netdur

On gemini.google.com, I see options labeled 'Fast' and 'Thinking.' The 'Thinking' option uses Gemini 3 Pro

hmm I wonder if we can collaborate somehow, I am trying to bootstrap https://hrflow.app

by that logic GPL wouldn’t be open source either since it also adds restrictions unlike the fuck license

nowadays github is filled with so-called open source projects under GPL3 but the maintainers want you to pay for a dual license

to be honest i still feel satisfied but when it comes to making something useful, in the past i gave up programming because of the endless repetitive tasks, you want to build something cool but wait you first need to make a auth system and by the time you finish that the cool idea is already dead because of how boring and repetitive it all is, ai coding made it fun again

Any idea what kind of server we’d need to handle 40M users if it’s just text chat (no files or audio)?

Yeah, ChatGPT is better overall, but with a twist, Gemini can actually be the best if you use AI Studio, tweak the config, and set up a good system prompt, kinda like how nano banana is SOTA, but Qwen-Edit feels more useful since it’s less censored, meanwhile ChatGPT is starting to feel slower and kinda showing its age, another example is Veo 3 being SOTA while the infamous Grok is technically worse but doing better, and OpenAI’s Sora is pretty much dead

edit: I run low profile service that localizing e-commerce photos, like taking Alibaba listings and swapping the model to look local alike, with nano banana I can’t automate it because I have to manually check if the output got blocked (anything with female skin is risky, underwear or cleavage is 100% blocked), but Qwen-Edit just does the job without fuss

I believe it was TikTok’s U.S. operations that were sold for $14B, The rest of the world is still under Chinese ownership... hmm, did I just say ‘the rest of the world is still under Chinese ownership’?

just unsloth on colab using A100 and dataset on google drive.

the model runs on H200 in ~20s, costing about $2.4/hr. on L4 it’s cheaper at ~$0.3/hr but takes ~85s to finish. overall, H200 ends up cheaper at volume. my scan has a separate issue though: each page has two columns, so text from the right side sometimes overflows into the left. OCR can’t really tell where sentences start and end unless the layout is split by column.

I’ve tried that too, trying to detect the scan layout to get better OCR, but it didn’t really beat a fine-tuned Qwen 2.5 VLM 7B. I’d say fine-tuning is the way to go

I upgraded to 2TB of storage since 15GB just wasn’t cutting it. Two years later, I’m already at a whopping 200GB!

Meta is run by people with no regard for ethics, and if that surprises you, that’s on you. Their whole model is just packaging and selling you with whatever tech they can grab. If you’re worried, don’t install Meta apps. I’ve got WhatsApp on Android and Instagram on iPad, They’re already getting eaten alive by TikTok and AI girlfriends

You’re conflating editing with rendering, YouTube didn’t overwrite creators's uploads, it applied an ML filter in the streaming/transcode pipeline, the same layer that already resizes, compresses, and tone-maps. That's not "editing my content" any more than your TV's sharpness setting edits a film, An "Enhance" toggle/label would be good UX, but calling it silent edits misdescribes what's happening

PS: this isn't "generative AI" It's basic ML enhancement (denoise/sharpen/tone-map)

This outrage feels odd, TV has "improved" movies for ages, youtube doing it with machine learning is the same idea, are we really upset because an ear looks a bit clearer?

Man, that whole ‘please make an RSS or JSON feed for me’ request reads like Richard Stallman emailing himself a webpage to print out later, let them use Twitter or whatever medium they wants

output tokens must be generated in order (autoregressive decoding), inputs don’t have that constraint, so prefill is parallel, with stronger kernels, KV-cache handling, and batching, Claude can outrun Gemini.

I love pricing pages, I avoid landing pages or whatever they want to me read and go directly to pricing page to get the meat of what they offer... then I look at price.

Mark Zuckberd must very very upset looking at this, I expect him to throw another billion dollars at google engineers

He speaks in movies terms, exactly what I say when I watch movie about programming

naive look, 2/3 of model, without multi-languages this shiuld be around 1B

it's big little planet or small big planet?

I have fine-tuned Gemma 3N 2B and it's pretty good, but loads slow on my S23U, once it's loaded though, it works fine

Also tried SmolVLM 256M and 500M, they load faster and you can embed them in assets, they work if you know what you're doing

Just keep in mind that smaller models don't perform as well due to their limited parameters

Also on Android, since you can't ship files larger than 2GB due to Java compression issues, you need to download models separately, then you can't load the model from the download folder, you have to copy it into the app's own folder, this means a Gemma 3N 2B model that's 3.14 GB would need at least 7 GB of free space on the user's phone

Working on https://github.com/netdur/llama_cpp_dart it is llama.cpp binding for Dart first then Flutter I am currently working on multimodal support, add vision and working on audio

you must be naive to think OpenAI does not train on your data, Altman is infamous for deceiving claims.

My understanding is that in multimodal models, both text and image vectors align to the same semantic space, this alignment seems to be the main difference from text-only models."

I followed this guide for fine-tuning: https://ai.google.dev/gemini-api/docs/model-tuning

Arabic OCR is a mess with historical texts. Take the word الف (alf/thousand) in dates like 1950 - in old documents, the ف (fa) had a dot below it, but modern OCR doesn't get this and outputs الد (alad), which is just gibberish in Arabic

Same problem with ق (qaf) written as ف (fa) in old Arabic

And don't get me started on merged letters! In محمد (Muhammad), sometimes the م (meem) sits right on top of the ح (haa), or appears as a little circle below the line. Modern OCR has no clue what to do with these

My solution? Run OCR first, then use LLMs to fix the mess based on context. The surprising part? In my tinkering, smaller fine-tuned models actually do BETTER at this specific task than the big general-purpose ones. They seem to learn the patterns of historical Arabic quirks more effectively. Pretty neat tradeoff of specialized knowledge vs. general intelligence

I have documents from the last 50 years that I need to digitalize, millions of them written in old Arabic. The OCR is not accurate due to handwritten documents, so I need to fine-tune a model on around 300k pairs of texts (OCR output and manually corrected versions)

I never heard it was bad at twitter! why is that?

Yes, there is Standard Arabic and various Arabic dialects. The Middle Eastern dialects are closer to Standard Arabic while still different. In Morocco, Darija appears to be an Arabic dialect but it's actually more of an Amazigh language that uses approximately 70% Arabic vocabulary, which can be misleading

seems on par or better than gpt4 mini