Systems engineer Yacine claims open-source AI models trail proprietary options by only four months
Victor Taelin joked that OpenAI is actually lagging behind.
Many users agree open source AI models are closing the gap to proprietary ones rapidly and celebrate the coming wave, while others criticize the term as inaccurate due to CUDA dependencies and limited model access.
Most Activity
@yacineMTB s/source/ai/
Remember, open source models are only ~4 months behind now

@Hydration_HQ codex

@yacineMTB How are they going to distill without access to fable now?

@yacineMTB What is the go to nowadays?

@steady_force @yacineMTB Slow down cause Chinese won't be able to use Fable to help train their open source ones.

@yacineMTB I just wanna run fable 5 locally on my iPhone 12. Is that too much to ask?

@yacineMTB next the US will tell nvidia to block chinese models on their infrastructure 😭

@lazyvibecoderx @yacineMTB Don't they get info directly through foreign anthropic employees or through other leaks about model architecture and training? If this is the case I assume they would want to get it out even more

@yacineMTB nice meme yacine

@yacineMTB I wont count as opensource any model which was built on CUDA and they never bothered to make it work on an opensource GPU system, even tho they keep saying in theory it can be. It can all be taken away. China found out.

@yacineMTB They really aren't

@tommyvedvik @yacineMTB @Hydration_HQ To what? 5.5 just gets shit done.

@yacineMTB virtually no difference in everyday tasks at this point too, for 70% of tasks the prompt is more important than the model

@yacineMTB Gonna be hard to distill a model you and your spies don’t have access to as a foreign national

@yacineMTB The biggest constraint is very quickly shifting from access to highly capable models to access to infrastructure.
When you need a $500k computer to run a 1T class model, the $200/month subscription feels cheap

@yacineMTB God bless communist China for keeping education free and accessible to all

@yacineMTB Any consensus on the best open source model out right now?

@yacineMTB It will be back on my Monday.
Or market will crash.
The biggest customer base for AI is in India.

@yacineMTB Did they have enough time to train though?

@yacineMTB Codex 5.4/5.5|Opus 4.6 region was the "rsi" boundary layer for self improvement of the models which the chinese more or less just hit. We should ideally see a similar ramp up now from them given they now have this capacity or near it