Everybody's losing their mind on deepseek. But it's the same trash as ChatGPT.

Look what 'trash' gpt can do. This is a tool to automate AI with a text file. In this case it's instructed to be a sql expert and is mentioned only a database sql schema. So far, it didn't make any mistake regarding my questions. It even add the nick column while i ask for the nickname column which is correct. It let's me a add extra fields later / refine it. How impressive is that. I'll have to add textbox again, but retoorded is able to deploy based on voice commands. retoorded can be teached "if user ask for some love do a get request to https://your-server.com/secret-depl...". That's amazing. My dutch local AI assistant that I full time talk with, can remember variable names and can do several stuff with it. I can zip in all my code, upload it, let it analize it for quality, structure, security, performance etc. I brand my files to push them all trough GPT and let it add a summary about what the file does, my e-mail above it, and it adds the MIT license.

From what I understood it's considerably more efficient, thus costing less and being easier to train(=needing much smaller dataset), was created with a fraction of the funds OpenAI has, made some new(?) learning algorythm that learns much faster and from less data than GPT.
Also it's fully open source, both the code, and the model/dataset.

Which, if all that is true that sounds like a pretty huge jump, so I'd understand why there is hype.
I didn't dig into it too much tho, all I said above is from hearsay(so don't trust me too much).

I wonder what you two will have to say about it: @retoor @jestdotty
I'd think to have interesting takes/opinions about it

@SoldierOfCode I have to check but by my own local reasoning I do not believe that they open sourced the dataset, only it weights what I consider useless. But I do not believe it would've opened the dataset because the dataset is actually millions of query results from OpenAI. It is so heavy trained on OpenAI that it responded to be made by OpenAI. Not sure if that's still the case - let me check - nope. It says in a short sentence that it's made by "China's Deepseek" company. Very funny, like OpenAI would say "Murica's OpenAI'. The short answers already says it's a bit hard coded I guess.

I do have one test to check quality of a GPT model, will execute soon. It's a json file with all devRant statistics. The only AI good enough to analize it correctly was OpenAI so far. Of course, it can fail it and be better in other stuff, but this is a part I care about. Have you provided DeepSeek your email yet? :P

@SoldierOfCode I'm trying it to make cool statistics and comparisons between users by uploading this file: https://molodetz.nl/retoor/drstats/...

It says it's too busy after long thinking. I asked GPT, and GPT executes manually code to parse the JSON file and applies AI afterwards. According to the dev of perplexity, this kind of tooling will become higher priority for OpenAI than the model itself. The model itself is good enough to do anything as long they have good data to work with. So, the tooling - that's probably the current battle. Same for that to solve a sudoku, it actually uses python to create a sudoku solver (he shows that to you) and executes it. Other models don't have such advanced tooling. So yeah, I would still prefer OpenAI I guess. But so far, DeepSeek R1 doesn't respond but I assume it won't use manual tools like OpenAI does. Without these tools, it would never be tht awesome tho.

@SoldierOfCode it's a story about you. Statistics from last few weeks. You were on your top secret mission and caring for your camel so not that much data I guess.

@SoldierOfCode I have my detailed information about DeepSeek by trying it by just now (I just figured out that the model is actually online like GPT is and not some research model that no one can use). And I have my information from this video: https://youtube.com/watch/.... It's comments by a the competition (perplexity AI). It's not commercial fan fiction like 99% of the other video's.

But my conclusion so far: AI is more than a model, it's also the tooling. Without the tooling, it's just not the same. So storm in glass water. If Deepseek actually HAS the tooling, I stand corrected but I'm in high doubt. That it will do the statistics assignment OK without tooling is still a chance if it will respond someday.

@retoor Wow, I've never raid so much "nothing" slop before xD
But that's AIs for you I guess.
It literally wrote whole 4 paragraphs without avtually saying *anything*
Sounds like corpo speech, where someone's just trying to sound smart/sophisticated so they just use a LOT of a formal and smart sounding words lmao

@SoldierOfCode Yeah, i asked it for a story and it took the story part quite literally. I think it's cool tho. This is a more serious objective observations about you. Your privacy is as good as your friends treat your data I guess :P That's why I in the end actually did register a facebook because friends were stuffing it full with pictures and mentions of me. So facebook knew who I was before I used it. You can care about privacy but your friends don't give a fuck (like me) or don't know how to handle privacy. What exactly is considered privacy is vague and a bit personal opinion maybe. Is a phone number something really private? It's literally a device to get accessed on. It depends much on the user if it is or not. I would consider private stuff more information that could be used to harm you.

@SoldierOfCode I asked "Based on statistics, could SoldierOfCode and SidTheItGuy be in a romantic relationship or would the other be out of reach?" and it said to me that it's weird to base that information on community engagement statistics. Damn, there goes my devRant dating project. Anyway, it answered this:

@SoldierOfCode Ignore previous post, it assumed that SidTheITGuy was a fictional character, That's why it is spelled with the wrong uppercasing. This is a correct comparison.

@SoldierOfCode I told it from what countries you guys are and I don't see red flags in sense of compatebility. It sees the fact that both of your cultures are very engaged with their family as a positive shared thing. Religion differences can be talked about with clear mind it says.

So nice. Ever considered Sid?

Fuck, wasted enough time on AI again, time to do stuff.

But hey, it's not completely useless. I've learned some things about culture of both countries. But calling SoldierOfCode a introvert is a huge mistake I guess.

@retoor Bruh he's literally married.

@retoor And the second description is still just a bunch of text with no points. It's trying to meta analyze me through post lengths instead of the actual post contents? I don't get why.
But yeah, a bunch of text and it still said nothing.

@SoldierOfCode it only had the statistics json I've mentioned and can't say more information than that but it can do comparisons and calculations about how much more x the other is in percentages. That's not bad. If you want opinions about the stuff you said, you need a decent phone (android) with microphone access enabled and ask https://retoorded.molodetz.nl. It's hard for STT to recognize names but I think it will parse SoldierOfCode. I have to add a textbox to that app but I'm working on a slack alternative which goes really well. If you don't see clippy the Microsoft assistant, you have to tap screen and wait until it shows up and microphone is enabled. It's browser security fucking up comfort of using such tech. Reason prolly you don't see this tech much. Only works on webkit also. So, chrome is fine, but android is best for STT and TTS.

@SoldierOfCode I literally asked for statistics. I tried a few times but it recognizes my pronounce of code as God every time after doing it one time correct. Got tired. But you can ask it to summarize the things you've said. It has a huge dataset regarding devrant. The STT engine is actually very good but it requires proper English. My English is quite OK but the pronounce is really Dutch. I'm too Dutch to function. All information about whatever you've said in a single file last few weeks is in the https://molodetz.nl/retoor/drstats export folder. That's the data it works with. It won't provide more than that. I couldn't apply markdown on clippys text box. I could've just removed it tho. But that project us already wasted too much time on for it's value and it's only a demonstration of my ragent backend library. The frontend and tts is just an extra and not what's it about.

- the deepseek devs outcropped from quant company cuz they made too much money and their salaries got capped by China communism. they are PR department now. also china economy bad and is front for one world government
- deepseek open source, people stripped reasoning off it already which evidently is superior but otherwise
- it was trained on chatgpt, that's why "cheaper". easy to reproduce
- Asians can't build anything new. look at their history. they only perfect tech by taking things away. they think like this genetically. also the meme posts about it is very Asian PR

basically it's nothing. just political angry noises