Chatgpt - Seriously good potential (or just some Internet fun)

[RxR]

I just uploaded a few of my working papers to deepseek and queried them using its reasoning engine. The results were pretty impressive - eg. inferring and giving reasonably accurate 'how to' procedures when these were only implied in the texts.

There was a semi-hallucination / overgreedy inference though. Which brings me to the point: using a reasoning engine to output descriptions and how-to's (for example) can indicate where the original text needs some sharpening or emphasis added (to prevent the hallucination). So they have a use as a prepublication editing aid.

 

[D.P.]

You can get output that rivals ChatGPT, and it can run locally on a 6GB laptop GPU - that isn't a big deal to you?

Deepseek r1 is a 671 billion parameter model. It isn't small by any means. The largest model from Meta is only 450bn.


R1 is.not going to run on your laptop, i'm not sure where you get that idea.

 

[adam cool dude]

Surely this is the call for us as a country to ditch America and cuddle up with the Chinese.

I think it is pretty obvious that China are the way forward in the next 30 years.

 

[darkgen]

Deepseek r1 is a 671 billion parameter model. It isn't small by any means. The largest model from Meta is only 450bn.


R1 is.not going to run on your laptop, i'm not sure where you get that idea.

You can run the distilled models with varying degrees of success and accuracy on a moderately powered nvidia GPU PC. The higher the vram plus system ram will increase your success. I've been playing with the 14b models via ollama and open webui with a 3080 (10gb) + 16gb of system ram. It's not fast but usable and I've been suitable impressed with various maths. logic and coding queries.

The reasoning and thinking output is rather interesting.

 

[pre]

Probably not quite the situation but...

 

[arknor]

chatgpt is so stupid...

been trying to solve the same problem for hours, chatgpt goes around in circles the whole time, makes stuff up on it's own, you tell it to read a link and it just ignores all info on there...

and t hen suddenly the penny drops....

If you're working with video instead of images, the prediction format and processing need adjustments.

​

• Label Studio sends video URLs, not images.
• You need to extract frames from the video, run YOLO on them, and return a prediction for each frame.
• The response format must match what Label Studio expects for video annotations.

I told it in the first place it was a video..... yet again making assumptions instead of actually listening to instructions.

I hope deep whatever kills of these crap american AIs, it seems to have a memory span of about 10 minutes too..

it's like its senile or something... it can lose track of what your doing, or info you gave it is totally irrelevant.

chatgpt always knows better than you do.

 

[Hagar]

Surely the logical conclusion of all the claims of distillation and training of various AI engines is that there will not be one best answer but a variety of samie's all producing similar if not identical output.

There can only be one correct solution to most problems.

 

[sityeuropa]

chatgpt is so stupid...

been trying to solve the same problem for hours, chatgpt goes around in circles the whole time, makes stuff up on it's own, you tell it to read a link and it just ignores all info on there...

and t hen suddenly the penny drops....



I told it in the first place it was a video..... yet again making assumptions instead of actually listening to instructions.

I hope deep whatever kills of these crap american AIs, it seems to have a memory span of about 10 minutes too..

it's like its senile or something... it can lose track of what your doing, or info you gave it is totally irrelevant.

chatgpt always knows better than you do.

What chatgpt model are you using exactly?

 

[NickK]

Yeah I don't think people understand this - there's lots of handwaving about censorship but it's not really sunk in that this is an open model, it doesn't have to be hosted in China, it doesn't require a censorship API, other third parties will host it (it's not just China vs do it yourself). Lots of buzz surrounding AI but plenty of people misunderstand it.

If the enc to weights diffused to text then the pretrained weights will recreated problematic text (ie bad code). Ie the weight drives the output.

You can remove the RAGs, APIs and everything else but the weight will cause a problem.

yes you can remove everything and use a MoE with opensourced code with no chinese connections or data but your need to dump the weights.

 

[arknor]

What chatgpt model are you using exactly?

gpt4o on the plus plan.

actually I found a really easy way to do what I wanted when I stopped being lazy and it meant editing one whole file.
then replacing a model that already existed.



this is what I was trying to do

Label Studio Documentation — Write your own ML backend

Set up your machine learning model to output and consume predictions in your data science and data labeling projects.

labelstud.io

ChatGPT4o literally found it impossible to write me a script compatible with that backend.

it couldn't figure out how to reply to predict() calls from label-studio properly.


I linked chatgpt the example yolo

label-studio-ml-backend/label_studio_ml/examples/yolo at master · HumanSignal/label-studio-ml-backend

Configs and boilerplates for Label Studio's Machine Learning backend - HumanSignal/label-studio-ml-backend

github.com

I told chagpt

label-studio-ml-backend/label_studio_ml/examples/yolo/control_models/video_rectangle.py at master · HumanSignal/label-studio-ml-backend

Configs and boilerplates for Label Studio's Machine Learning backend - HumanSignal/label-studio-ml-backend

github.com

seemed to translate yolo formar > studio label json

I linked chatgpt https://github.com/HumanSignal/labe...bel_studio_ml/examples/yolo/README_DEVELOP.md

I even linked it the studio label sdk at one point....


surely it should have been able to figure it out?

I ended up editing video_rectangle.py so it points to
model_path = "best.pt"

my own model..



chatgpt couldn't figure out how to export with interpolated frames bounding boxes either.
I googled and figured out how to do it my self... I think I just copy and pasted the code after googling too...

from label_studio_sdk import Client
# Connect to Label Studio
ls = Client(url='http://localhost:8080', api_key='MYKEY')
# Get your project
project = ls.get_project(1)
# Create an export snapshot with keyframe interpolation enabled
export_result = project.export_snapshot_create(
title='Export with Interpolated Keyframes',
interpolate_key_frames=True
)
# Get the export ID
export_id = export_result['id']
# Wait for the export to complete (you may need to implement a waiting mechanism)
# Download the export
status, filename = project.export_snapshot_download(
export_id, export_type='YOLOv8', path='.'
)

the one thing chatgpt did manage after dozens of attempts was how to convert that json snapshot to yolo format. (export to yolo_obb seems bugged with video files)
finally after many days I have label studio set up with a custom trained model that can pre-annotate, and I can easily retrain.


sometimes itr's brilliant but then its like it secretly switches to a dumb model and literally becomes senile.

I literally ask if it remembers what we ared oing? how can it ask for debug output then when it;s provided, it acts like your starting a whole new topic?

 

[Lopéz]

Seeing plenty of discussion on Reddit about people cancelling their GPT4 plans to use DeepSeek instead.

Feels somewhat premature, I have been trying to use DeepSeek as an alternative/comparison over the last week and I am lucky if I get 1 response from 30 queries. "The server is busy. Please try again later"

 

[Koalaboy]

I just uploaded a few of my working papers to deepseek and queried them using its reasoning engine. The results were pretty impressive - eg. inferring and giving reasonably accurate 'how to' procedures when these were only implied in the texts.

There was a semi-hallucination / overgreedy inference though. Which brings me to the point: using a reasoning engine to output descriptions and how-to's (for example) can indicate where the original text needs some sharpening or emphasis added (to prevent the hallucination). So they have a use as a prepublication editing aid.

This is the key feature for me also.

I've been using Deepseek v3 r1 32B distilled (qwen) locally since it came out, for general conversation and also code, on a Mac Studio. It's not 'super fast' but getting answers within a minute is still much quicker and more successful than random internet searches.
It still hallucinates at times, but seeing the reasoning allows me to drill in further with more questions and context, generally getting accuracy improvements.

Many people seem to just use the online apps (which are more directly censored/adapted) and ask a single question rather than have a conversation. I treat it like a professor or politician, where you need to build up the context in a discussion and then that usually gives much better results overall. Like humans, as much context as possible is important.

 

[sityeuropa]

gpt4o on the plus plan.

Any reason why you are not using o1 for coding purposes? or did you run out?

 

[Ayahuasca]

Yeah 4o is pretty poor now compared to some of the latest models, I mean, o3 mini has just been released and the "full" o3 version should be available shortly.

 

[NickK]

So now I've finished my course - this afternoon I'm have a play with Claude in developing some code for a personal neural network research project.

Seems a little more 'with it' than chatgpt.

 

[arknor]

Feels somewhat premature, I have been trying to use DeepSeek as an alternative/comparison over the last week and I am lucky if I get 1 response from 30 queries. "The server is busy. Please try again later"

people are installing it locally with 0lama or whatever, it saves them paying chatgpt.

if you haver 10gb gpu memory you can probably run some of the smaller models, if you have a 4090 with 24gb you can probably run one of the medium ones
theres guides on youtube for setting it up, it seems like a few minutes process

Any reason why you are not using o1 for coding purposes? or did you run out?

honestly no idea, but surely chatgpt4 should have been able to solve it? it wasn't exactly complex code

if it knows how to convert yolo to json, why can't it do the reverse int he expected format predict() was asking for? even with the readme documentation and examples it failed.

it claims it reads websites and file links, it does the "analysing" thing but it's answers don't seem to be based on your own info provided.

I usually just use github copilot with claude.



for the record I can't really code, I understand what the code is doing though. python doesn't seem that complicated, I should probably just learn it... chatgpt estimated about 6-12 months to learn everything I wanted

I'm just messing around with ai/machine learning, computer vision stuff and games

So now I've finished my course - this afternoon I'm have a play with Claude in developing some code for a personal neural network research project.

theres a programming forum on here, idk if it gets much use though

you using chatgpt copilot? in visual studio claude can directly edit the code

I think it wass like 10$ a month for personal but they do a 30days trial

GitHub Copilot · Your AI pair programmer

GitHub Copilot works alongside you directly in your editor, suggesting whole lines or entire functions for you.

github.com

Your free one-time 30 days trial expires in 24 days. You'll be billed $10.00/month after the trial ends on Feb 23, 2025. Read billing documentation for more details.

 

[dowie]

If the enc to weights diffused to text then the pretrained weights will recreated problematic text (ie bad code). Ie the weight drives the output.

You can remove the RAGs, APIs and everything else but the weight will cause a problem.

yes you can remove everything and use a MoE with opensourced code with no chinese connections or data but your need to dump the weights.

I think you're missing the point that was re: the API and being hosted in China! That's rather fundamental here - regardless of what model you have if you've got a censor in front of it and CCP laws to adhere to then you need to account for that too!

 

[NickK]

That was the obvious point - I was referring to the “chinaless” version but using weights from the china version.

Can you imagine someone using the non-china version in their company thinking it was all safe.. also everyone complains about legacy code bases but given all the consultants simply shove the old java into code whispeer to port to AWS and then everyone wonders why they the. Have a massive problem as nobody still understands the code.. except AWs’s own tooling.. and when you want changes the only way is to use the AWS tooling..

 

[dowie]

That was the obvious point - I was referring to the “chinaless” version but using weights from the china version.

Then this makes no sense:

There's a number of gotchas with DeepSeek:

[...]
2. Data goes in.. stays in and remains theirs for training etc.
[...]

So nobody can use it realistically for the company..

That only applies if you're using their API/using the model when hosted by them in China.

Can you imagine someone using the non-china version in their company thinking it was all safe.. also everyone complains about legacy code bases but given all the consultants simply shove the old java into code whispeer to port to AWS and then everyone wonders why they the. Have a massive problem as nobody still understands the code.. except AWs’s own tooling.. and when you want changes the only way is to use the AWS tooling..

You're not being clear there at all, you seem to be talking about a potential AWS issue there - I don't really see the relevance to this specific model vs some other LLM?

 

[NickK]

Then this makes no sense:
That only applies if you're using their API/using the model when hosted by them in China.

My understanding is their full opensource provided the weights so you have a full text in-text out capability? Happy to be corrected but that is what I've seen discussed - I should have a look at the sources myself.