok buje rely kora jai.
damn
send system prompt at the beginning when chatting with ai
Output instructions:
- when modifying code, only show the changed part
- output only copy-pasteable code
- don't explain unless asked, don't add inline comments
- suggest alternative solutions if exists with only one-line explanation
Coding style:
- use modern syntax e.g. const, arrow functions, async-await, try-catch
- make sure it's TypeScript safe
- follow my .prettierrc config
"""
"semi": false,
"singleQuote": true,
"trailingComma": "none",
"arrowParens": "avoid"
"""
deepseek r2 coming soon
it was planned for may, but they are rushing, probably because it wont be superior to claude 3.7 anyways
cooked
imagine a foss claude 3.7 in 2026
hmm
rewriting entire app claude 3.7 + cursor
not the only one. February 2025 update 100% copilot https://www.reddit.com/r/programming/comments/1gf0vf9/vscode_october_2024_version_195_90_of_the/
cargo but for c++
fk it, i think i am onto creating a new one
ami c++ dia eita korsi
pros - js style syntax
cons - manually writing headers file (biggest time waste), no modules, slow compile time, no standard package manager tai higher chance of abi mistmatch issues
rust check korlam - otirikto boilerplate code, ja confusing
ekon babtasi zig e rewrite kormu. zig JS style + no headers
distopean
wtf
ultimate ai gf π
why does she talk like that
trained perticularly to be gf or smth
she even rememers what you spoke about last round
fk vcing randos
imma vc this bot now
Nah after 30 min vc, both them model boring, Gemma 27b anyways
Some progress, you need to pass in build spec, kind of cargo like exp, just need the build spec, cuz... You'd be importing codebase that uses other build system
hmm.
10x faster typescript
yes ts lsp is horribly slow
sometimes i wait upto 5 seconds for suggestions to appear for large codebases
sometimes it doesnt even show up
hmm.
cursor ai manus theke kom na. paid plan worth
web search ditase, error read kortase, terminal e command run kore shetar output kemon check kore code improvement kortase. shob automatic jotokon porjonto perfect na hoise.
ami age gajakuri cline agent use korsilam. agent kototuk real hoite pare feel ta pai nai
nice
should be available in insider edition
not pushed to main one yet it seems
there is this agent mode, i wonder why avante.nvim doesnt have anything like that rn, it is easy to run terminal commands in neovim, monitor contents of that buffer, feed to ai
codebase so large, bro is cooked
real
@cloud badge kire bro, cursor er output ee erom hoy?
π
Maybe 1200 line ei aito code changes ig
That is a big file, big component, not ideal
na. amarta 10k lines TS codebase. single file e na jodio.. modular
single file beshi boro hole manually specific context dile kaj korte pare
Select lines + Ctrl+I to add context to sidebar chat
Select lines + Ctrl+K for asking inline directly
mention function in chat @functionname
I get it, output can be max of ~~800 lines total each file
It doesn't have a agent watching over it, like dang stopped at line 802, continue where you left off shit
hgmm
is Zen browser worth the hype?
Yes
yea i dont use it
but zen is good
you may like qute, if you want a different breed of browsers
Yeah
It's trying to be like arc but unfortunately has pretty much surpassed it
Real
nonononononono
I use Vivaldi btw
Yuoh
Oh nice then
its bad
sponsers zen glazers
brave is better
Nah its good
google didnt win, in my testing google gemini 2.5 pro is behind claude 3.7 thought
why youtubers jumping every time a new ai comes out π
day 20 of using cracked cursor claude 3.7
Where is the monkey with cat pfp
That was a W pfp
Yo
agula ki eikhane lol
cat
where is fish
in cat

no i meant cat ate the fish
:)

so i tried zen browser again today
its really bad
the same people who use notion use this goofy ah browser
average hyprland notion zen goof ah
thinks they are productive but all they do is waste time
bro dawg, try qute
slim, minimalistic, same adblock as brave
All the minimalistic browsers make the tabs small and it feels weird
Partly cuz of muscle memory
I want those big buttons man
amar kase small tabs aro balo, space bare.
mouse unsupported main issue. ekane ekta button click korte age read korte hobe pore 2 bar click korte hobe
shob kaj keyboard dia hoi na
mouse ase toh
yea
How is it bad though
Your browser isnt gonna make you productive
Its better than having to customize firefox
It is much easier to style
And has vertical tabs before firefox implemented them
atleast you can organize your tabs and stuff where in firefox it gets very cluttered
But yeah calling it a goof doesnt make sense
Its not similar to hyprland
In hyprland you customize everything yourself while in this browser you dont really need to
zen/firefox goof
every browser try korsi. 5-10 din pore abar chrome/chromium e come back kori. devtools, performance, syncing, less clutter ui, less bloat, extensions marketplace shob dik bibechona kore
yea
brave is w
after removing all those bs tho
Chromium got the biggest nerf in extension market
brave still has strong adblock, not extention level
baked into brave
same with qute, same adblock, just with python bindings
Dont they just get the filters from ubo
How to become a dev
Learn to solve problems
kibabe? adbloker?
Yeah manifest v3
ublock latest version 2 days ago use kortasi
manifest v2 ekhono purapuri remove kore nai
2 bosor dore suntasi. jai hok hoile ublocklite e switch kormu, ahamori kisu na
eventually shob browser mv3 adopt korbe
Firefox will keep maintaining mv2
And they are taking actions
I don't have issues man, I got skills
Then start vibe coding
What coding now
brave's extention will still exist
But the extension store wont
vim binds?
@graceful gustwhy are you blabing about firefox
i use brave
sure buddy mald more
kind of
chrome store e jodi alternative adblocker na takto tokon bujtam biggest nerf. ublocklite, adguard mv3 onek option ase.
arekdike firefox ekta dying browser, niche user der jekane extension store e ekta na hazar extension nai
ar ja ase oitao outdated
firefox store
YouTube can sometimes even detect the original ublock origin. The mv3 versions are obviously weaker. But yeah, Firefox is kind of dying in terms of market share, then again only other option is a chromium based browser which lets google the ecosystem
firecox will continue to have manifest v2
ubo
who tf uses adguard adblokcer
we are not feral european setlers in north america in 1600s
Then?
that is a browser extention specification
manifest v2 is more permissive
manifest v3 doesnt let extentions do shit like network blocking, have "too many operations" or act as adblocker
anything to do with adblocking
basically
well it is still predicting
human brain nijei simple prediction machine.
ekane jodi prediction means pattern matching/autocomplete? tahole bhul
ai does plan, reason, generate the unseen != autocomplete prediction
still predicting
it is predicting the next token
it is made to do 1 thing, predict the next token, similar to the training data
to my understanding
it can think, even in thinking, it is predicting next token
with the additional context(thinking) it still predicts
building engineers der 1 tai matro kaj, ekta brick er upor arekta brick raka, etato amio pari
artist der 1 tai matro kaj. kagojer upor ronger thuli sporsho korano
That's just overly simplified gist
thanks sherlock 4 leting aware
ai predict the next token is overly simplified gist
GPT ba modern ai ke jodi boli "give me 10 words" tokon 10 words ei dibe.
etai cognitive ability.
"next token prediction" algorithm pre-ai era tei chilo kintu kunutar cognitive ability ki ase?
prediction diye "cognitive" achive possible na. 2 ta alada system
ai next token predict kore prottekta word er embedding (unique number representation) er similarity score calculate kore.
ekon main issue holo je ei prediction outside theke control kora jai na karon ai nijer icche moto max similarity words pick korbe.
ei outside control tai cognitive ability. e.g. safety control, jailbreak control, cognitive control (give 100 words), hallucination control. egula ekekta complex system.
oitai claude article e bolse "more than token prediction"
If ai takes over, it is confirmed that i will be dead and you will become one of them
I was letting you know that there are more into it
And not just limited to what you have said
We need that
and you aware of entire context? oita "analogy" example.
hmm
No
I'm an AI and that action was done automatically
affirmative
how i live knowing claude 3.7 thought exists
normal > thinking
You're a senior AI engineer who'll be helping me with coding.
# Instructions
- Output only the modified diff parts
- Preserve original comments and you avoid adding more inline comments
- Keep explanation concise/one-line
- Suggest simpler, performant solutions when available
- Refactor code if it doesn't follow these rules
# Coding style:
- Avoid abbreviating variable/function names (e.g. Use "event" over "e")
- Use == over ===
- Use `Boolean(variable)` over `!!variable`
- Single-line ifs
- Follow `.prettierrc`
"""
semi: false
singleQuote: true
trailingComma: none
arrowParens: avoid
"""
## TypeScript
- Use modern ES14+ syntax (e.g. const arrow functions, async/await, try/catch, logical OR assignment ||=)
- Use type over interface
- Avoid explicit return types when it can infer automatically
## Vue 3:
- File structure order: <template> first, then <script>, then <style>
- Always use shorthands in templates (e.g. `prop` instead of `:prop={true}`)
- PascalCase for component names, camelCase for props
- Use modern v3.5+ syntax (e.g. defineModel, defineEmits named tuple)
## Tailwind:
- Use following spacing scale: 2,4,6,8 for p/m/gap
- Use size- when same width/height instead of w-/h-
- Put CSS keyframes in <style>
## shadcn-vue:
- Avoid for= and id= in <Label> and <Input> components
- No need for mr-/ml- in icons inside <Button>
for me
refer to this before wrtiing svelte
https://svelte.dev/docs/svelte/v5-migration-guide
hmm
infinite scroll userscript still going strong
Organisms therefore have the best developer, with the best code
organisms ki
how do we know you are not a horse
and not mad at hydrogen powered horse robot from japan
In this video, we will discuss:
β
Leaked design from the Robotic Horse from Kawasaki
β
Predicted specifications and superior features
β
Possible engines and technologies used
Will the 2025 Kawasaki Corleo be the dream bike for speed lovers? Don't forget to Like, Comment, and Subscribe for the latest updates on the automotive world!
Sour...
can horse wash after poopin
that reminded me of pre-ai era movie
Can computers clean their temp folder?
on their own?
on linux yes
wait no
that's not a Computer
it's an OS
hmm
let me ask this
Can a computer clean its insides from Dust?
except for machines that are built specifically to clean computers or any other things
my computer does
oh nahh
after ai
In this video, Atlas is demonstrating policies developed using reinforcement learning with references from human motion capture and animation. This work was done as part of a research partnership between Boston Dynamics and the Robotics and AI Institute (RAI Institute).
#BostonDynamics #ReinforcementLearning #RLPolicy
BREAKING: Google just released Firebase.Studio! In this unedited reaction video, I explore this brand new AI-powered development environment that lets you build mobile apps, web apps, and backends using Gemini 2.5 Pro. Firebase.Studio integrates Gemini AI with Firebase's powerful backend services, creating an all-in-one solution for rapidly buil...
new design
Do you like it?
100%. eta material 3. ami android lolipop theke google design follow kori.
koekdin age ami coding webapp screenshot disilam oitar ui o erokom
modern design = flatness
age google "shadow, border" techniques diye title bar, body, sidebar seperation korto.
pore material 3 te google deklo color shades diyei eta kora jai ja flatness barai. kubi creative idea
ekon color shades er arekta issue ase ota holo foreground abong background "contrast" maintain thake na. ei karone google pore new "HCT" color model invent kore. ja "Material You" e use kore. ami etar ekta typescript library korsi https://github.com/rakibdev/material-colors
material 3 te colors google ojota overuse kore na (e.g. material 2 title bar e color high contrast toiri korse). ete content er upor focus kome jai
not a fan of this autistic design π
hmm. toile kun design er fan? example?
ami only UI design ei na. google er hardware design o follow kori. pattern same.
amar mothe google products er high market share er ekta boro reason tader design.
pixel 9 phone
- material 3 er rounded corners, pastel colors dise
- tiny "dot" camera dise, waterdrop or iphone notch na. kubi carefully thought out
pastel colors
Ager ta practical design silo
Aita autistic design, inconvinently centered for no reason
search bar ta? center tho shobjaigai claude, gpt
isnt it annoying
why does it move
it should be where it stays
i am not fan of certering literally everything
π
understandable. centering material 3 er ongsho na jodio
me in hell, to the guy who invented centering everything
Programming is hard, never have I ever comprehended this until I tried raw dogging rust ππ
Good god, how do people even use c, atleast compiler in rust prevents you from creating lose lose situations
You compile, then you have to deal with this dumb aah errors π, how would you even go about debugging them
It is not that bad π, but dang
ohshit after 1.5 months of crackin got ip banned from cursor
anyway, cloudflare warp working for now
I found C easier to write than Rust
You usually compile with debug information and plop it in GDB (or a GDB frontend or clangd)
Yes but you need to manage race conditions
I believe I have seen some fked up nested pointer shit
Atleast here the pointer nesting makes sense
so?
you're not free from race conditions even on Rust
You are
It will be picked up on compile time
it is literally impossible to detect all forms of race conditions
Unless it shoot yourself in the foot too hard by circular dependencies
Yes but most of it is picked up
it is very easy to create race conditions in rust, all you're free from is a very strictly defined set of race conditions
Yes the most common set of race conditions
i haven't written rust in a few months so i can't remember them right now
Yes possible but you won't be doing them as often is what I am saying
no?
Rust just gets deadlocked
Also how tf do you get dependencies bro
True
thus it's considered "safe"
but it's still a race condition
and they're still there
and they will cause issues
How does one get dependencies for c/cpp dawg
there is also this https://conan.io/
Conan is an open source, decentralized and multi-platform package manager for C and C++ that allows you to create and share all your native binaries.
(which uses CMake and stuff under the hood)
Ig I will work on that when I have time, it supports adding cmake mason static header dependencies pain free
just use conan at that point
It doesn't support all packages, very restricted set of packages
wdym?
It doesn't have very many packages
CMake? lol
How do you even use it, it has almost nothing
you can use anything with CMake
Why does it say, like 2000 libraries or smth
Last I checked
say where
CMake doesn't have a central package storage, because you don't need one
you're probably misremembering something
Hmm no it doesn't say that
Yea
You always need one
π
clearly not
Literally nobody would be using rust otherwise
Very steep learning curve compared to c/cpp
C++ is objectively harder (or not, might be a subjective opinion lmao)
Bro what π€₯
but anyway, Rust isn't standardized yet
π
that's why there aren't any crates host other than crates.io (but u can still use Git links directly)
How is c++ harder
ππ
I used it, just havnt got around to using cmake or any build systems
Obscure resources, less friendly learning environment, slapped together features that aren't as carefully thought of as Rust's thus results in devilish syntax
just a few that bothered me
ask someone who has been writing C++ professionally and they will say Rust is easier and nicer
In the grand scheme of things, rust is easier, but it has a far steeper learning curve than cpp
only if you've never written C or C++ before
Bro I have
in that case, you don't really count
what's the most complex thing you've written?
In cpp?
or C
okay man
Lol
that's not enough
Probably, but c'mon, it doesn't seem like cpp is hard
Just maybe design structure may be
abandon seem
Connan can technically import and use any cpp library?
C++ is nice if you're using it like C but only utilize the richer STDLIB
know a few people who use it like that
what is a friend
i aint reading all that witchcraft
so they're with benefits
Lol
as long as its available on Conan, otherwise u need to submit it
so technically, yes
Bro then mine is superior
You just need a build recipe, that's sll
All
Supports fetching from git over https or ssh, and fetching archives/ static file too
CMake does that regardless of conan
Cmake feels like witchcraft but true
FetchContent_Declare(
raylib
DOWNLOAD_EXTRACT_TIMESTAMP OFF
URL https://github.com/raysan5/raylib/archive/refs/tags/${RAYLIB_VERSION}.tar.gz
FIND_PACKAGE_ARGS ${RAYLIB_VERSION}
)
FetchContent_MakeAvailable(raylib)
target_link_libraries(${PROJECT_NAME} raylib)
for example
you might like xmake
it uses Lua
has similar features
otherwise just vendor dependencies with git submodules (best way IMO)
c++ easier. c++ e ase oop, auto memory management, less boilerplate (e.g. async-await) etc...
c te coding theke language er shathe fight korte hobe beshi
by auto memory management, do you mean RAII